Cleanup timing macros for portability across compilers. Rename the --enable-timing configure option to be --enable-pmix-timing so it doesn't pickup external timing requests. Remove a stale function reference in PMIx so it can compile with timing enabled.
Signed-off-by: Ralph Castain <rhc@open-mpi.org>
Этот коммит содержится в:
Ralph Castain
Boris Karasev
родитель
36a0e71f2d
Коммит
95ae0d1df3
1
.gitignore
поставляемый
1
.gitignore
поставляемый
@ -244,6 +244,7 @@ ompi/mpiext/cuda/c/mpiext_cuda_c.h
|
||||
ompi/tools/mpisync/mpisync
|
||||
ompi/tools/mpisync/mpirun_prof
|
||||
ompi/tools/mpisync/ompi_timing_post
|
||||
ompi/tools/mpisync/mpisync.1
|
||||
|
||||
ompi/tools/ompi_info/ompi_info
|
||||
ompi/tools/ompi_info/ompi_info.1
|
||||
|
||||
@ -1,6 +1,6 @@
|
||||
/*
|
||||
* Copyright (C) 2014 Artem Polyakov <artpol84@gmail.com>
|
||||
* Copyright (c) 2014 Intel, Inc. All rights reserved.
|
||||
* Copyright (c) 2014-2017 Intel, Inc. All rights reserved.
|
||||
* $COPYRIGHT$
|
||||
*
|
||||
* Additional copyrights may follow
|
||||
@ -139,7 +139,6 @@ int main(int argc, char **argv)
|
||||
|
||||
MPI_Gather(hname,sizeof(hname),MPI_CHAR,hnames,sizeof(hname),MPI_CHAR, 0, MPI_COMM_WORLD);
|
||||
MPI_Gather(send,2,MPI_DOUBLE,measure,2, MPI_DOUBLE, 0, MPI_COMM_WORLD);
|
||||
char tmpname[128];
|
||||
FILE *fp = fopen(filename,"w");
|
||||
if( fp == NULL ){
|
||||
fprintf(stderr, "Fail to open the file %s. Abort\n", filename);
|
||||
|
||||
@ -13,27 +13,27 @@ typedef struct {
|
||||
char *prefix;
|
||||
} ompi_timing_val_t;
|
||||
|
||||
typedef struct {
|
||||
ompi_timing_val_t *val;
|
||||
int use;
|
||||
struct ompi_timing_list_t *next;
|
||||
} ompi_timing_list_t;
|
||||
typedef struct {
|
||||
ompi_timing_val_t *val;
|
||||
int use;
|
||||
struct ompi_timing_list_t *next;
|
||||
} ompi_timing_list_t;
|
||||
|
||||
typedef struct ompi_timing_t {
|
||||
double ts;
|
||||
const char *prefix;
|
||||
int size;
|
||||
int cnt;
|
||||
int error;
|
||||
int enabled;
|
||||
opal_timing_ts_func_t get_ts;
|
||||
ompi_timing_list_t *timing;
|
||||
ompi_timing_list_t *cur_timing;
|
||||
} ompi_timing_t;
|
||||
typedef struct ompi_timing_t {
|
||||
double ts;
|
||||
const char *prefix;
|
||||
int size;
|
||||
int cnt;
|
||||
int error;
|
||||
int enabled;
|
||||
opal_timing_ts_func_t get_ts;
|
||||
ompi_timing_list_t *timing;
|
||||
ompi_timing_list_t *cur_timing;
|
||||
} ompi_timing_t;
|
||||
|
||||
#define OMPI_TIMING_INIT(_size) \
|
||||
ompi_timing_t OMPI_TIMING; \
|
||||
OMPI_TIMING.prefix = __FUNCTION__; \
|
||||
OMPI_TIMING.prefix = __func__; \
|
||||
OMPI_TIMING.size = _size; \
|
||||
OMPI_TIMING.get_ts = opal_timing_ts_func(OPAL_TIMING_AUTOMATIC_TIMER); \
|
||||
OMPI_TIMING.cnt = 0; \
|
||||
@ -55,156 +55,164 @@ typedef struct {
|
||||
} \
|
||||
}
|
||||
|
||||
#define OMPI_TIMING_ITEM_EXTEND ({ \
|
||||
if (OMPI_TIMING.enabled) { \
|
||||
OMPI_TIMING.cur_timing->next = (struct ompi_timing_list_t*)malloc(sizeof(ompi_timing_list_t)); \
|
||||
OMPI_TIMING.cur_timing = (ompi_timing_list_t*)OMPI_TIMING.cur_timing->next; \
|
||||
memset(OMPI_TIMING.cur_timing, 0, sizeof(ompi_timing_list_t)); \
|
||||
OMPI_TIMING.cur_timing->val = malloc(sizeof(ompi_timing_val_t) * OMPI_TIMING.size); \
|
||||
} \
|
||||
})
|
||||
#define OMPI_TIMING_ITEM_EXTEND \
|
||||
do { \
|
||||
if (OMPI_TIMING.enabled) { \
|
||||
OMPI_TIMING.cur_timing->next = (struct ompi_timing_list_t*)malloc(sizeof(ompi_timing_list_t)); \
|
||||
OMPI_TIMING.cur_timing = (ompi_timing_list_t*)OMPI_TIMING.cur_timing->next; \
|
||||
memset(OMPI_TIMING.cur_timing, 0, sizeof(ompi_timing_list_t)); \
|
||||
OMPI_TIMING.cur_timing->val = malloc(sizeof(ompi_timing_val_t) * OMPI_TIMING.size); \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define OMPI_TIMING_FINALIZE ({ \
