2005-03-14 23:57:21 +03:00
|
|
|
/*
|
2007-03-17 02:11:45 +03:00
|
|
|
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
|
2005-11-05 22:57:48 +03:00
|
|
|
* University Research and Technology
|
|
|
|
|
* Corporation. All rights reserved.
|
|
|
|
|
* Copyright (c) 2004-2005 The University of Tennessee and The University
|
|
|
|
|
* of Tennessee Research Foundation. All rights
|
|
|
|
|
* reserved.
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
2005-03-14 23:57:21 +03:00
|
|
|
* University of Stuttgart. All rights reserved.
|
2005-03-24 15:43:37 +03:00
|
|
|
* Copyright (c) 2004-2005 The Regents of the University of California.
|
|
|
|
|
* All rights reserved.
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
* Copyright (c) 2006 Los Alamos National Security, LLC. All rights
|
|
|
|
|
* reserved.
|
2007-01-30 01:01:28 +03:00
|
|
|
* Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
*
|
2005-03-14 23:57:21 +03:00
|
|
|
* $COPYRIGHT$
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
*
|
2005-03-14 23:57:21 +03:00
|
|
|
* Additional copyrights may follow
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
*
|
2005-03-14 23:57:21 +03:00
|
|
|
* $HEADER$
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
/** @file **/
|
|
|
|
|
|
|
|
|
|
#include "orte_config.h"
|
|
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
#include <sys/types.h>
|
|
|
|
|
#ifdef HAVE_UNISTD_H
|
|
|
|
|
#include <unistd.h>
|
|
|
|
|
#endif
|
|
|
|
|
|
2006-02-12 04:33:29 +03:00
|
|
|
#include "orte/orte_constants.h"
|
2005-03-14 23:57:21 +03:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
#include "opal/event/event.h"
|
|
|
|
|
#include "opal/util/output.h"
|
|
|
|
|
#include "opal/util/show_help.h"
|
|
|
|
|
#include "opal/threads/mutex.h"
|
2006-02-12 04:33:29 +03:00
|
|
|
#include "opal/runtime/opal.h"
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
#include "opal/runtime/opal_cr.h"
|
|
|
|
|
|
|
|
|
|
#include "opal/mca/mca.h"
|
|
|
|
|
#include "opal/mca/base/base.h"
|
|
|
|
|
#include "opal/mca/base/mca_base_param.h"
|
|
|
|
|
#include "opal/util/os_path.h"
|
|
|
|
|
#include "opal/util/cmd_line.h"
|
|
|
|
|
#include "opal/util/malloc.h"
|
|
|
|
|
|
|
|
|
|
#include "orte/dss/dss.h"
|
|
|
|
|
#include "orte/mca/rml/base/base.h"
|
|
|
|
|
#include "orte/mca/rml/base/rml_contact.h"
|
|
|
|
|
#include "orte/mca/routed/base/base.h"
|
|
|
|
|
#include "orte/mca/routed/routed.h"
|
|
|
|
|
#include "orte/mca/errmgr/base/base.h"
|
|
|
|
|
#include "orte/mca/grpcomm/base/base.h"
|
|
|
|
|
#include "orte/mca/iof/base/base.h"
|
|
|
|
|
#include "orte/mca/ns/base/base.h"
|
|
|
|
|
#include "orte/mca/sds/base/base.h"
|
|
|
|
|
#include "orte/mca/gpr/base/base.h"
|
|
|
|
|
#include "orte/mca/ras/base/base.h"
|
|
|
|
|
#include "orte/mca/rds/base/base.h"
|
|
|
|
|
#include "orte/mca/pls/base/base.h"
|
|
|
|
|
#include "orte/mca/rmgr/base/base.h"
|
|
|
|
|
#include "orte/mca/odls/base/base.h"
|
|
|
|
|
|
|
|
|
|
#include "orte/mca/rmaps/base/base.h"
|
|
|
|
|
#if OPAL_ENABLE_FT == 1
|
|
|
|
|
#include "orte/mca/snapc/base/base.h"
|
|
|
|
|
#endif
|
|
|
|
|
#include "orte/mca/filem/base/base.h"
|
|
|
|
|
#include "orte/mca/schema/base/base.h"
|
|
|
|
|
#include "orte/mca/smr/base/base.h"
|
|
|
|
|
#include "orte/util/univ_info.h"
|
|
|
|
|
#include "orte/util/proc_info.h"
|
|
|
|
|
#include "orte/util/session_dir.h"
|
|
|
|
|
#include "orte/util/sys_info.h"
|
|
|
|
|
#include "orte/util/universe_setup_file_io.h"
|
|
|
|
|
|
|
|
|
|
/* these are to be cleaned up for 2.0 */
|
|
|
|
|
#include "orte/mca/ras/base/ras_private.h"
|
|
|
|
|
#include "orte/mca/rmgr/base/rmgr_private.h"
|
|
|
|
|
|
2006-02-12 04:33:29 +03:00
|
|
|
#include "orte/runtime/runtime.h"
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
#include "orte/runtime/orte_wait.h"
|
|
|
|
|
#include "orte/runtime/params.h"
|
2005-03-14 23:57:21 +03:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
#include "orte/runtime/orte_cr.h"
|
2005-03-14 23:57:21 +03:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
int orte_init(bool infrastructure)
|
2005-03-14 23:57:21 +03:00
|
|
|
{
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
int ret;
|
|
|
|
|
char *error = NULL;
|
|
|
|
|
char *jobid_str = NULL;
|
|
|
|
|
char *procid_str = NULL;
|
|
|
|
|
char *contact_path = NULL;
|
|
|
|
|
|
2007-01-30 01:01:28 +03:00
|
|
|
if (orte_initialized) {
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
/* initialize the opal layer */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = opal_init())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* register handler for errnum -> string conversion */
|
|
|
|
|
opal_error_register("ORTE", ORTE_ERR_BASE, ORTE_ERR_MAX, orte_err2str);
|
|
|
|
|
|
|
|
|
|
/* Register all MCA Params */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_register_params(infrastructure))) {
|
|
|
|
|
error = "orte_register_params";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Ensure the system_info structure is instantiated and initialized */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_sys_info())) {
|
|
|
|
|
error = "orte_sys_info";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Ensure the process info structure is instantiated and initialized */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_proc_info())) {
