1
1
openmpi/opal/mca/btl/usnic/btl_usnic_stats.c

540 строки
19 KiB
C
Исходник Обычный вид История

Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
/*
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
#include "opal_config.h"
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
#include <unistd.h>
#include <stdlib.h>
#include "opal/util/output.h"
#include "opal/mca/base/mca_base_var.h"
#include "opal/mca/base/mca_base_pvar.h"
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
#include "opal/util/proc.h"
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
#include "btl_usnic.h"
#include "btl_usnic_module.h"
#include "btl_usnic_stats.h"
/*
* Local variables
*/
static mca_base_var_type_t pvar_type = MCA_BASE_VAR_TYPE_MAX;
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
static inline void usnic_stats_reset(opal_btl_usnic_module_t *module)
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
{
int i;
module->stats.num_total_sends =
module->stats.num_resends =
module->stats.num_chunk_sends =
module->stats.num_frag_sends =
module->stats.num_ack_recvs =
module->stats.num_total_recvs =
module->stats.num_unk_recvs =
module->stats.num_dup_recvs =
module->stats.num_oow_low_recvs =
module->stats.num_oow_high_recvs =
module->stats.num_frag_recvs =
module->stats.num_chunk_recvs =
module->stats.num_badfrag_recvs =
module->stats.num_ack_sends =
module->stats.num_recv_reposts =
module->stats.num_crc_errors =
module->stats.num_old_dup_acks =
module->stats.num_dup_acks =
module->stats.num_fast_retrans =
module->stats.num_timeout_retrans =
module->stats.max_sent_window_size =
module->stats.max_rcvd_window_size =
module->stats.pml_module_sends =
module->stats.pml_send_callbacks =
0;
for (i=0; i<USNIC_NUM_CHANNELS; ++i) {
module->mod_channels[i].num_channel_sends = 0;
}
}
/* Prints a few terse statistics lines via opal_output(0,...). The first
* line will be prefixed with the string "prefix". If "reset_stats" is true
* then the statistics will be reset after printing.
*
* NOTE: this routine ignores the setting of stats_enable, so it can be used
* for debugging routines even when normal stats reporting is not enabled.
*/
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
void opal_btl_usnic_print_stats(
opal_btl_usnic_module_t *module,
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
const char *prefix,
bool reset_stats)
{
char tmp[128], str[2048];
/* The usuals */
snprintf(str, sizeof(str), "%s:MCW:%3u, ST(P+D)/F/C/R(T+F)/A:%8lu(%8u+%8u)/%8lu/%8lu/%4lu(%4lu+%4lu)/%8lu, RcvTot/Chk/F/C/L/H/D/BF/A:%8lu/%c%c/%8lu/%8lu/%4lu+%2lu/%4lu/%4lu/%6lu OA/DA %4lu/%4lu CRC:%4lu ",
prefix,
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
((orte_process_name_t*)&opal_proc_local_get()->proc_name)->vpid,
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
module->stats.num_total_sends,
module->mod_channels[USNIC_PRIORITY_CHANNEL].num_channel_sends,
module->mod_channels[USNIC_DATA_CHANNEL].num_channel_sends,
module->stats.num_frag_sends,
module->stats.num_chunk_sends,
module->stats.num_resends,
module->stats.num_timeout_retrans,
module->stats.num_fast_retrans,
module->stats.num_ack_sends,
module->stats.num_total_recvs,
(module->stats.num_total_recvs -
module->stats.num_recv_reposts) == 0 ? 'g' : 'B',
(module->stats.num_total_recvs -
module->stats.num_frag_recvs -
module->stats.num_chunk_recvs -
module->stats.num_badfrag_recvs -
module->stats.num_oow_low_recvs -
module->stats.num_oow_high_recvs -
module->stats.num_dup_recvs -
module->stats.num_ack_recvs -
module->stats.num_unk_recvs) == 0 ? 'g' : 'B',
module->stats.num_frag_recvs,
module->stats.num_chunk_recvs,
module->stats.num_oow_low_recvs,
module->stats.num_oow_high_recvs,
module->stats.num_dup_recvs,
module->stats.num_badfrag_recvs,
module->stats.num_ack_recvs,
module->stats.num_old_dup_acks,
module->stats.num_dup_acks,
module->stats.num_crc_errors);
/* If our PML calls were 0, then show send and receive window
extents instead */
if (module->stats.pml_module_sends +
module->stats.pml_send_callbacks == 0) {
int64_t send_unacked, su_min = WINDOW_SIZE * 2, su_max = 0;
int64_t recv_depth, rd_min = WINDOW_SIZE * 2, rd_max = 0;
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
opal_btl_usnic_endpoint_t *endpoint;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
opal_list_item_t *item;
rd_min = su_min = WINDOW_SIZE * 2;
rd_max = su_max = 0;
item = opal_list_get_first(&module->all_endpoints);
while (item != opal_list_get_end(&(module->all_endpoints))) {
endpoint = container_of(item, mca_btl_base_endpoint_t,