|
||||
if (OMPI_TIMING.enabled) { \
|
||||
ompi_timing_list_t *t = OMPI_TIMING.timing, *tmp; \
|
||||
while ( NULL != t) { \
|
||||
tmp = t; \
|
||||
t = t->next; \
|
||||
free(tmp->val); \
|
||||
free(tmp); \
|
||||
} \
|
||||
OMPI_TIMING.timing = NULL; \
|
||||
OMPI_TIMING.cur_timing = NULL; \
|
||||
OMPI_TIMING.cnt = 0; \
|
||||
} \
|
||||
})
|
||||
#define OMPI_TIMING_FINALIZE \
|
||||
do { \
|
||||
if (OMPI_TIMING.enabled) { \
|
||||
ompi_timing_list_t *t = OMPI_TIMING.timing, *tmp; \
|
||||
while ( NULL != t) { \
|
||||
tmp = t; \
|
||||
t = (ompi_timing_list_t*)t->next; \
|
||||
free(tmp->val); \
|
||||
free(tmp); \
|
||||
} \
|
||||
OMPI_TIMING.timing = NULL; \
|
||||
OMPI_TIMING.cur_timing = NULL; \
|
||||
OMPI_TIMING.cnt = 0; \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define OMPI_TIMING_NEXT(fmt, ...) ({ \
|
||||
if (!OMPI_TIMING.error && OMPI_TIMING.enabled) { \
|
||||
char *f = strrchr(__FILE__, '/') + 1; \
|
||||
int len = 0; \
|
||||
if (OMPI_TIMING.cur_timing->use >= OMPI_TIMING.size){ \
|
||||
OMPI_TIMING_ITEM_EXTEND; \
|
||||
} \
|
||||
len = snprintf(OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].desc, \
|
||||
OPAL_TIMING_STR_LEN, fmt, ##__VA_ARGS__); \
|
||||
if (len >= OPAL_TIMING_STR_LEN) { \
|
||||
OMPI_TIMING.error = 1; \
|
||||
} \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].file = f; \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].prefix = __FUNCTION__; \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use++].ts = \
|
||||
OMPI_TIMING.get_ts() - OMPI_TIMING.ts; \
|
||||
OMPI_TIMING.cnt++; \
|
||||
OMPI_TIMING.ts = OMPI_TIMING.get_ts(); \
|
||||
} \
|
||||
})
|
||||
#define OMPI_TIMING_NEXT(...) \
|
||||
do { \
|
||||
if (!OMPI_TIMING.error && OMPI_TIMING.enabled) { \
|
||||
char *f = strrchr(__FILE__, '/') + 1; \
|
||||
int len = 0; \
|
||||
if (OMPI_TIMING.cur_timing->use >= OMPI_TIMING.size){ \
|
||||
OMPI_TIMING_ITEM_EXTEND; \
|
||||
} \
|
||||
len = snprintf(OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].desc, \
|
||||
OPAL_TIMING_STR_LEN, ##__VA_ARGS__); \
|
||||
if (len >= OPAL_TIMING_STR_LEN) { \
|
||||
OMPI_TIMING.error = 1; \
|
||||
} \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].file = strdup(f); \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].prefix = strdup(__func__); \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use++].ts = \
|
||||
OMPI_TIMING.get_ts() - OMPI_TIMING.ts; \
|
||||
OMPI_TIMING.cnt++; \
|
||||
OMPI_TIMING.ts = OMPI_TIMING.get_ts(); \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define OMPI_TIMING_APPEND(filename,func,desc,ts) { \
|
||||
if (OMPI_TIMING.cur_timing->use >= OMPI_TIMING.size){ \
|
||||
OMPI_TIMING_ITEM_EXTEND; \
|
||||
} \
|
||||
int len = snprintf(OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].desc, \
|
||||
OPAL_TIMING_STR_LEN, "%s", desc); \
|
||||
if (len >= OPAL_TIMING_STR_LEN) { \
|
||||
OMPI_TIMING.error = 1; \
|
||||
} \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].prefix = func; \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].file = filename; \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use++].ts = \
|
||||
OMPI_TIMING.get_ts() - OMPI_TIMING.ts; \
|
||||
OMPI_TIMING.cnt++; \
|
||||
OMPI_TIMING.ts = OMPI_TIMING.get_ts(); \
|
||||
}
|
||||
#define OMPI_TIMING_APPEND(filename,func,desc,ts) \
|
||||
do { \
|
||||
if (OMPI_TIMING.cur_timing->use >= OMPI_TIMING.size){ \
|
||||
OMPI_TIMING_ITEM_EXTEND; \
|
||||
} \
|
||||
int len = snprintf(OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].desc, \
|
||||
OPAL_TIMING_STR_LEN, "%s", desc); \
|
||||
