[dcmf] Re: [PATCH] This test will repeatedly call the low-level critical-section functions for performance testing.

Sameer Kumar sameerk at us.ibm.com
Tue Feb 5 12:02:13 CST 2008


Joe,
      If the lock is called once per message transaction the rate will be
4.7 MMPS,  if its called twice it will  be 2.3 MMPS and  four times 1.2
MMPS.  So we need to investigate how many times the lock is called per
procnull message send and recv.  May be someone in Argonne can answer that.


                  sameer.




                                                                           
             Joseph                                                        
             Ratterman/Rochest                                             
             er/IBM at IBMUS                                               To 
             Sent by:                  Joseph                              
             dcmf-bounces at list         Ratterman/Rochester/IBM at IBMUS       
             s.anl-external.or                                          cc 
             g                         DCMF <dcmf at lists.anl-external.org>  
                                                                   Subject 
                                       [dcmf] Re: [PATCH] This test will   
             02/05/2008 12:23          repeatedly call the low-level       
             PM                        critical-section functions for      
                                       performance testing.                
                                                                           
                                                                           
                                                                           
                                                                           
                                                                           
                                                                           





Here are the results from running this test:

$ mpirun -np 1 -nofree -mode SMP ./build-tests/perf/dcmf/CS.cnk
DCMF_THREAD_SINGLE: Called Enter/Exit 10000 times at 78.0386 cycles each.
DCMF_THREAD_FUNNELED: Called Enter/Exit 10000 times at 78.0054 cycles each.

DCMF_THREAD_SERIALIZED: Called Enter/Exit 10000 times at 78.0029 cycles
each.
DCMF_THREAD_MULTIPLE: Called Enter/Exit 10000 times at 180.063 cycles each.

$ mpirun -np 1 -nofree -mode DUAL ./build-tests/perf/dcmf/CS.cnk
DCMF_THREAD_SINGLE: Called Enter/Exit 10000 times at 78.0378 cycles each.
DCMF_THREAD_FUNNELED: Called Enter/Exit 10000 times at 78.0032 cycles each.

DCMF_THREAD_SERIALIZED: Called Enter/Exit 10000 times at 78.0019 cycles
each.
DCMF_THREAD_MULTIPLE: Called Enter/Exit 10000 times at 196.044 cycles each.


While this is a doubling in the time it takes to lock/unlock, that alone
wouldn't drop the one process/thread performance from 4.47 to 1 MMPS.  We
will look into it more after we get the benchmark.


Thanks,
Joe Ratterman



                                                                           
 Joseph                                                                    
 Ratterman/Rochester/IBM@                                                  
 IBMUS                                                                     
                                                                        To 
                                   DCMF <dcmf at lists.anl-external.org>      
 02/05/08 11:19 AM                                                      cc 
                                   Joseph Ratterman/Rochester/IBM at IBMUS    
                                                                   Subject 
                                   [PATCH] This test will repeatedly call  
                                   the low-level critical-section          
                                   functions for performance testing.      
                                                                           
                                                                           
                                                                           
                                                                           
                                                                           
                                                                           
                                                                           





This is helpful when trying to understand performance degradations in
MPI_THREAD_MULTIPLE.

Signed-off-by: Joe Ratterman <jratt at us.ibm.com>
---
sys/tests/perf/Makefile.in            |    2 +-
sys/tests/perf/dcmf/CS.c              |   64 ++++++++++++++++++++++++++++++
+++
sys/tests/perf/{ => dcmf}/Makefile.in |    4 +-
3 files changed, 67 insertions(+), 3 deletions(-)
create mode 100644 sys/tests/perf/dcmf/CS.c
copy sys/tests/perf/{ => dcmf}/Makefile.in (96%)

diff --git a/sys/tests/perf/Makefile.in b/sys/tests/perf/Makefile.in
index 22a7ac7..4266989 100644
--- a/sys/tests/perf/Makefile.in
+++ b/sys/tests/perf/Makefile.in
@@ -12,6 +12,6 @@
# end_generated_IBM_copyright_prolog                               #

VPATH                                  = @abs_srcdir@
-SUBDIRS                                  = mpi spi mpid
+SUBDIRS                                  = mpi spi mpid dcmf
TESTS                                  =
include @abs_top_builddir@/Make.rules
diff --git a/sys/tests/perf/dcmf/CS.c b/sys/tests/perf/dcmf/CS.c
new file mode 100644
index 0000000..080f5df
--- /dev/null
+++ b/sys/tests/perf/dcmf/CS.c
@@ -0,0 +1,64 @@
+/* begin_generated_IBM_copyright_prolog                             */
+/*                                                                  */
+/* ---------------------------------------------------------------- */
+/* (C)Copyright IBM Corp.  2007, 2008                               */
+/* IBM CPL License                                                  */
+/* ---------------------------------------------------------------- */
+/*                                                                  */
+/* end_generated_IBM_copyright_prolog                               */
+/**
+ * \file perf/dcmf/CS.c
+ * \brief Test the performance of the low-level critical-section functions
+ */
+
+
+#include <tests.h>
+#define NUM 10000
+DCMF_Configure_t config;
+
+
+double time_CS(uint32_t x)
+{
+  uint64_t start, stop;
+  uint32_t i;
+
+  start = DCMF_Timebase();
+  for (i=0; i<x; ++i) {
+    DCMF_CriticalSection_enter(0);
+    DCMF_CriticalSection_exit(0);
+  }
+  stop  = DCMF_Timebase();
+
+  return (double)(stop-start)/(double)x;
+}
+
+
+#define time_run(c) time_run_long(c, #c)
+void time_run_long(DCMF_Thread thread_level, char* thread_string)
+{
+  double time;
+  DCMF_Result rc;
+
+  config.thread_level = thread_level;
+  rc = DCMF_Messager_configure (&config, &config);
+  assert(rc == DCMF_SUCCESS);
+  assert(config.thread_level == thread_level);
+  time = time_CS(NUM);
+  printf("%s: Called Enter/Exit %u times at %g cycles each.\n",
thread_string, NUM, time);
+}
+
+
+int main()
+{
+  config.interrupts = DCMF_INTERRUPTS_OFF;
+
+  MPI_INIT;
+
+  time_run(DCMF_THREAD_SINGLE);
+  time_run(DCMF_THREAD_FUNNELED);
+  time_run(DCMF_THREAD_SERIALIZED);
+  time_run(DCMF_THREAD_MULTIPLE);
+
+  MPI_FINALIZE;
+  return (0);
+}
diff --git a/sys/tests/perf/Makefile.in b/sys/tests/perf/dcmf/Makefile.in
similarity index 96%
copy from sys/tests/perf/Makefile.in
copy to sys/tests/perf/dcmf/Makefile.in
index 22a7ac7..4c474b6 100644
--- a/sys/tests/perf/Makefile.in
+++ b/sys/tests/perf/dcmf/Makefile.in
@@ -12,6 +12,6 @@
# end_generated_IBM_copyright_prolog                               #

VPATH                                  = @abs_srcdir@
-SUBDIRS                                  = mpi spi mpid
-TESTS                                  =
+SUBDIRS                                  =
+TESTS                                  = CS.c
include @abs_top_builddir@/Make.rules
--
1.5.4

_______________________________________________
dcmf mailing list
dcmf at lists.anl-external.org
http://lists.anl-external.org/cgi-bin/mailman/listinfo/dcmf





More information about the dcmf mailing list