endpoint_endpoint_li);
item = opal_list_get_next(item);
/* Number of un-acked sends (i.e., sends for which we're
still waiting for ACK) */
send_unacked =
endpoint->endpoint_next_seq_to_send -
endpoint->endpoint_ack_seq_rcvd - 1;
if (send_unacked > su_max) su_max = send_unacked;
if (send_unacked < su_min) su_min = send_unacked;
/* Receive window depth (i.e., difference between highest
seq received and the next message we haven't ACKed
yet) */
recv_depth =
endpoint->endpoint_highest_seq_rcvd -
endpoint->endpoint_next_contig_seq_to_recv;
if (recv_depth > rd_max) rd_max = recv_depth;
if (recv_depth < rd_min) rd_min = recv_depth;
}
snprintf(tmp, sizeof(tmp), "PML S:%1ld, Win!A/R:%4ld/%4ld %4ld/%4ld",
module->stats.pml_module_sends,
su_min, su_max,
rd_min, rd_max);
} else {
snprintf(tmp, sizeof(tmp), "PML S/CB/Diff:%4lu/%4lu=%4ld",
module->stats.pml_module_sends,
module->stats.pml_send_callbacks,
module->stats.pml_module_sends -
module->stats.pml_send_callbacks);
}
strncat(str, tmp, sizeof(str) - strlen(str) - 1);
opal_output(0, "%s", str);
if (reset_stats) {
usnic_stats_reset(module);
}
}
/*
* Callback routine for libevent
*/
static void usnic_stats_callback(int fd, short flags, void *arg)
{
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
opal_btl_usnic_module_t *module = (opal_btl_usnic_module_t*) arg;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
char tmp[128];
if (!mca_btl_usnic_component.stats_enabled) {
return;
}
snprintf(tmp, sizeof(tmp), "%4lu", ++module->stats.report_num);
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
opal_btl_usnic_print_stats(module, tmp,
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
/*reset=*/mca_btl_usnic_component.stats_relative);
/* In OMPI v1.6, we have to re-add this event (because there's an
old libevent in OMPI v1.6) */
opal_event_add(&(module->stats.timer_event),
&(module->stats.timeout));
}
/*
* Initialize usnic module statistics
*/
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
int opal_btl_usnic_stats_init(opal_btl_usnic_module_t *module)
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
{
if (mca_btl_usnic_component.stats_enabled) {
usnic_stats_reset(module);
module->stats.timeout.tv_sec = mca_btl_usnic_component.stats_frequency;
module->stats.timeout.tv_usec = 0;
opal_event_set(opal_event_base, &(module->stats.timer_event),
-1, EV_TIMEOUT | EV_PERSIST,
&usnic_stats_callback, module);
opal_event_add(&(module->stats.timer_event),
&(module->stats.timeout));
}
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
return OPAL_SUCCESS;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
}
/*
* Finalize usnic module statistics
*/
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
int opal_btl_usnic_stats_finalize(opal_btl_usnic_module_t *module)
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
{
/* Disable the stats callback event, and then call the stats
callback manually to display the final stats */
if (mca_btl_usnic_component.stats_enabled) {
opal_event_del(&(module->stats.timer_event));
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
opal_btl_usnic_print_stats(module, "final", /*reset_stats=*/false);
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
}
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
return OPAL_SUCCESS;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
}
/************************************************************************/
/*
* Function called by the pvar base upon MPI_T_pvar_handle_alloc,
* handle_start, and handle_stop.
*/
static int usnic_pvar_notify(struct mca_base_pvar_t *pvar,
mca_base_pvar_event_t event,
void *obj, int *count)
{
if (MCA_BASE_PVAR_HANDLE_BIND == event) {
*count = mca_btl_usnic_component.num_modules;
}
/* Don't care about the other events */
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
return OPAL_SUCCESS;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
}
/*
* Function called by the pvar base when a user wants to read the
* value of an MPI_T performance variable.
*/
static int usnic_pvar_read(const struct mca_base_pvar_t *pvar,
void *value, void *bound_obj)
{
size_t i;
size_t offset = (size_t) pvar->ctx;
uint64_t *array = (uint64_t*) value;
for (i = 0; i < mca_btl_usnic_component.num_modules; ++i) {
char *base = (char*) &(mca_btl_usnic_component.usnic_active_modules[i]->stats);
array[i] = *((uint64_t*) (base + offset));
}
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
return OPAL_SUCCESS;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
}
/*
* Register an MPI_T performance variable of type CLASS_HIGHWATERMARK.