if (len >= OPAL_TIMING_STR_LEN) { \
|
||||
OMPI_TIMING.error = 1; \
|
||||
} \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].prefix = func; \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use].file = filename; \
|
||||
OMPI_TIMING.cur_timing->val[OMPI_TIMING.cur_timing->use++].ts = \
|
||||
OMPI_TIMING.get_ts() - OMPI_TIMING.ts; \
|
||||
OMPI_TIMING.cnt++; \
|
||||
OMPI_TIMING.ts = OMPI_TIMING.get_ts(); \
|
||||
} while(0)
|
||||
|
||||
#define OMPI_TIMING_IMPORT_OPAL_PREFIX(_prefix, func) { \
|
||||
if (!OMPI_TIMING.error && OMPI_TIMING.enabled) { \
|
||||
int cnt = OPAL_TIMING_ENV_CNT(func); \
|
||||
int i; \
|
||||
OMPI_TIMING.error = OPAL_TIMING_ENV_ERROR_PREFIX(_prefix, func); \
|
||||
for(i = 0; i < cnt; i++){ \
|
||||
char *desc, *filename; \
|
||||
double ts = OPAL_TIMING_ENV_GETDESC_PREFIX(_prefix, &filename, func, i, &desc); \
|
||||
OMPI_TIMING_APPEND(filename, func, desc, ts); \
|
||||
} \
|
||||
} \
|
||||
}
|
||||
#define OMPI_TIMING_IMPORT_OPAL_PREFIX(_prefix, func) \
|
||||
do { \
|
||||
if (!OMPI_TIMING.error && OMPI_TIMING.enabled) { \
|
||||
int cnt; \
|
||||
int i; \
|
||||
double ts; \
|
||||
OPAL_TIMING_ENV_CNT(func, cnt); \
|
||||
OPAL_TIMING_ENV_ERROR_PREFIX(_prefix, func, OMPI_TIMING.error); \
|
||||
for(i = 0; i < cnt; i++){ \
|
||||
char *desc, *filename; \
|
||||
OPAL_TIMING_ENV_GETDESC_PREFIX(_prefix, &filename, func, i, &desc, ts); \
|
||||
OMPI_TIMING_APPEND(filename, func, desc, ts); \
|
||||
} \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define OMPI_TIMING_IMPORT_OPAL(func) \
|
||||
OMPI_TIMING_IMPORT_OPAL_PREFIX("", func)
|
||||
|
||||
|
||||
|
||||
#define OMPI_TIMING_OUT ({ \
|
||||
if (OMPI_TIMING.enabled) { \
|
||||
int i, size, rank; \
|
||||
MPI_Comm_size(MPI_COMM_WORLD, &size); \
|
||||
MPI_Comm_rank(MPI_COMM_WORLD, &rank); \
|
||||
int error = 0; \
|
||||
\
|
||||
MPI_Reduce(&OMPI_TIMING.error, &error, 1, \
|
||||
MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD); \
|
||||
\
|
||||
if (error) { \
|
||||
if (0 == rank) { \
|
||||
printf("==OMPI_TIMING== error: something went wrong, timings doesn't work\n"); \
|
||||
} \
|
||||
} \
|
||||
else { \
|
||||
double *avg = (double*)malloc(sizeof(double) * OMPI_TIMING.cnt); \
|
||||
double *min = (double*)malloc(sizeof(double) * OMPI_TIMING.cnt); \
|
||||
double *max = (double*)malloc(sizeof(double) * OMPI_TIMING.cnt); \
|
||||
char **desc = (char**)malloc(sizeof(char*) * OMPI_TIMING.cnt); \
|
||||
char **prefix = (char**)malloc(sizeof(char*) * OMPI_TIMING.cnt); \
|
||||
char **file = (char**)malloc(sizeof(char*) * OMPI_TIMING.cnt); \
|
||||
\
|
||||
if( OMPI_TIMING.cnt > 0 ) { \
|
||||
OMPI_TIMING.ts = OMPI_TIMING.get_ts(); \
|
||||
ompi_timing_list_t *timing = OMPI_TIMING.timing; \
|
||||
i = 0; \
|
||||
do { \
|
||||
int use; \
|
||||
for (use = 0; use < timing->use; use++) { \
|
||||
MPI_Reduce(&timing->val[use].ts, avg + i, 1, \
|
||||
MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD); \
|
||||
MPI_Reduce(&timing->val[use].ts, min + i, 1, \
|
||||
MPI_DOUBLE, MPI_MIN, 0, MPI_COMM_WORLD); \
|
||||
MPI_Reduce(&timing->val[use].ts, max + i, 1, \
|
||||
MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD); \
|
||||
desc[i] = timing->val[use].desc; \
|
||||
prefix[i] = timing->val[use].prefix; \
|
||||
file[i] = timing->val[use].file; \
|
||||
i++; \
|
||||
} \
|
||||
timing = (ompi_timing_list_t*)timing->next; \
|
||||
} while (timing != NULL); \
|
||||
\
|
||||
if( 0 == rank ){ \
|
||||
if (OMPI_TIMING.timing->next) { \
|
||||
printf("==OMPI_TIMING== warning: added the extra timings allocation that might misrepresent the results.\n" \
|
||||
"==OMPI_TIMING== Increase the inited size of timings to avoid extra allocation during runtime.\n"); \