|
|
|
|
|
error = "orte_proc_info";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Ensure the universe_info structure is instantiated and initialized */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_univ_info())) {
|
|
|
|
|
error = "orte_univ_info";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Initialize the data storage service.
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_dss_open())) {
|
|
|
|
|
error = "orte_dss_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Open the name services to ensure access to local functions
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_ns_base_open())) {
|
|
|
|
|
error = "orte_ns_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Open the error manager to activate error logging - needs local name services */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_errmgr_base_open())) {
|
|
|
|
|
error = "orte_errmgr_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/***** ERROR LOGGING NOW AVAILABLE *****/
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Internal startup
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_wait_init())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_wait_init";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Runtime Messaging Layer
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rml_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rml_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Runtime Messaging Layer
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rml_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rml_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Routed system
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_routed_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_routed_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Routed system
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_routed_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_routed_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Group communications
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_grpcomm_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_grpcomm_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_grpcomm_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_grpcomm_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Registry
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_gpr_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Initialize the daemon launch system so those types
|
|
|
|
|
* are registered (needed by the sds to talk to its
|
|
|
|
|
* local daemon)
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_odls_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_odls_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Initialize schema utilities
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_schema_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_schema_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Initialize and select the Startup Discovery Service. This must
|
|
|
|
|
* be done here since some environments have different requirements
|
|
|
|
|
* for detecting/connecting to a universe. Note that this does
|
|
|
|
|
* *not* set our name - that will come later
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_sds_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_sds_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_sds_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_sds_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Try to connect to the universe. If we don't find one and are a
|
|
|
|
|
* singleton, this will startup a new HNP and define our name
|
|
|
|
|
* within it - in which case, we will skip the name discovery
|
|
|
|
|
* process since we already have one
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_sds_base_contact_universe())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_sds_base_contact_universe";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Name Server
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_ns_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_ns_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Registry
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_gpr_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* set contact info for ns/gpr */
|
|
|
|
|
if(NULL != orte_process_info.ns_replica_uri) {
|
|
|
|
|
orte_rml.set_contact_info(orte_process_info.ns_replica_uri);
|
|
|
|
|
}
|
|
|
|
|
if(NULL != orte_process_info.gpr_replica_uri) {
|
|
|
|
|
orte_rml.set_contact_info(orte_process_info.gpr_replica_uri);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Set my name */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_sds_base_set_name())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_sds_base_set_name";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* all done with sds - clean up and call it a day */
|
|
|
|
|
orte_sds_base_close();
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Now that we know for certain if we are an HNP and/or a daemon,
|
|
|
|
|
* setup the resource management frameworks. This includes
|
|
|
|
|
* selecting the daemon launch framework - that framework "knows"
|
|
|
|
|
* what to do if it isn't in a daemon.
