maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

Improving EPCC performance when linking with hwloc 

When linking with libhwloc, the ORDERED EPCC test slows down on big
machines (> 48 cores). Performance analysis showed that a cache thrash
was occurring and this padding helps alleviate the problem.

Also, inside the main spin-wait loop in kmp_wait_release.h, we can eliminate
the references to the global shared variables by instead creating a local
variable, oversubscribed and instead checking that.

Differential Revision: http://reviews.llvm.org/D22093

llvm-svn: 274894
This commit is contained in:
Jonathan Peyton 2016-07-08 17:43:21 +00:00
parent 50d307680f
commit 4d3c21307c
3 changed files with 16 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
openmp/runtime/src/kmp.h
View File

@ -1706,6 +1706,12 @@ typedef struct dispatch_shared_info {
volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1) volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
kmp_int32 doacross_num_done; // count finished threads kmp_int32 doacross_num_done; // count finished threads
#endif #endif
#if KMP_USE_HWLOC
// When linking with libhwloc, the ORDERED EPCC test slows down on big
// machines (> 48 cores). Performance analysis showed that a cache thrash
// was occurring and this padding helps alleviate the problem.
char padding[64];
#endif
} dispatch_shared_info_t; } dispatch_shared_info_t;
typedef struct kmp_disp { typedef struct kmp_disp {
@ -2567,7 +2573,7 @@ typedef struct KMP_ALIGN_CACHE kmp_base_team {
int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via omp_set_num_threads() call int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via omp_set_num_threads() call
// Read/write by workers as well ----------------------------------------------------------------------- // Read/write by workers as well -----------------------------------------------------------------------
#if KMP_ARCH_X86 || KMP_ARCH_X86_64 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && !KMP_USE_HWLOC
// Using CACHE_LINE=64 reduces memory footprint, but causes a big perf regression of epcc 'parallel' // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf regression of epcc 'parallel'
// and 'barrier' on fxe256lin01. This extra padding serves to fix the performance of epcc 'parallel' // and 'barrier' on fxe256lin01. This extra padding serves to fix the performance of epcc 'parallel'
// and 'barrier' when CACHE_LINE=64. TODO: investigate more and get rid if this padding. // and 'barrier' when CACHE_LINE=64. TODO: investigate more and get rid if this padding.

6
openmp/runtime/src/kmp_dispatch.cpp
View File

@ -180,6 +180,12 @@ struct dispatch_shared_info_template {
kmp_uint32 *doacross_flags; // array of iteration flags (0/1) kmp_uint32 *doacross_flags; // array of iteration flags (0/1)
kmp_int32 doacross_num_done; // count finished threads kmp_int32 doacross_num_done; // count finished threads
#endif #endif
#if KMP_USE_HWLOC
// When linking with libhwloc, the ORDERED EPCC test slowsdown on big
// machines (> 48 cores). Performance analysis showed that a cache thrash
// was occurring and this padding helps alleviate the problem.
char padding[64];
#endif
}; };
/* ------------------------------------------------------------------------ */ /* ------------------------------------------------------------------------ */

4
openmp/runtime/src/kmp_wait_release.h
View File

@ -97,6 +97,7 @@ __kmp_wait_template(kmp_info_t *this_thr, C *flag, int final_spin
kmp_uint32 hibernate; kmp_uint32 hibernate;
int th_gtid; int th_gtid;
int tasks_completed = FALSE; int tasks_completed = FALSE;
int oversubscribed;
KMP_FSYNC_SPIN_INIT(spin, NULL); KMP_FSYNC_SPIN_INIT(spin, NULL);
if (flag->done_check()) { if (flag->done_check()) {
@ -166,6 +167,7 @@ __kmp_wait_template(kmp_info_t *this_thr, C *flag, int final_spin
hibernate - __kmp_global.g.g_time.dt.t_value)); hibernate - __kmp_global.g.g_time.dt.t_value));
} }
oversubscribed = (TCR_4(__kmp_nth) > __kmp_avail_proc);
KMP_MB(); KMP_MB();
// Main wait spin loop // Main wait spin loop
@ -201,7 +203,7 @@ __kmp_wait_template(kmp_info_t *this_thr, C *flag, int final_spin
} }
// If we are oversubscribed, or have waited a bit (and KMP_LIBRARY=throughput), then yield // If we are oversubscribed, or have waited a bit (and KMP_LIBRARY=throughput), then yield
KMP_YIELD(TCR_4(__kmp_nth) > __kmp_avail_proc); KMP_YIELD(oversubscribed);
// TODO: Should it be number of cores instead of thread contexts? Like: // TODO: Should it be number of cores instead of thread contexts? Like:
// KMP_YIELD(TCR_4(__kmp_nth) > __kmp_ncores); // KMP_YIELD(TCR_4(__kmp_nth) > __kmp_ncores);
// Need performance improvement data to make the change... // Need performance improvement data to make the change...