|
||||
} \
|
||||
\
|
||||
printf("------------------ %s ------------------\n", \
|
||||
OMPI_TIMING.prefix); \
|
||||
for(i=0; i< OMPI_TIMING.cnt; i++){ \
|
||||
avg[i] /= size; \
|
||||
printf("[%s:%s:%s]: %lf / %lf / %lf\n", \
|
||||
file[i], prefix[i], desc[i], avg[i], min[i], max[i]); \
|
||||
} \
|
||||
printf("[%s:overhead]: %lf \n", OMPI_TIMING.prefix, \
|
||||
OMPI_TIMING.get_ts() - OMPI_TIMING.ts); \
|
||||
} \
|
||||
} \
|
||||
free(avg); \
|
||||
free(min); \
|
||||
free(max); \
|
||||
free(desc); \
|
||||
free(prefix); \
|
||||
free(file); \
|
||||
} \
|
||||
} \
|
||||
})
|
||||
#define OMPI_TIMING_OUT \
|
||||
do { \
|
||||
if (OMPI_TIMING.enabled) { \
|
||||
int i, size, rank; \
|
||||
MPI_Comm_size(MPI_COMM_WORLD, &size); \
|
||||
MPI_Comm_rank(MPI_COMM_WORLD, &rank); \
|
||||
int error = 0; \
|
||||
\
|
||||
MPI_Reduce(&OMPI_TIMING.error, &error, 1, \
|
||||
MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD); \
|
||||
\
|
||||
if (error) { \
|
||||
if (0 == rank) { \
|
||||
printf("==OMPI_TIMING== error: something went wrong, timings doesn't work\n"); \
|
||||
} \
|
||||
} \
|
||||
else { \
|
||||
double *avg = (double*)malloc(sizeof(double) * OMPI_TIMING.cnt); \
|
||||
double *min = (double*)malloc(sizeof(double) * OMPI_TIMING.cnt); \
|
||||
double *max = (double*)malloc(sizeof(double) * OMPI_TIMING.cnt); \
|
||||
char **desc = (char**)malloc(sizeof(char*) * OMPI_TIMING.cnt); \
|
||||
char **prefix = (char**)malloc(sizeof(char*) * OMPI_TIMING.cnt); \
|
||||
char **file = (char**)malloc(sizeof(char*) * OMPI_TIMING.cnt); \
|
||||
\
|
||||
if( OMPI_TIMING.cnt > 0 ) { \
|
||||
OMPI_TIMING.ts = OMPI_TIMING.get_ts(); \
|
||||
ompi_timing_list_t *timing = OMPI_TIMING.timing; \
|
||||
i = 0; \
|
||||
do { \
|
||||
int use; \
|
||||
for (use = 0; use < timing->use; use++) { \
|
||||
MPI_Reduce(&timing->val[use].ts, avg + i, 1, \
|
||||
MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD); \
|
||||
MPI_Reduce(&timing->val[use].ts, min + i, 1, \
|
||||
MPI_DOUBLE, MPI_MIN, 0, MPI_COMM_WORLD); \
|
||||
MPI_Reduce(&timing->val[use].ts, max + i, 1, \
|
||||
MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD); \
|
||||
desc[i] = timing->val[use].desc; \
|
||||
prefix[i] = timing->val[use].prefix; \
|
||||
file[i] = timing->val[use].file; \
|
||||
i++; \
|
||||
} \
|
||||
timing = (ompi_timing_list_t*)timing->next; \
|
||||
} while (timing != NULL); \
|
||||
\
|
||||
if( 0 == rank ){ \
|
||||
if (OMPI_TIMING.timing->next) { \
|
||||
printf("==OMPI_TIMING== warning: added the extra timings allocation that might misrepresent the results.\n" \
|
||||
"==OMPI_TIMING== Increase the inited size of timings to avoid extra allocation during runtime.\n"); \
|
||||
} \
|
||||
\
|
||||
printf("------------------ %s ------------------\n", \
|
||||
OMPI_TIMING.prefix); \
|
||||
for(i=0; i< OMPI_TIMING.cnt; i++){ \
|
||||
avg[i] /= size; \
|
||||
printf("[%s:%s:%s]: %lf / %lf / %lf\n", \
|
||||
file[i], prefix[i], desc[i], avg[i], min[i], max[i]); \
|
||||
} \
|
||||
printf("[%s:overhead]: %lf \n", OMPI_TIMING.prefix, \
|
||||
OMPI_TIMING.get_ts() - OMPI_TIMING.ts); \
|
||||
} \
|
||||
} \
|
||||
free(avg); \
|
||||
free(min); \
|
||||
free(max); \
|
||||
free(desc); \
|
||||
free(prefix); \
|
||||
free(file); \
|
||||
} \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#else
|
||||
#define OMPI_TIMING_INIT(size)
|
||||
|
||||
@ -49,7 +49,19 @@ AC_DEFUN([MCA_opal_pmix_pmix2x_CONFIG],[
|
||||
opal_pmix_pmix2x_sm_flag=--disable-dstore
|
||||
fi
|
||||
|
||||
opal_pmix_pmix2x_args="--with-pmix-symbol-rename=OPAL_MCA_PMIX2X_ $opal_pmix_pmix2x_sm_flag --without-tests-examples --disable-visibility --enable-embedded-libevent --with-libevent-header=\\\"opal/mca/event/$opal_event_base_include\\\""