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rds_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rds_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rds_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rds_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_ras_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_ras_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_ras_base_find_available())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_ras_base_find_available";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rmaps_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rmaps_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rmaps_base_find_available())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rmaps_base_find_available";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_pls_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_pls_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_pls_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_pls_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_odls_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_odls_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rmgr_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rmgr_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rmgr_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rmgr_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* setup the state monitor
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_smr_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_smr_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_smr_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_smr_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* setup the errmgr -- open has been done way before
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_errmgr_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_errmgr_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* enable communication with the rml */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rml.enable_comm())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rml.enable_comm";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* if I'm the seed, set the seed uri to be me! */
|
|
|
|
|
if (orte_process_info.seed) {
|
|
|
|
|
if (NULL != orte_universe_info.seed_uri) {
|
|
|
|
|
free(orte_universe_info.seed_uri);
|
|
|
|
|
}
|
|
|
|
|
orte_universe_info.seed_uri = orte_rml.get_contact_info();
|
|
|
|
|
/* and make sure that the daemon flag is NOT set so that
|
|
|
|
|
* components unique to non-HNP orteds can be selected
|
|
|
|
|
*/
|
|
|
|
|
orte_process_info.daemon = false;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* setup my session directory */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_ns.get_jobid_string(&jobid_str, orte_process_info.my_name))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_ns.get_jobid_string";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_ns.get_vpid_string(&procid_str, orte_process_info.my_name))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_ns.get_vpid_string";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (orte_debug_flag) {
|
|
|
|
|
opal_output(0, "%s setting up session dir with",
|
|
|
|
|
ORTE_NAME_PRINT(orte_process_info.my_name));
|
|
|
|
|
if (NULL != orte_process_info.tmpdir_base) {
|
|
|
|
|
opal_output(0, "\ttmpdir %s", orte_process_info.tmpdir_base);
|
|
|
|
|
}
|
|
|
|
|
opal_output(0, "\tuniverse %s", orte_universe_info.name);
|
|
|
|
|
opal_output(0, "\tuser %s", orte_system_info.user);
|
|
|
|
|
opal_output(0, "\thost %s", orte_system_info.nodename);
|
|
|
|
|
opal_output(0, "\tjobid %s", jobid_str);
|
|
|
|
|
opal_output(0, "\tprocid %s", procid_str);
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_session_dir(true,
|
|
|
|
|
orte_process_info.tmpdir_base,
|
|
|
|
|
orte_system_info.user,
|
|
|
|
|
orte_system_info.nodename, NULL,
|
|
|
|
|
orte_universe_info.name,
|
|
|
|
|
jobid_str, procid_str))) {
|
|
|
|
|
if (jobid_str != NULL) free(jobid_str);
|
|
|
|
|
if (procid_str != NULL) free(procid_str);
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_session_dir";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (NULL != jobid_str) {
|
|
|
|
|
free(jobid_str);
|
|
|
|
|
}
|
|
|
|
|
if (NULL != procid_str) {
|
|
|
|
|
free(procid_str);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Once the session directory location has been established, set
|
|
|
|
|
the opal_output default file location to be in the
|
|
|
|
|
