From e7d7d4a985338bffc0574da0889c604489a84c13 Mon Sep 17 00:00:00 2001 From: etortilopez Date: Wed, 1 Oct 2014 11:05:29 +0200 Subject: [PATCH] Fixes in adaptive algorithm --- src/simix/smx_context_raw.c | 111 ++++++++++++++++++++++++++++-------- src/simix/smx_private.h | 7 ++- 2 files changed, 91 insertions(+), 27 deletions(-) diff --git a/src/simix/smx_context_raw.c b/src/simix/smx_context_raw.c index 89e19b44d9..23b08dfd5e 100644 --- a/src/simix/smx_context_raw.c +++ b/src/simix/smx_context_raw.c @@ -33,6 +33,7 @@ static xbt_os_thread_key_t raw_worker_id_key; /* thread-specific storage for the #define SCHED_ROUND_LIMIT 5 xbt_os_timer_t round_time; double par_time,seq_time; +double par_ratio,seq_ratio; static unsigned int par_proc_that_ran = 0,seq_proc_that_ran = 0; /* Counters of processes that have run in SCHED_ROUND_LIMIT scheduling rounds */ static unsigned int seq_sched_round, par_sched_round; /* Amount of SR that ran serial/parallel*/ #endif @@ -210,6 +211,15 @@ static unsigned int ssr_count = 0; static char new_sr = 0; #endif +#ifdef TIME_BENCH_ENTIRE_SRS +static unsigned int sr_count = 0; +static xbt_os_timer_t timer; +#endif + +#ifdef ADAPTIVE_THRESHOLD +int reached_seq_limit, reached_par_limit; +#endif + static void smx_ctx_raw_wrapper(smx_ctx_raw_t context); static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory); static smx_context_t smx_ctx_raw_create_context(xbt_main_func_t code, int argc, @@ -272,6 +282,18 @@ void SIMIX_ctx_raw_factory_init(smx_context_factory_t *factory) (*factory)->runall = smx_ctx_raw_runall_serial; (*factory)->suspend = smx_ctx_raw_suspend_serial; } +#ifdef TIME_BENCH_ENTIRE_SRS + (*factory)->runall = smx_ctx_raw_runall; + (*factory)->suspend = NULL; + timer = xbt_os_timer_new(); +#endif + +#ifdef ADAPTIVE_THRESHOLD + round_time = xbt_os_timer_new(); + reached_seq_limit = 0; + reached_par_limit = 0; +#endif + #ifdef TIME_BENCH_PER_SR timer = xbt_os_timer_new(); #endif @@ -582,19 +604,30 @@ static void smx_ctx_raw_runall_parallel(void) static void smx_ctx_raw_runall(void) { unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run); + reached_seq_limit = (seq_sched_round % SCHED_ROUND_LIMIT == 0); + reached_par_limit = (par_sched_round % SCHED_ROUND_LIMIT == 0); - if(seq_sched_round % SCHED_ROUND_LIMIT == 0 && par_sched_round % SCHED_ROUND_LIMIT == 0){ - seq_sched_round = 1; - par_sched_round = 1; - if((seq_time / (double)seq_proc_that_ran) > (par_time / (double)par_proc_that_ran)){ + if(reached_par_limit){ + par_sched_round = 0; + par_ratio = (par_proc_that_ran != 0) ? (par_time / (double)par_proc_that_ran) : 0; + par_time = 0; par_proc_that_ran = 0; + } + + if(reached_seq_limit){ + seq_sched_round = 0; + seq_ratio = (seq_proc_that_ran != 0) ? (seq_time / (double)seq_proc_that_ran) : 0; + seq_time = 0; seq_proc_that_ran = 0; + } + + if(reached_seq_limit && reached_par_limit){ + if(seq_ratio > par_ratio){ SIMIX_context_set_parallel_threshold(SIMIX_context_get_parallel_threshold() - 1); } else { SIMIX_context_set_parallel_threshold(SIMIX_context_get_parallel_threshold() + 1); } - par_time = 0; par_proc_that_ran = 0; - seq_time = 0; seq_proc_that_ran = 0; } - round_time = xbt_os_timer_new(); + + XBT_CRITICAL("Adaptive Algorithm. Parallel Threshold is: %d. Processes: %d", SIMIX_context_get_parallel_threshold(), nb_processes); if (nb_processes >= SIMIX_context_get_parallel_threshold()) { XBT_DEBUG("Runall // %lu", nb_processes); simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel; @@ -602,8 +635,8 @@ static void smx_ctx_raw_runall(void) smx_ctx_raw_runall_parallel(); xbt_os_cputimer_stop(round_time); par_time += xbt_os_timer_elapsed(round_time); - par_proc_that_ran += xbt_dynar_length(simix_global->process_that_ran); - par_sched_round += 1; + par_proc_that_ran += nb_processes; + par_sched_round++; } else { XBT_DEBUG("Runall serial %lu", nb_processes); simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial; @@ -615,8 +648,8 @@ static void smx_ctx_raw_runall(void) #endif xbt_os_cputimer_stop(round_time); seq_time += xbt_os_timer_elapsed(round_time); - seq_proc_that_ran += xbt_dynar_length(simix_global->process_that_ran); - seq_sched_round += 1; + seq_proc_that_ran += nb_processes; + seq_sched_round++; } } @@ -624,19 +657,49 @@ static void smx_ctx_raw_runall(void) static void smx_ctx_raw_runall(void) { +#ifdef TIME_BENCH_ENTIRE_SRS + sr_count++; + timer = xbt_os_timer_new(); + double elapsed = 0; +#endif unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run); - if (nb_processes >= SIMIX_context_get_parallel_threshold()) { - XBT_DEBUG("Runall // %lu", nb_processes); - simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel; - smx_ctx_raw_runall_parallel(); - } else { - XBT_DEBUG("Runall serial %lu", nb_processes); - simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial; - #ifdef TIME_BENCH_PER_SR - smx_ctx_raw_runall_serial(simix_global->process_to_run); - #else - smx_ctx_raw_runall_serial(); - #endif - } + if (SIMIX_context_is_parallel() && SIMIX_context_get_parallel_threshold()context_factory->suspend = smx_ctx_raw_suspend_parallel; + + #ifdef TIME_BENCH_ENTIRE_SRS + xbt_os_cputimer_start(timer); + #endif + + smx_ctx_raw_runall_parallel(); + + #ifdef TIME_BENCH_ENTIRE_SRS + xbt_os_cputimer_stop(timer); + elapsed = xbt_os_timer_elapsed(timer); + #endif + } else { + XBT_DEBUG("Runall serial %lu", nb_processes); + simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial; + + #ifdef TIME_BENCH_PER_SR + smx_ctx_raw_runall_serial(simix_global->process_to_run); + #else + + #ifdef TIME_BENCH_ENTIRE_SRS + xbt_os_cputimer_start(timer); + #endif + + smx_ctx_raw_runall_serial(); + + #ifdef TIME_BENCH_ENTIRE_SRS + xbt_os_cputimer_stop(timer); + elapsed = xbt_os_timer_elapsed(timer); + #endif + #endif + } + +#ifdef TIME_BENCH_ENTIRE_SRS + XBT_CRITICAL("Total time SR %u = %f, %d", sr_count, elapsed, nb_processes); +#endif } #endif diff --git a/src/simix/smx_private.h b/src/simix/smx_private.h index 11fe9e7614..b061c5da59 100644 --- a/src/simix/smx_private.h +++ b/src/simix/smx_private.h @@ -30,9 +30,10 @@ #include "smx_new_api_private.h" /* Define only for SimGrid benchmarking purposes */ -//#define TIME_BENCH_PER_SR /* this aims at measuring the time spent in each scheduling round per each thread. The code is thus run in sequential to bench separately each SSR */ -//#define TIME_BENCH_AMDAHL /* this aims at measuring the porting of time that could be parallelized at maximum (to get the optimal speedup by applying the amdahl law). */ -//#define ADAPTIVE_THRESHOLD /* this is to enable the adaptive threshold algorithm in raw contexts*/ +//#define TIME_BENCH_PER_SR /* this aims at measuring the time spent in each scheduling round per each thread. The code is thus run in sequential to bench separately each SSR */ +//#define TIME_BENCH_AMDAHL /* this aims at measuring the porting of time that could be parallelized at maximum (to get the optimal speedup by applying the amdahl law). */ +//#define ADAPTIVE_THRESHOLD /* this is to enable the adaptive threshold algorithm in raw contexts*/ +//#define TIME_BENCH_ENTIRE_SRS /* more general benchmark than TIME_BENCH_PER_SR. It aims to measure the total time spent in a whole scheduling round (including synchro costs)*/ #ifdef TIME_BENCH_PER_SR void smx_ctx_raw_new_sr(void); -- 2.20.1