|
||||
AC_ARG_ENABLE([pmix-timing],
|
||||
[AC_HELP_STRING([--enable-pmix-timing],
|
||||
[Enable PMIx timing measurements (default: disabled)])])
|
||||
AC_MSG_CHECKING([if PMIx timing is enabled])
|
||||
if test "$enable_pmix_timing" == "yes"; then
|
||||
AC_MSG_RESULT([yes])
|
||||
opal_pmix_pmix2x_timing_flag=--enable-pmix-timing
|
||||
else
|
||||
AC_MSG_RESULT([no (disabled)])
|
||||
opal_pmix_pmix2x_timing_flag=--disable-pmix-timing
|
||||
fi
|
||||
|
||||
opal_pmix_pmix2x_args="--with-pmix-symbol-rename=OPAL_MCA_PMIX2X_ $opal_pmix_pmix2x_sm_flag $opal_pmix_pmix2x_timing_flag --without-tests-examples --disable-visibility --enable-embedded-libevent --with-libevent-header=\\\"opal/mca/event/$opal_event_base_include\\\""
|
||||
AS_IF([test "$enable_debug" = "yes"],
|
||||
[opal_pmix_pmix2x_args="--enable-debug $opal_pmix_pmix2x_args"
|
||||
CFLAGS="$OPAL_CFLAGS_BEFORE_PICKY $OPAL_VISIBILITY_CFLAGS -g"],
|
||||
|
||||
@ -950,18 +950,18 @@ AC_MSG_RESULT([$with_ident_string])
|
||||
# Timing support
|
||||
#
|
||||
AC_MSG_CHECKING([if want developer-level timing support])
|
||||
AC_ARG_ENABLE(timing,
|
||||
AC_HELP_STRING([--enable-timing],
|
||||
AC_ARG_ENABLE(pmix-timing,
|
||||
AC_HELP_STRING([--enable-pmix-timing],
|
||||
[enable developer-level timing code (default: disabled)]))
|
||||
if test "$enable_timing" = "yes"; then
|
||||
if test "$enable_pmix_timing" = "yes"; then
|
||||
AC_MSG_RESULT([yes])
|
||||
WANT_TIMING=1
|
||||
WANT_PMIX_TIMING=1
|
||||
else
|
||||
AC_MSG_RESULT([no])
|
||||
WANT_TIMING=0
|
||||
WANT_PMIX_TIMING=0
|
||||
fi
|
||||
|
||||
AC_DEFINE_UNQUOTED([PMIX_ENABLE_TIMING], [$WANT_TIMING],
|
||||
AC_DEFINE_UNQUOTED([PMIX_ENABLE_TIMING], [$WANT_PMIX_TIMING],
|
||||
[Whether we want developer-level timing support or not])
|
||||
|
||||
#
|
||||
|
||||
@ -21,7 +21,7 @@
|
||||
* and Technology (RIST). All rights reserved.
|
||||
* Copyright (c) 2015 Mellanox Technologies, Inc.
|
||||
* All rights reserved.
|
||||
* Copyright (c) 2016 Intel, Inc. All rights reserved.
|
||||
* Copyright (c) 2016-2017 Intel, Inc. All rights reserved.
|
||||
* $COPYRIGHT$
|
||||
*
|
||||
* Additional copyrights may follow
|
||||
@ -37,7 +37,6 @@
|
||||
#include "src/util/timings.h"
|
||||
|
||||
#if PMIX_ENABLE_TIMING
|
||||
char *pmix_timing_sync_file = NULL;
|
||||
char *pmix_timing_output = NULL;
|
||||
bool pmix_timing_overhead = true;
|
||||
#endif
|
||||
@ -56,16 +55,6 @@ pmix_status_t pmix_register_params(void)
|
||||
pmix_register_done = true;
|
||||
|
||||
#if PMIX_ENABLE_TIMING
|
||||
pmix_timing_sync_file = NULL;
|
||||
(void) pmix_mca_base_var_register ("pmix", "pmix", NULL, "timing_sync_file",
|
||||
"Clock synchronisation information generated by mpisync tool. You don't need to touch this if you use mpirun_prof tool.",
|
||||
PMIX_MCA_BASE_VAR_TYPE_STRING, NULL, 0, 0,
|
||||
PMIX_INFO_LVL_9, PMIX_MCA_BASE_VAR_SCOPE_ALL,
|
||||
&pmix_timing_sync_file);
|
||||
if( pmix_timing_clocksync_read(pmix_timing_sync_file) ){
|
||||
pmix_output(0, "Cannot read file %s containing clock synchronisation information\n", pmix_timing_sync_file);
|
||||
}
|
||||
|
||||
pmix_timing_output = NULL;
|
||||
(void) pmix_mca_base_var_register ("pmix", "pmix", NULL, "timing_output",
|
||||
"The name of output file for timing information. If this parameter is not set then output will be directed into PMIX debug channel.",
|
||||
|
||||
@ -1,6 +1,6 @@
|
||||
/*
|
||||
* Copyright (C) 2014 Artem Polyakov <artpol84@gmail.com>
|
||||
* Copyright (c) 2014 Intel, Inc. All rights reserved.
|
||||
* Copyright (c) 2014-2017 Intel, Inc. All rights reserved.