proc-specific session directory. */
|
|
|
|
|
opal_output_set_output_file_info(orte_process_info.proc_session_dir,
|
|
|
|
|
"output-", NULL, NULL);
|
|
|
|
|
|
|
|
|
|
/* if i'm the seed, get my contact info and write my setup file for others to find */
|
|
|
|
|
if (orte_process_info.seed) {
|
|
|
|
|
contact_path = opal_os_path(false, orte_process_info.universe_session_dir,
|
|
|
|
|
"universe-setup.txt", NULL);
|
|
|
|
|
if (orte_debug_flag) {
|
|
|
|
|
opal_output(0, "%s contact_file %s",
|
|
|
|
|
ORTE_NAME_PRINT(orte_process_info.my_name), contact_path);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_write_universe_setup_file(contact_path, &orte_universe_info))) {
|
|
|
|
|
if (orte_debug_flag) {
|
|
|
|
|
opal_output(0, "%s couldn't write setup file", ORTE_NAME_PRINT(orte_process_info.my_name));
|
|
|
|
|
}
|
|
|
|
|
} else if (orte_debug_flag) {
|
|
|
|
|
opal_output(0, "%s wrote setup file", ORTE_NAME_PRINT(orte_process_info.my_name));
|
|
|
|
|
}
|
|
|
|
|
free(contact_path);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Initialize the selected modules now that all components/name are available.
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_ns.init())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_ns.init";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.init())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_gpr.init";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* setup I/O forwarding system
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_iof_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_iof_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_iof_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_iof_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (orte_process_info.seed) {
|
|
|
|
|
/* if we are an HNP, we have to setup an app_context for ourselves so
|
|
|
|
|
* various frameworks can find their required info
|
|
|
|
|
*/
|
|
|
|
|
orte_app_context_t *app;
|
|
|
|
|
orte_gpr_value_t *value;
|
|
|
|
|
|
|
|
|
|
app = OBJ_NEW(orte_app_context_t);
|
|
|
|
|
app->app = strdup("orted");
|
|
|
|
|
app->num_procs = 1;
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rmgr.store_app_context(ORTE_PROC_MY_NAME->jobid, &app, 1))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "HNP store app context";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
OBJ_RELEASE(app);
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_value(&value,
|
|
|
|
|
ORTE_GPR_OVERWRITE|ORTE_GPR_TOKENS_AND,
|
|
|
|
|
"orte-job-0", 2,
|
|
|
|
|
0))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "HNP create value";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* store the node name and the daemon's name */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(value->keyvals[0]), ORTE_NODE_NAME_KEY,
|
|
|
|
|
ORTE_STRING, orte_system_info.nodename))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "HNP create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(value->keyvals[1]), ORTE_PROC_NAME_KEY,
|
|
|
|
|
ORTE_NAME, ORTE_PROC_MY_NAME))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "HNP create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* set the tokens */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_schema.get_proc_tokens(&(value->tokens), &(value->num_tokens), ORTE_PROC_MY_NAME))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "HNP get proc tokens";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* insert values */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.put(1, &value))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "HNP put values";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
OBJ_RELEASE(value);
|
|
|
|
|
|
|
|
|
|
/* and set our state to LAUNCHED */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_smr.set_proc_state(ORTE_PROC_MY_NAME, ORTE_PROC_STATE_LAUNCHED, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "HNP could not set launched state";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (orte_process_info.singleton) {
|
|
|
|
|
/* since all frameworks are now open and active, walk through
|
|
|
|
|
* the spawn sequence to ensure that we collect info on all
|
|
|
|
|
* available resources. Although we are a singleton and hence
|
|
|
|
|
* don't need to be spawned, we may choose to dynamically spawn
|
|
|
|
|
* additional processes. If we do that, then we need to know
|
|
|
|
|
* about any resources that have been allocated to us - executing
|
|
|
|
|
* the RDS and RAS frameworks is the only way to get that info.
