pd.if.org Git - liblfds/blob - liblfds/liblfds7.1.0/test_and_benchmark/libbenchmark/src/libbenchmark_benchmarks_freelist_push1_then_pop1/libbenchmark_benchmarks_freelist_pthread_mutex_push1_then_pop1.c

   1 /***** includes *****/
   2 #include "libbenchmark_benchmarks_freelist_internal.h"
   3
   4 /***** structs *****/
   5 struct libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_per_thread_benchmark_state
   6 {
   7   lfds710_pal_uint_t
   8     operation_count;
   9 };
  10
  11
  12
  13
  14
  15 /****************************************************************************/
  16 void libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_init( struct libbenchmark_topology_state *ts,
  17                                                                     struct lfds710_list_aso_state *logical_processor_set,
  18                                                                     struct libshared_memory_state *ms,
  19                                                                     enum libbenchmark_topology_numa_mode numa_mode,
  20                                                                     struct libbenchmark_threadset_state *tsets )
  21 {
  22   lfds710_pal_uint_t
  23     loop,
  24     number_logical_processors,
  25     number_logical_processors_in_numa_node,
  26     largest_number_logical_processors_in_numa_node = 0;
  27
  28   struct lfds710_list_asu_element
  29     *lasue = NULL,
  30     *lasue_lp = NULL;
  31
  32   struct libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_per_thread_benchmark_state
  33     *ptbs;
  34
  35   struct libbenchmark_datastructure_freelist_pthread_mutex_element
  36     *fe;
  37
  38   struct libbenchmark_datastructure_freelist_pthread_mutex_state
  39     *fs = NULL;
  40
  41   struct libbenchmark_threadset_per_numa_state
  42     *pns,
  43     *largest_pns = NULL;
  44
  45   struct libbenchmark_threadset_per_thread_state
  46     *pts;
  47
  48   struct libbenchmark_topology_node_state
  49     *numa_node_for_lp;
  50
  51   LFDS710_PAL_ASSERT( ts != NULL );
  52   LFDS710_PAL_ASSERT( logical_processor_set != NULL );
  53   LFDS710_PAL_ASSERT( ms != NULL );
  54   // TRD : numa_mode can be any value in its range
  55   LFDS710_PAL_ASSERT( tsets != NULL );
  56
  57   libbenchmark_threadset_init( tsets, ts, logical_processor_set, ms, libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_thread, NULL );
  58
  59   switch( numa_mode )
  60   {
  61     case LIBBENCHMARK_TOPOLOGY_NUMA_MODE_SMP:
  62       fs = libshared_memory_alloc_from_most_free_space_node( ms, sizeof(struct libbenchmark_datastructure_freelist_pthread_mutex_state), LFDS710_PAL_ATOMIC_ISOLATION_IN_BYTES );
  63       libbenchmark_datastructure_freelist_pthread_mutex_init( fs, NULL );
  64       lfds710_list_aso_query( logical_processor_set, LFDS710_LIST_ASO_QUERY_GET_POTENTIALLY_INACCURATE_COUNT, NULL, (void *) &number_logical_processors );
  65       fe = libshared_memory_alloc_from_most_free_space_node( ms, sizeof(struct libbenchmark_datastructure_freelist_pthread_mutex_element) * number_logical_processors, LFDS710_PAL_ATOMIC_ISOLATION_IN_BYTES );
  66       for( loop = 0 ; loop < number_logical_processors ; loop++ )
  67         libbenchmark_datastructure_freelist_pthread_mutex_push( fs, &fe[loop] );
  68       while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_thread_states,lasue) )
  69       {
  70         pts = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue );
  71         ptbs = libshared_memory_alloc_from_most_free_space_node( ms, sizeof(struct libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_per_thread_benchmark_state), LFDS710_PAL_ATOMIC_ISOLATION_IN_BYTES );
  72         pts->users_per_thread_state = ptbs;
  73       }
  74     break;
  75
  76     case LIBBENCHMARK_TOPOLOGY_NUMA_MODE_NUMA:
  77       /* TRD : init the freelist from the NUMA node with most processors from the current set
  78                or, if equal threads, with lowest NUMA
  79                iterate over the NUMA node list
  80                for each NUMA node, allocate one freelist element per thread on that node
  81                and push those elements onto the stack
  82
  83                the loop over the threads, and give each one the freelist state as it's user state
  84       */
  85
  86       while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_numa_states,lasue) )
  87       {
  88         pns = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue );
  89
  90         lasue_lp = NULL;
  91         number_logical_processors_in_numa_node = 0;
  92
  93         while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_thread_states,lasue_lp) )
  94         {
  95           pts = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue_lp );
  96
  97           libbenchmark_topology_query( ts, LIBBENCHMARK_TOPOLOGY_QUERY_GET_NUMA_NODE_FOR_LOGICAL_PROCESSOR, pts->tns_lp, &numa_node_for_lp );
  98
  99           if( LIBBENCHMARK_TOPOLOGY_NODE_GET_NUMA_ID(*numa_node_for_lp) == pns->numa_node_id )
 100             number_logical_processors_in_numa_node++;
 101         }
 102
 103         if( number_logical_processors_in_numa_node > largest_number_logical_processors_in_numa_node )
 104           largest_pns = pns;
 105       }
 106
 107       fs = libshared_memory_alloc_from_specific_node( ms, largest_pns->numa_node_id, sizeof(struct libbenchmark_datastructure_freelist_pthread_mutex_state), LFDS710_PAL_ATOMIC_ISOLATION_IN_BYTES );
 108       libbenchmark_datastructure_freelist_pthread_mutex_init( fs, NULL );
 109
