2004-11-20 22:12:43 +03:00
|
|
|
/*
|
2010-03-24 00:28:02 +03:00
|
|
|
* Copyright (c) 2004-2010 The Trustees of Indiana University and Indiana
|
2005-11-05 22:57:48 +03:00
|
|
|
* University Research and Technology
|
|
|
|
|
* Corporation. All rights reserved.
|
2006-08-23 07:32:36 +04:00
|
|
|
* Copyright (c) 2004-2006 The University of Tennessee and The University
|
2005-11-05 22:57:48 +03:00
|
|
|
* of Tennessee Research Foundation. All rights
|
|
|
|
|
* reserved.
|
2004-11-28 23:09:25 +03:00
|
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
|
|
|
|
* 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.
|
2007-10-15 19:41:36 +04:00
|
|
|
* Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
|
2004-11-22 04:38:40 +03:00
|
|
|
* $COPYRIGHT$
|
|
|
|
|
*
|
|
|
|
|
* Additional copyrights may follow
|
|
|
|
|
*
|
2004-11-20 22:12:43 +03:00
|
|
|
* $HEADER$
|
|
|
|
|
*/
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
#ifndef MCA_RML_OOB_RML_OOB_H
|
|
|
|
|
#define MCA_RML_OOB_RML_OOB_H
|
|
|
|
|
|
2009-03-04 18:35:54 +03:00
|
|
|
#include "orte_config.h"
|
2007-07-20 05:34:02 +04:00
|
|
|
#include "opal/threads/condition.h"
|
|
|
|
|
#include "opal/threads/mutex.h"
|
Update libevent to the 2.0 series, currently at 2.0.7rc. We will update to their final release when it becomes available. Currently known errors exist in unused portions of the libevent code. This revision passes the IBM test suite on a Linux machine and on a standalone Mac.
This is a fairly intrusive change, but outside of the moving of opal/event to opal/mca/event, the only changes involved (a) changing all calls to opal_event functions to reflect the new framework instead, and (b) ensuring that all opal_event_t objects are properly constructed since they are now true opal_objects.
Note: Shiqing has just returned from vacation and has not yet had a chance to complete the Windows integration. Thus, this commit almost certainly breaks Windows support on the trunk. However, I want this to have a chance to soak for as long as possible before I become less available a week from today (going to be at a class for 5 days, and thus will only be sparingly available) so we can find and fix any problems.
Biggest change is moving the libevent code from opal/event to a new opal/mca/event framework. This was done to make it much easier to update libevent in the future. New versions can be inserted as a new component and tested in parallel with the current version until validated, then we can remove the earlier version if we so choose. This is a statically built framework ala installdirs, so only one component will build at a time. There is no selection logic - the sole compiled component simply loads its function pointers into the opal_event struct.
I have gone thru the code base and converted all the libevent calls I could find. However, I cannot compile nor test every environment. It is therefore quite likely that errors remain in the system. Please keep an eye open for two things:
1. compile-time errors: these will be obvious as calls to the old functions (e.g., opal_evtimer_new) must be replaced by the new framework APIs (e.g., opal_event.evtimer_new)
2. run-time errors: these will likely show up as segfaults due to missing constructors on opal_event_t objects. It appears that it became a typical practice for people to "init" an opal_event_t by simply using memset to zero it out. This will no longer work - you must either OBJ_NEW or OBJ_CONSTRUCT an opal_event_t. I tried to catch these cases, but may have missed some. Believe me, you'll know when you hit it.
There is also the issue of the new libevent "no recursion" behavior. As I described on a recent email, we will have to discuss this and figure out what, if anything, we need to do.
This commit was SVN r23925.