|
||||
* $COPYRIGHT$
|
||||
*
|
||||
* Additional copyrights may follow
|
||||
@ -39,47 +39,46 @@ typedef struct {
|
||||
|
||||
opal_timing_ts_func_t opal_timing_ts_func(opal_timer_type_t type);
|
||||
|
||||
#define OPAL_TIMING_ENV_START_TYPE(func, type, prefix) ({ \
|
||||
opal_timing_env_t h; \
|
||||
char *ptr = NULL; \
|
||||
char *_prefix = prefix; \
|
||||
int n; \
|
||||
if( NULL == prefix ){ \
|
||||
_prefix = ""; \
|
||||
} \
|
||||
h.error = 0; \
|
||||
n = snprintf(h.id, OPAL_TIMING_STR_LEN, "%s%s", _prefix, func); \
|
||||
if( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h.error = 1; \
|
||||
} \
|
||||
n = sprintf(h.cntr_env,"OMPI_TIMING_%s%s_CNT", prefix, h.id); \
|
||||
if( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h.error = 1; \
|
||||
} \
|
||||
ptr = getenv(h.id); \
|
||||
if( NULL == ptr || strcmp(ptr, "1")){ \
|
||||
h.enabled = 0; \
|
||||
} \
|
||||
h.get_ts = opal_timing_ts_func(type); \
|
||||
ptr = getenv("OPAL_TIMING_ENABLE"); \
|
||||
if (NULL != ptr) { \
|
||||
h.enabled = atoi(ptr); \
|
||||
} \
|
||||
h.cntr = 0; \
|
||||
ptr = getenv(h.id); \
|
||||
if( NULL != ptr ){ \
|
||||
h.cntr = atoi(ptr); \
|
||||
} \
|
||||
h.ts = h.get_ts(); \
|
||||
if ( 0 != h.error ){ \
|
||||
h.enabled = 0; \
|
||||
} \
|
||||
h; \
|
||||
})
|
||||
#define OPAL_TIMING_ENV_START_TYPE(func, _nm, type, prefix) \
|
||||
do { \
|
||||
char *ptr = NULL; \
|
||||
char *_prefix = prefix; \
|
||||
int n; \
|
||||
if( NULL == prefix ){ \
|
||||
_prefix = ""; \
|
||||
} \
|
||||
(_nm)->error = 0; \
|
||||
n = snprintf((_nm)->id, OPAL_TIMING_STR_LEN, "%s%s", _prefix, func); \
|
||||
if( n > OPAL_TIMING_STR_LEN ){ \
|
||||
(_nm)->error = 1; \
|
||||
} \
|
||||
n = sprintf((_nm)->cntr_env,"OMPI_TIMING_%s%s_CNT", prefix, (_nm)->id); \
|
||||
if( n > OPAL_TIMING_STR_LEN ){ \
|
||||
(_nm)->error = 1; \
|
||||
} \
|
||||
ptr = getenv((_nm)->id); \
|
||||
if( NULL == ptr || strcmp(ptr, "1")){ \
|
||||
(_nm)->enabled = 0; \
|
||||
} \
|
||||
(_nm)->get_ts = opal_timing_ts_func(type); \
|
||||
ptr = getenv("OPAL_TIMING_ENABLE"); \
|
||||
if (NULL != ptr) { \
|
||||
(_nm)->enabled = atoi(ptr); \
|
||||
} \
|
||||
(_nm)->cntr = 0; \
|
||||
ptr = getenv((_nm)->id); \
|
||||
if( NULL != ptr ){ \
|
||||
(_nm)->cntr = atoi(ptr); \
|
||||
} \
|
||||
(_nm)->ts = (_nm)->get_ts(); \
|
||||
if ( 0 != (_nm)->error ){ \
|
||||
(_nm)->enabled = 0; \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define OPAL_TIMING_ENV_INIT(name) \
|
||||
opal_timing_env_t name ## _val, *name = &(name ## _val); \
|
||||
*name = OPAL_TIMING_ENV_START_TYPE(__FUNCTION__, OPAL_TIMING_AUTOMATIC_TIMER, "");
|
||||
opal_timing_env_t name ## _val, *name = &(name ## _val); \
|
||||
OPAL_TIMING_ENV_START_TYPE(__func__, name, OPAL_TIMING_AUTOMATIC_TIMER, "");
|
||||
|
||||
/* We use function names for identification
|
||||
* however this might be a problem for the private
|
||||
@ -88,111 +87,114 @@ opal_timing_ts_func_t opal_timing_ts_func(opal_timer_type_t type);
|
||||
* Use prefix to do a finer-grained identification if needed
|
||||
*/
|
||||
#define OPAL_TIMING_ENV_INIT_PREFIX(prefix, name) \
|
||||
opal_timing_env_t name ## _val, *name = &(name ## _val); \
|
||||
*name = OPAL_TIMING_ENV_START_TYPE(__FUNCTION__, OPAL_TIMING_AUTOMATIC_TIMER, prefix);
|
||||
do { \
|
||||
opal_timing_env_t name ## _val, *name = &(name ## _val); \
|
||||
*name = OPAL_TIMING_ENV_START_TYPE(__func__, OPAL_TIMING_AUTOMATIC_TIMER, prefix); \
|
||||
} while(0)
|
||||
|
||||
#define OPAL_TIMING_ENV_NEXT(h, fmt, ...) ({ \
|
||||
int n; \
|
||||
char buf1[OPAL_TIMING_STR_LEN], buf2[OPAL_TIMING_STR_LEN]; \
|
||||
double time; \
|
||||
char *filename; \
|
||||
if( h->enabled ){ \
|
||||
/* enabled codepath */ \
|
||||
time = h->get_ts() - h->ts; \
|
||||
n = snprintf(buf1, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s_DESC_%d", h->id, h->cntr); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
n = snprintf(buf2, OPAL_TIMING_STR_LEN, fmt, ## __VA_ARGS__ ); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
setenv(buf1, buf2, 1); \
|
||||
n = snprintf(buf1, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s_VAL_%d", h->id, h->cntr); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
n = snprintf(buf2, OPAL_TIMING_STR_LEN, "%lf", time); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
setenv(buf1, buf2, 1); \
|
||||
filename = strrchr(__FILE__, '/') + 1; \
|
||||
n = snprintf(buf1, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s_FILE_%d", h->id, h->cntr); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
n = snprintf(buf2, OPAL_TIMING_STR_LEN, "%s", filename); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
setenv(buf1, buf2, 1); \
|
||||
h->cntr++; \
|
||||
sprintf(buf1, "%d", h->cntr); \
|
||||
setenv(h->cntr_env, buf1, 1); \
|
||||
/* We don't include env operations into the consideration.