|
|
|
|
|
*
|
|
|
|
|
* THIS ONLY SHOULD BE DONE FOR SINGLETONS - DO NOT DO IT
|
|
|
|
|
* FOR ANY OTHER CASE
|
|
|
|
|
*/
|
|
|
|
|
orte_gpr_value_t *values[2];
|
|
|
|
|
orte_std_cntr_t one=1, zero=0;
|
|
|
|
|
orte_proc_state_t init=ORTE_PROC_STATE_INIT;
|
|
|
|
|
orte_vpid_t lrank=0, vone=1;
|
|
|
|
|
char *segment;
|
|
|
|
|
|
|
|
|
|
/* ensure we read the allocation, even if we are not sitting on a node that is within it */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_ras.allocate_job(ORTE_PROC_MY_NAME->jobid, NULL))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not allocate job";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* let's deal with the procs first - start by getting their job segment name */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_schema.get_job_segment_name(&segment, ORTE_PROC_MY_NAME->jobid))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create job segment name";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_value(&(values[0]),
|
|
|
|
|
ORTE_GPR_OVERWRITE|ORTE_GPR_TOKENS_AND,
|
|
|
|
|
segment, 7, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create gpr value";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[0]->keyvals[0]), ORTE_PROC_RANK_KEY, ORTE_VPID, &(ORTE_PROC_MY_NAME->vpid)))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[0]->keyvals[1]), ORTE_PROC_NAME_KEY, ORTE_NAME, ORTE_PROC_MY_NAME))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[0]->keyvals[2]), ORTE_NODE_NAME_KEY, ORTE_STRING, orte_system_info.nodename))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[0]->keyvals[3]), ORTE_PROC_APP_CONTEXT_KEY, ORTE_STD_CNTR, &zero))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[0]->keyvals[4]), ORTE_PROC_STATE_KEY, ORTE_PROC_STATE, &init))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[0]->keyvals[5]), ORTE_PROC_LOCAL_RANK_KEY, ORTE_VPID, &lrank))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[0]->keyvals[6]), ORTE_NODE_NUM_PROCS_KEY, ORTE_STD_CNTR, &one))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_schema.get_proc_tokens(&(values[0]->tokens), &(values[0]->num_tokens), ORTE_PROC_MY_NAME))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not get gpr tokens";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_value(&values[1],
|
|
|
|
|
ORTE_GPR_OVERWRITE|ORTE_GPR_TOKENS_AND,
|
|
|
|
|
segment, 3, 1))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create gpr value";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[1]->keyvals[0]), ORTE_PROC_NUM_AT_INIT, ORTE_STD_CNTR, &one))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[1]->keyvals[1]), ORTE_JOB_VPID_START_KEY, ORTE_VPID, &(ORTE_PROC_MY_NAME->vpid)))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.create_keyval(&(values[1]->keyvals[2]), ORTE_JOB_VPID_RANGE_KEY, ORTE_VPID, &vone))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not create keyval";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
values[1]->tokens[0] = strdup(ORTE_JOB_GLOBALS); /* counter is in the job's globals container */
|
|
|
|
|
|
|
|
|
|
/* insert all values in one call */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_gpr.put(2, values))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not put its launch data";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
OBJ_RELEASE(values[0]);
|
|
|
|
|
OBJ_RELEASE(values[1]);
|
|
|
|
|
free(segment);
|
|
|
|
|
|
|
|
|
|
/* wireup our io */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_iof.iof_pull(ORTE_PROC_MY_NAME, ORTE_NS_CMP_JOBID, ORTE_IOF_STDOUT, 1))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not setup iof";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_iof.iof_pull(ORTE_PROC_MY_NAME, ORTE_NS_CMP_JOBID, ORTE_IOF_STDERR, 2))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not setup iof";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_iof.iof_push(ORTE_PROC_MY_NAME, ORTE_NS_CMP_JOBID, ORTE_IOF_STDIN, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not setup iof";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* setup the errmgr, as required */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_errmgr.register_job(ORTE_PROC_MY_NAME->jobid))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not setup errmgr callbacks";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* set our state to LAUNCHED */
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_smr.set_proc_state(ORTE_PROC_MY_NAME, ORTE_PROC_STATE_LAUNCHED, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "singleton could not set launched state";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* setup the routed info - the selected routed component
|
|
|
|
|
* will know what to do. Some may put us in a blocking
|
|
|
|
|
* receive here so they can get ALL of the contact info
|
|
|
|
|
* from our peers. Others may just find the local daemon's
|
|
|
|
|
* contact info and immediately return.