2010-10-24 22:35:54 +04:00
|
|
|
#include "opal/mca/event/event.h"
|
2006-02-12 04:33:29 +03:00
|
|
|
#include "orte/mca/rml/rml.h"
|
2007-07-20 05:34:02 +04:00
|
|
|
#include "orte/mca/oob/oob.h"
|
2008-02-28 04:57:57 +03:00
|
|
|
#include "opal/dss/dss_types.h"
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
BEGIN_C_DECLS
|
|
|
|
|
|
|
|
|
|
struct orte_rml_oob_module_t {
|
|
|
|
|
struct orte_rml_module_t super;
|
|
|
|
|
mca_oob_t *active_oob;
|
|
|
|
|
opal_list_t exceptions;
|
|
|
|
|
opal_mutex_t exceptions_lock;
|
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
|
|
|
opal_list_t queued_routing_messages;
|
|
|
|
|
opal_mutex_t queued_lock;
|
2007-10-15 19:41:36 +04:00
|
|
|
opal_event_t *timer_event;
|
|
|
|
|
struct timeval timeout;
|
2007-07-20 05:34:02 +04:00
|
|
|
};
|
|
|
|
|
typedef struct orte_rml_oob_module_t orte_rml_oob_module_t;
|
2004-11-20 22:12:43 +03:00
|
|
|
|
2006-08-23 07:32:36 +04:00
|
|
|
ORTE_MODULE_DECLSPEC extern orte_rml_component_t mca_rml_oob_component;
|
2007-07-20 05:34:02 +04:00
|
|
|
extern orte_rml_oob_module_t orte_rml_oob_module;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
typedef enum {
|
|
|
|
|
ORTE_RML_BLOCKING_SEND,
|
|
|
|
|
ORTE_RML_NONBLOCKING_IOV_SEND,
|
|
|
|
|
ORTE_RML_NONBLOCKING_BUFFER_SEND,
|
|
|
|
|
ORTE_RML_BLOCKING_RECV,
|
|
|
|
|
ORTE_RML_NONBLOCKING_IOV_RECV,
|
|
|
|
|
ORTE_RML_NONBLOCKING_BUFFER_RECV
|
|
|
|
|
} orte_rml_oob_msg_type_t;
|
|
|
|
|
|
|
|
|
|
struct orte_rml_oob_msg_header_t {
|
|
|
|
|
orte_process_name_t origin;
|
|
|
|
|
orte_process_name_t destination;
|
|
|
|
|
int tag;
|
|
|
|
|
};
|
|
|
|
|
typedef struct orte_rml_oob_msg_header_t orte_rml_oob_msg_header_t;
|
|
|
|
|
|
2007-08-29 00:54:52 +04:00
|
|
|
#define ORTE_RML_OOB_MSG_HEADER_HTON(hdr) \
|
|
|
|
|
do { \
|
|
|
|
|
ORTE_PROCESS_NAME_HTON((hdr).origin); \
|
|
|
|
|
ORTE_PROCESS_NAME_HTON((hdr).destination); \
|
|
|
|
|
(hdr).tag = htonl((hdr).tag); \
|
|
|
|
|
} while (0)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#define ORTE_RML_OOB_MSG_HEADER_NTOH(hdr) \
|
|
|
|
|
do { \
|
|
|
|
|
ORTE_PROCESS_NAME_NTOH((hdr).origin); \
|
|
|
|
|
ORTE_PROCESS_NAME_NTOH((hdr).destination); \
|
|
|
|
|
(hdr).tag = ntohl((hdr).tag); \
|
|
|
|
|
} while (0)
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
struct orte_rml_oob_msg_t {
|
|
|
|
|
opal_object_t super;
|
|
|
|
|
|
|
|
|
|
opal_mutex_t msg_lock;
|
|
|
|
|
opal_condition_t msg_cond;
|
|
|
|
|
|
|
|
|
|
orte_rml_oob_msg_type_t msg_type;
|
|
|
|
|
int msg_status;
|
|
|
|
|
volatile bool msg_complete;
|
|
|
|
|
bool msg_persistent;
|
2004-11-20 22:12:43 +03:00
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
union {
|
|
|
|
|
orte_rml_callback_fn_t iov;
|
|
|
|
|
orte_rml_buffer_callback_fn_t buffer;
|
|
|
|
|
} msg_cbfunc;
|
|
|
|
|
void *msg_cbdata;
|
|
|
|
|
|
|
|
|
|
struct iovec *msg_data;
|
|
|
|
|
|
|
|
|
|
/** buffer for non-blocking buffer sends */
|
2008-02-28 04:57:57 +03:00
|
|
|
opal_buffer_t msg_recv_buffer;
|
2007-07-20 05:34:02 +04:00
|
|
|
/** pointer to user buffer for buffered sends */
|
2008-02-28 04:57:57 +03:00
|
|
|
opal_buffer_t *user_buffer;
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
orte_rml_oob_msg_header_t msg_header;
|