|
||||
* Hopefully this will help to make measurements more accurate.
|
||||
*/ \
|
||||
h->ts = h->get_ts(); \
|
||||
} \
|
||||
if (h->error) { \
|
||||
n = snprintf(buf1, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s_ERROR", h->id);\
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
n = snprintf(buf2, OPAL_TIMING_STR_LEN, "%d", h->error); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
setenv(buf1, buf2, 1); \
|
||||
} \
|
||||
})
|
||||
#define OPAL_TIMING_ENV_NEXT(h, ...) \
|
||||
do { \
|
||||
int n; \
|
||||
char buf1[OPAL_TIMING_STR_LEN], buf2[OPAL_TIMING_STR_LEN]; \
|
||||
double time; \
|
||||
char *filename; \
|
||||
if( h->enabled ){ \
|
||||
/* enabled codepath */ \
|
||||
time = h->get_ts() - h->ts; \
|
||||
n = snprintf(buf1, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s_DESC_%d", h->id, h->cntr); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
n = snprintf(buf2, OPAL_TIMING_STR_LEN, __VA_ARGS__ ); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
setenv(buf1, buf2, 1); \
|
||||
n = snprintf(buf1, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s_VAL_%d", h->id, h->cntr); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
n = snprintf(buf2, OPAL_TIMING_STR_LEN, "%lf", time); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
setenv(buf1, buf2, 1); \
|
||||
filename = strrchr(__FILE__, '/') + 1; \
|
||||
n = snprintf(buf1, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s_FILE_%d", h->id, h->cntr); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
n = snprintf(buf2, OPAL_TIMING_STR_LEN, "%s", filename); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
setenv(buf1, buf2, 1); \
|
||||
h->cntr++; \
|
||||
sprintf(buf1, "%d", h->cntr); \
|
||||
setenv(h->cntr_env, buf1, 1); \
|
||||
/* We don't include env operations into the consideration.
|
||||
* Hopefully this will help to make measurements more accurate.
|
||||
*/ \
|
||||
h->ts = h->get_ts(); \
|
||||
} \
|
||||
if (h->error) { \
|
||||
n = snprintf(buf1, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s_ERROR", h->id);\
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
n = snprintf(buf2, OPAL_TIMING_STR_LEN, "%d", h->error); \
|
||||
if ( n > OPAL_TIMING_STR_LEN ){ \
|
||||
h->error = 1; \
|
||||
} \
|
||||
setenv(buf1, buf2, 1); \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
/* This function supposed to be called from the code that will
|
||||
* do the postprocessing, i.e. OMPI timing portion that will
|
||||
* do the reduction of accumulated values
|
||||
*/
|
||||
#define OPAL_TIMING_ENV_CNT_PREFIX(prefix, func) ({ \
|
||||
char ename[OPAL_TIMING_STR_LEN]; \
|
||||
int cnt = 0; \
|
||||
char *ptr = NULL; \
|
||||
int n = snprintf(ename, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s%s_CNT", prefix, func); \
|
||||
if ( n <= OPAL_TIMING_STR_LEN ){ \
|
||||
ptr = getenv(ename); \
|
||||
if( NULL != ptr ){ cnt = atoi(ptr); }; \
|
||||
} \
|
||||
cnt; \
|
||||
})
|
||||
#define OPAL_TIMING_ENV_CNT_PREFIX(prefix, func, _cnt) \
|
||||
do { \
|
||||
char ename[OPAL_TIMING_STR_LEN]; \
|
||||
char *ptr = NULL; \
|
||||
int n = snprintf(ename, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s%s_CNT", prefix, func); \
|
||||
(_cnt) = 0; \
|
||||
if ( n <= OPAL_TIMING_STR_LEN ){ \
|
||||
ptr = getenv(ename); \
|
||||
if( NULL != ptr ){ (_cnt) = atoi(ptr); }; \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define OPAL_TIMING_ENV_ERROR_PREFIX(prefix, func) ({ \
|
||||
char ename[OPAL_TIMING_STR_LEN]; \
|
||||
int error = 0; \
|
||||
char *ptr = NULL; \
|
||||