*/
static void register_pvar_highwater(char *name, char *desc, size_t offset)
{
int rc __opal_attribute_unused__;
rc = mca_base_component_pvar_register(&mca_btl_usnic_component.super.btl_version,
name, desc,
OPAL_INFO_LVL_5,
MCA_BASE_PVAR_CLASS_HIGHWATERMARK,
pvar_type,
NULL, /* enumeration */
MCA_BASE_VAR_BIND_NO_OBJECT,
(MCA_BASE_PVAR_FLAG_READONLY |
MCA_BASE_PVAR_FLAG_CONTINUOUS),
usnic_pvar_read,
NULL, /* write function */
usnic_pvar_notify,
(void *) offset);
assert(rc >= 0);
}
/*
* Function called by the pvar base when a user wants to read the
* devices enum value. The array is a simple list of 0..num_modules,
* which will map to the strings in the devices_enum
* setup_mpit_pvar_type().
*/
static int usnic_pvar_enum_read(const struct mca_base_pvar_t *pvar,
void *value, void *bound_obj)
{
size_t i;
int *array = (int *) value;
for (i = 0; i < mca_btl_usnic_component.num_modules; ++i) {
array[i] = i;
}
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
return OPAL_SUCCESS;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
}
/*
* Register an MPI_T performance variable of type CLASS_COUNTER.
*/
static void register_pvar_counter(char *name, char *desc, size_t offset)
{
int rc __opal_attribute_unused__;
rc = mca_base_component_pvar_register(&mca_btl_usnic_component.super.btl_version,
name, desc,
OPAL_INFO_LVL_5,
MCA_BASE_PVAR_CLASS_COUNTER,
pvar_type,
NULL, /* enumeration */
MCA_BASE_VAR_BIND_NO_OBJECT,
(MCA_BASE_PVAR_FLAG_READONLY |
MCA_BASE_PVAR_FLAG_CONTINUOUS),
usnic_pvar_read,
NULL, /* write function */
usnic_pvar_notify,
(void *) offset);
assert(rc >= 0);
}
/*
* Find the MPI_T type corresponding to our uint64_t counters and
* highwatermarks.
*/
static bool setup_mpit_pvar_type(void)
{
/* Our stats variables are uint64_t's, so find a pvar type that is
compatible */
if (sizeof(uint64_t) == sizeof(unsigned int)) {
pvar_type = MCA_BASE_VAR_TYPE_UNSIGNED_INT;
} else if (sizeof(uint64_t) == sizeof(unsigned long)) {
pvar_type = MCA_BASE_VAR_TYPE_UNSIGNED_LONG;
#ifdef HAVE_UNSIGNED_LONG_LONG
} else if (sizeof(uint64_t) == sizeof(unsigned long long)) {
pvar_type = MCA_BASE_VAR_TYPE_UNSIGNED_LONG_LONG;
#endif
}
/* Let the caller know if we found a compatible type or not */
if (MCA_BASE_VAR_TYPE_MAX == pvar_type) {
return false;
}
return true;
}
/*
* Setup the usnic_X device enumeration pvar
*/
static void setup_mpit_pvars_enum(void)
{
size_t i;
int rc __opal_attribute_unused__;
mca_base_var_enum_value_t *devices;
static mca_base_var_enum_t *devices_enum;
struct ibv_device *device;
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
opal_btl_usnic_module_t *m;
unsigned char *c;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
devices = calloc(mca_btl_usnic_component.num_modules + 1,
sizeof(*devices));
assert(devices != NULL);
for (i = 0; i < mca_btl_usnic_component.num_modules; ++i) {
char *str;
m = mca_btl_usnic_component.usnic_active_modules[i];
c = (unsigned char*) &m->if_ipv4_addr;
device = m->device;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
devices[i].value = i;
rc = asprintf(&str, "%s,%s,%hhu.%hhu.%hhu.%hhu/%" PRIu32,
ibv_get_device_name(device),
m->if_name,
c[0], c[1], c[2], c[3],
m->if_cidrmask);
assert(rc > 0);
devices[i].string = str;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
}
devices[i].string = NULL;
rc = mca_base_var_enum_create("btl_usnic", devices, &devices_enum);
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
assert(OPAL_SUCCESS == rc);
rc = mca_base_component_pvar_register(&mca_btl_usnic_component.super.btl_version,
"devices",
"Enumeration representing which slot in btl_usnic_* MPI_T pvar value arrays correspond to which usnic_X Linux device",
OPAL_INFO_LVL_5,
MCA_BASE_PVAR_CLASS_STATE,
MCA_BASE_VAR_TYPE_INT,
devices_enum,
MCA_BASE_VAR_BIND_NO_OBJECT,
(MCA_BASE_PVAR_FLAG_READONLY |
MCA_BASE_PVAR_FLAG_CONTINUOUS),
usnic_pvar_enum_read,
NULL, /* write function */
usnic_pvar_notify,
NULL /* context */);
assert(rc >= 0);
/* Free the strings (mca_base_var_enum_create() strdup()'ed them
into private storage, so we don't need them any more) */
for (i = 0; i < mca_btl_usnic_component.num_modules; ++i) {
free((char*) devices[i].string);
}
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
/* The devices_enum has been RETAIN'ed by the pvar, so we can