|
|
|
|
};
|
|
|
|
|
typedef struct orte_rml_oob_msg_t orte_rml_oob_msg_t;
|
|
|
|
|
OBJ_CLASS_DECLARATION(orte_rml_oob_msg_t);
|
|
|
|
|
|
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
|
|
|
struct orte_rml_oob_queued_msg_t {
|
|
|
|
|
opal_list_item_t super;
|
|
|
|
|
struct iovec payload[1];
|
|
|
|
|
};
|
|
|
|
|
typedef struct orte_rml_oob_queued_msg_t orte_rml_oob_queued_msg_t;
|
|
|
|
|
OBJ_CLASS_DECLARATION(orte_rml_oob_queued_msg_t);
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
int orte_rml_oob_init(void);
|
|
|
|
|
int orte_rml_oob_fini(void);
|
2007-03-17 02:11:45 +03:00
|
|
|
int orte_rml_oob_ft_event(int state);
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
int orte_rml_oob_send(orte_process_name_t* peer,
|
|
|
|
|
struct iovec *msg,
|
|
|
|
|
int count,
|
|
|
|
|
int tag,
|
|
|
|
|
int flags);
|
|
|
|
|
int orte_rml_oob_send_nb(orte_process_name_t* peer,
|
|
|
|
|
struct iovec* msg,
|
|
|
|
|
int count,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags,
|
|
|
|
|
orte_rml_callback_fn_t cbfunc,
|
|
|
|
|
void* cbdata);
|
|
|
|
|
int orte_rml_oob_send_buffer(orte_process_name_t* peer,
|
2008-02-28 04:57:57 +03:00
|
|
|
opal_buffer_t* buffer,
|
2007-07-20 05:34:02 +04:00
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags);
|
|
|
|
|
|
|
|
|
|
int orte_rml_oob_send_buffer_nb(orte_process_name_t* peer,
|
2008-02-28 04:57:57 +03:00
|
|
|
opal_buffer_t* buffer,
|
2007-07-20 05:34:02 +04:00
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags,
|
|
|
|
|
orte_rml_buffer_callback_fn_t cbfunc,
|
|
|
|
|
void* cbdata);
|
|
|
|
|
|
|
|
|
|
int orte_rml_oob_recv(orte_process_name_t* peer,
|
|
|
|
|
struct iovec *msg,
|
|
|
|
|
int count,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags);
|
|
|
|
|
int orte_rml_oob_recv_nb(orte_process_name_t* peer,
|
|
|
|
|
struct iovec* msg,
|
|
|
|
|
int count,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags,
|
|
|
|
|
orte_rml_callback_fn_t cbfunc,
|
|
|
|
|
void* cbdata);
|
|
|
|
|
int orte_rml_oob_recv_buffer(orte_process_name_t* peer,
|
2008-02-28 04:57:57 +03:00
|
|
|
opal_buffer_t *buf,
|
2007-07-20 05:34:02 +04:00
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags);
|
|
|
|
|
int orte_rml_oob_recv_buffer_nb(orte_process_name_t* peer,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags,
|
|
|
|
|
orte_rml_buffer_callback_fn_t cbfunc,
|
|
|
|
|
void* cbdata);
|
|
|
|
|
int orte_rml_oob_recv_cancel(orte_process_name_t* peer,
|
|
|
|
|
orte_rml_tag_t tag);
|
|
|
|
|
|
|
|
|
|
int orte_rml_oob_ping(const char* uri,
|
|
|
|
|
const struct timeval* tv);
|
|
|
|
|
|
|
|
|
|
char* orte_rml_oob_get_uri(void);
|
|
|
|
|
int orte_rml_oob_set_uri(const char*);
|
|
|
|
|
|
|
|
|
|
int orte_rml_oob_add_exception(orte_rml_exception_callback_t cbfunc);
|
|
|
|
|
int orte_rml_oob_del_exception(orte_rml_exception_callback_t cbfunc);
|
|
|
|
|
void orte_rml_oob_exception_callback(const orte_process_name_t *peer,
|
|
|
|
|
orte_rml_exception_t exception);
|
|
|
|
|
|
|
|
|
|
|
2010-03-24 00:28:02 +03:00
|
|
|
int orte_rml_oob_purge(orte_process_name_t *peer);
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
END_C_DECLS
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
#endif
|