|
|
|
|
|
*/
|
|
|
|
|
if( !orte_universe_info.console ) {
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_routed.init_routes(ORTE_PROC_MY_NAME->jobid, NULL))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_routed.init_routes";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
2005-03-14 23:57:21 +03:00
|
|
|
}
|
|
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
/*
|
|
|
|
|
* Setup the FileM
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_filem_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_filem_base_open";
|
|
|
|
|
goto error;
|
2005-03-14 23:57:21 +03:00
|
|
|
}
|
2005-09-03 01:07:21 +04:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
if (ORTE_SUCCESS != (ret = orte_filem_base_select())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_filem_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#if OPAL_ENABLE_FT == 1
|
|
|
|
|
/*
|
|
|
|
|
* Setup the SnapC
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_snapc_base_open())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_snapc_base_open";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_snapc_base_select(orte_process_info.seed, !orte_process_info.daemon))) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_snapc_base_select";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Need to figure out if we are an application or part of ORTE */
|
|
|
|
|
if(infrastructure ||
|
|
|
|
|
orte_process_info.seed ||
|
|
|
|
|
orte_process_info.daemon) {
|
|
|
|
|
/* ORTE doesn't need the OPAL CR stuff */
|
|
|
|
|
opal_cr_set_enabled(false);
|
|
|
|
|
}
|
|
|
|
|
else {
|
|
|
|
|
/* The application does however */
|
|
|
|
|
opal_cr_set_enabled(true);
|
|
|
|
|
}
|
|
|
|
|
#else
|
|
|
|
|
opal_cr_set_enabled(false);
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Initalize the CR setup
|
|
|
|
|
* Note: Always do this, even in non-FT builds.
|
|
|
|
|
* If we don't some user level tools may hang.
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_cr_init())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_cr_init";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Since we are now finished with init, change the state to running */
|
|
|
|
|
orte_universe_info.state = ORTE_UNIVERSE_STATE_RUNNING;
|
|
|
|
|
|
|
|
|
|
/* startup the receive if we are not the HNP - unless we are a singleton,
|
|
|
|
|
* in which case we must start it up in case we do a comm_spawn!
|
|
|
|
|
*/
|
|
|
|
|
if (orte_process_info.singleton ||
|
|
|
|
|
(!orte_process_info.seed && orte_process_info.daemon)) {
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_rml_base_comm_start())) {
|
|
|
|
|
ORTE_ERROR_LOG(ret);
|
|
|
|
|
error = "orte_rml_base_comm_start";
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* All done */
|
|
|
|
|
orte_initialized = true;
|
2005-03-14 23:57:21 +03:00
|
|
|
return ORTE_SUCCESS;
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
|
|
|
|
|
|
|
|
error:
|
|
|
|
|
if (ret != ORTE_SUCCESS) {
|
|
|
|
|
opal_show_help("help-orte-runtime",
|
|
|
|
|
"orte_init:startup:internal-failure",
|
|
|
|
|
true, error, ORTE_ERROR_NAME(ret), ret);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return ret;
|
2005-03-14 23:57:21 +03:00
|
|
|
}
|