RELEASE it here, and the enum will be destroyed when the pvar
is destroyed. */
OBJ_RELEASE(devices_enum);
}
/*
* Setup high watermark MPI_T performance variables
*/
static void setup_mpit_pvars_highwatermark(void)
{
#define REGISTERHW(field, desc) \
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
register_pvar_highwater(#field, (desc), offsetof(opal_btl_usnic_module_stats_t, field))
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
REGISTERHW(max_sent_window_size,
"Maximum number of entries in all send windows from this peer");
REGISTERHW(max_rcvd_window_size,
"Maximum number of entries in all receive windows to this peer");
}
/*
* Setup counter MPI_T performance variables
*/
static void setup_mpit_pvars_counters(void)
{
#define REGISTERC(field, desc) \
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
register_pvar_counter(#field, (desc), offsetof(opal_btl_usnic_module_stats_t, field))
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
REGISTERC(num_total_sends,
"Total number of sends (MPI data, ACKs, retransmissions, etc.)");
REGISTERC(num_resends,
"Total number of all retransmissions");
REGISTERC(num_timeout_retrans,
"Number of times chunk retransmissions have occured because an ACK was not received within the timeout");
REGISTERC(num_fast_retrans,
"Number of times chunk retransmissions have occured because due to a repeated ACK");
REGISTERC(num_chunk_sends,
"Number of sends that were part of a larger MPI message fragment (i.e., the MPI message was so long that it had to be split into multiple MTU/network sends)");
REGISTERC(num_frag_sends,
"Number of sends where the entire MPI message fragment fit into a single MTU/network send");
REGISTERC(num_ack_sends,
"Number of ACKs sent (i.e., usNIC-BTL-to-usNIC-BTL control messages)");
REGISTERC(num_total_recvs,
"Total number of receives completed");
REGISTERC(num_unk_recvs,
"Number of receives with an unknown source or type, and therefore ignored by the usNIC BTL (this should never be >0)");
REGISTERC(num_dup_recvs,
"Number of duplicate receives");
REGISTERC(num_oow_low_recvs,
"Number of times a receive was out of the sliding window (on the low side)");
REGISTERC(num_oow_high_recvs,
"Number of times a receive was out of the sliding window (on the high side)");
REGISTERC(num_frag_recvs,
"Number of receives where the entire MPI message fragment fit into a single MTU/network send");
REGISTERC(num_chunk_recvs,
"Number of receives that were part of a larger MPI message fragment (i.e., this receive was reassembled into a larger MPI message fragment)");
REGISTERC(num_badfrag_recvs,
"Number of chunks received that had a bad fragment ID (this should never be >0)");
REGISTERC(num_ack_recvs,
"Total number of ACKs received");
REGISTERC(num_old_dup_acks,
"Number of old duplicate ACKs received (i.e., before the current expected ACK)");
REGISTERC(num_dup_acks,
"Number of duplicate ACKs received (i.e., the current expected ACK)");
REGISTERC(num_recv_reposts,
"Number of times buffers have been reposted for receives");
REGISTERC(num_crc_errors,
"Number of times receives were aborted because of a CRC error");
REGISTERC(pml_module_sends,
"Number of times the PML has called down to send a message");
REGISTERC(pml_send_callbacks,
"Number of times the usNIC BTL has called up to the PML to complete a send");
}
/*
* Initialize MPI_T performance variables
*/
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
int opal_btl_usnic_setup_mpit_pvars(void)
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
{
/* If we cannot find a compatible pvar type, we're done (i.e.,
don't register any pvars) */
if (!setup_mpit_pvar_type()) {
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
return OPAL_SUCCESS;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
}
/* Setup the usnic_X device enumeration pvar */
setup_mpit_pvars_enum();
/* Register watermark pvars */
setup_mpit_pvars_highwatermark();
/* If our counter stats are relative, don't report them through
MPI_T, because MPI_T expects counters to be monotonically
rising. */
if (!mca_btl_usnic_component.stats_relative) {
setup_mpit_pvars_counters();
}
/* All done */
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
return OPAL_SUCCESS;
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
}