int n = snprintf(ename, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s%s_ERROR", prefix, func); \
|
||||
if ( n <= OPAL_TIMING_STR_LEN ){ \
|
||||
ptr = getenv(ename); \
|
||||
if( NULL != ptr ){ error = atoi(ptr); }; \
|
||||
} \
|
||||
error; \
|
||||
})
|
||||
#define OPAL_TIMING_ENV_ERROR_PREFIX(prefix, func, _err) \
|
||||
do { \
|
||||
char ename[OPAL_TIMING_STR_LEN]; \
|
||||
(_err) = 0; \
|
||||
char *ptr = NULL; \
|
||||
int n = snprintf(ename, OPAL_TIMING_STR_LEN, "OMPI_TIMING_%s%s_ERROR", prefix, func); \
|
||||
if ( n <= OPAL_TIMING_STR_LEN ){ \
|
||||
ptr = getenv(ename); \
|
||||
if( NULL != ptr ){ (_err) = atoi(ptr); }; \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define OPAL_TIMING_ENV_CNT(func) \
|
||||
OPAL_TIMING_ENV_CNT_PREFIX("", func)
|
||||
#define OPAL_TIMING_ENV_CNT(func, _cnt) \
|
||||
OPAL_TIMING_ENV_CNT_PREFIX("", func, _cnt)
|
||||
|
||||
#define OPAL_TIMING_ENV_GETDESC_PREFIX(prefix, filename, func, i, desc) ({ \
|
||||
char vname[OPAL_TIMING_STR_LEN]; \
|
||||
double ts = 0.0; \
|
||||
sprintf(vname, "OMPI_TIMING_%s%s_FILE_%d", prefix, func, i); \
|
||||
*filename = getenv(vname); \
|
||||
sprintf(vname, "OMPI_TIMING_%s%s_DESC_%d", prefix, func, i); \
|
||||
*desc = getenv(vname); \
|
||||
sprintf(vname, "OMPI_TIMING_%s%s_VAL_%d", prefix, func, i); \
|
||||
char *ptr = getenv(vname); \
|
||||
if ( NULL != ptr ) { \
|
||||
sscanf(ptr,"%lf", &ts); \
|
||||
} \
|
||||
ts; \
|
||||
})
|
||||
#define OPAL_TIMING_ENV_GETDESC_PREFIX(prefix, filename, func, i, desc, _t) \
|
||||
do { \
|
||||
char vname[OPAL_TIMING_STR_LEN]; \
|
||||
(_t) = 0.0; \
|
||||
sprintf(vname, "OMPI_TIMING_%s%s_FILE_%d", prefix, func, i); \
|
||||
*filename = getenv(vname); \
|
||||
sprintf(vname, "OMPI_TIMING_%s%s_DESC_%d", prefix, func, i); \
|
||||
*desc = getenv(vname); \
|
||||
sprintf(vname, "OMPI_TIMING_%s%s_VAL_%d", prefix, func, i); \
|
||||
char *ptr = getenv(vname); \
|
||||
if ( NULL != ptr ) { \
|
||||
sscanf(ptr,"%lf", &(_t)); \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define OPAL_TIMING_ENV_GETDESC(file, func, index, desc) \
|
||||
OPAL_TIMING_ENV_GETDESC_PREFIX("", file, func, index, desc)
|
||||
|
||||
@ -343,9 +343,6 @@ void mca_oob_tcp_send_handler(int sd, short flags, void *cbdata)
|
||||
static int read_bytes(mca_oob_tcp_peer_t* peer)
|
||||
{
|
||||
int rc;
|
||||
#if OPAL_ENABLE_TIMING
|
||||
int to_read = peer->recv_msg->rdbytes;
|
||||
#endif
|
||||
|
||||
/* read until all bytes recvd or error */
|
||||
while (0 < peer->recv_msg->rdbytes) {
|
||||
@ -431,9 +428,6 @@ void mca_oob_tcp_recv_handler(int sd, short flags, void *cbdata)
|
||||
mca_oob_tcp_peer_t* peer = (mca_oob_tcp_peer_t*)cbdata;
|
||||
int rc;
|
||||
orte_rml_send_t *snd;
|
||||
#if OPAL_ENABLE_TIMING
|
||||
bool timing_same_as_hdr = false;
|
||||
#endif
|
||||
|
||||
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
|
||||
"%s:tcp:recv:handler called for peer %s",
|
||||
@ -503,13 +497,7 @@ void mca_oob_tcp_recv_handler(int sd, short flags, void *cbdata)
|
||||
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
|
||||
"%s:tcp:recv:handler read hdr",
|
||||
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
||||
#if OPAL_ENABLE_TIMING
|
||||
int to_recv = peer->recv_msg->rdbytes;
|
||||
#endif
|
||||
if (ORTE_SUCCESS == (rc = read_bytes(peer))) {
|
||||
#if OPAL_ENABLE_TIMING
|
||||
timing_same_as_hdr = true;
|
||||
#endif
|
||||
/* completed reading the header */
|
||||
peer->recv_msg->hdr_recvd = true;
|
||||
/* convert the header */
|
||||
|
||||
Загрузка…
x
Ссылка в новой задаче
Block a user