 110       lasue = NULL;
 111
 112       while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_numa_states,lasue) )
 113       {
 114         pns = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue );
 115
 116         /* TRD : for each NUMA node, figure out how many LPs in the current set are in that NUMA node
 117                  and allocate then the correct number of elements from this NUMA node (1 per LP)
 118         */
 119
 120         lasue_lp = NULL;
 121         number_logical_processors_in_numa_node = 0;
 122
 123         while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_thread_states,lasue_lp) )
 124         {
 125           pts = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue_lp );
 126
 127           libbenchmark_topology_query( ts, LIBBENCHMARK_TOPOLOGY_QUERY_GET_NUMA_NODE_FOR_LOGICAL_PROCESSOR, pts->tns_lp, &numa_node_for_lp );
 128
 129           if( LIBBENCHMARK_TOPOLOGY_NODE_GET_NUMA_ID(*numa_node_for_lp) == pns->numa_node_id )
 130             number_logical_processors_in_numa_node++;
 131         }
 132
 133         fe = libshared_memory_alloc_from_specific_node( ms, pns->numa_node_id, sizeof(struct libbenchmark_datastructure_freelist_pthread_mutex_element) * number_logical_processors_in_numa_node, LFDS710_PAL_ATOMIC_ISOLATION_IN_BYTES );
 134         for( loop = 0 ; loop < number_logical_processors_in_numa_node ; loop++ )
 135           libbenchmark_datastructure_freelist_pthread_mutex_push( fs, &fe[loop] );
 136       }
 137
 138       lasue = NULL;
 139
 140       while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_thread_states,lasue) )
 141       {
 142         pts = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue );
 143         ptbs = libshared_memory_alloc_from_specific_node( ms, largest_pns->numa_node_id, sizeof(struct libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_per_thread_benchmark_state), LFDS710_PAL_ATOMIC_ISOLATION_IN_BYTES );
 144         pts->users_per_thread_state = ptbs;
 145       }
 146     break;
 147
 148     case LIBBENCHMARK_TOPOLOGY_NUMA_MODE_NUMA_BUT_NOT_USED:
 149       /* TRD : freelist state in the NUMA node with most threads from the current set
 150                or, if equal threads, with lowest NUMA
 151                all elements alloced from that node as well
 152       */
 153
 154       while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_numa_states,lasue) )
 155       {
 156         pns = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue );
 157
 158         lasue_lp = NULL;
 159         number_logical_processors_in_numa_node = 0;
 160
 161         while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_thread_states,lasue_lp) )
 162         {
 163           pts = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue_lp );
 164
 165           libbenchmark_topology_query( ts, LIBBENCHMARK_TOPOLOGY_QUERY_GET_NUMA_NODE_FOR_LOGICAL_PROCESSOR, pts->tns_lp, &numa_node_for_lp );
 166
 167           if( LIBBENCHMARK_TOPOLOGY_NODE_GET_NUMA_ID(*numa_node_for_lp) == pns->numa_node_id )
 168             number_logical_processors_in_numa_node++;
 169         }
 170
 171         if( number_logical_processors_in_numa_node > largest_number_logical_processors_in_numa_node )
 172           largest_pns = pns;
 173       }
 174
 175       fs = libshared_memory_alloc_from_specific_node( ms, largest_pns->numa_node_id, sizeof(struct libbenchmark_datastructure_freelist_pthread_mutex_state), LFDS710_PAL_ATOMIC_ISOLATION_IN_BYTES );
 176       libbenchmark_datastructure_freelist_pthread_mutex_init( fs, NULL );
 177
 178       lfds710_list_aso_query( logical_processor_set, LFDS710_LIST_ASO_QUERY_GET_POTENTIALLY_INACCURATE_COUNT, NULL, (void *) &number_logical_processors );
 179       fe = libshared_memory_alloc_from_specific_node( ms, largest_pns->numa_node_id, sizeof(struct libbenchmark_datastructure_freelist_pthread_mutex_element) * number_logical_processors, LFDS710_PAL_ATOMIC_ISOLATION_IN_BYTES );
 180       for( loop = 0 ; loop < number_logical_processors ; loop++ )
 181         libbenchmark_datastructure_freelist_pthread_mutex_push( fs, &fe[loop] );
 182
 183       lasue = NULL;
 184
 185       while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_thread_states,lasue) )
 186       {
 187         pts = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue );
 188         ptbs = libshared_memory_alloc_from_specific_node( ms, largest_pns->numa_node_id, sizeof(struct libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_per_thread_benchmark_state), LFDS710_PAL_ATOMIC_ISOLATION_IN_BYTES );
 189         pts->users_per_thread_state = ptbs;
 190       }
 191     break;
 192   }
 193
 194   tsets->users_threadset_state = fs;
 195
 196   return;
 197 }
 198
 199
 200
 201
 202
 203 /****************************************************************************/
 204 libshared_pal_thread_return_t LIBSHARED_PAL_THREAD_CALLING_CONVENTION libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_thread( void *libbenchmark_threadset_per_thread_state )
 205 {
 206   int long long unsigned
 207     current_time = 0,
 208     end_time,
 209     time_units_per_second;
 210
 211   lfds710_pal_uint_t
 212     operation_count = 0,
 213     time_loop = 0;
 214
 215   struct libbenchmark_datastructure_freelist_pthread_mutex_state
 216     *fs;
 217
 218   struct libbenchmark_datastructure_freelist_pthread_mutex_element
 219     *fe;
 220
 221   struct libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_per_thread_benchmark_state
 222     *ptbs;
 223
 224   struct libbenchmark_threadset_per_thread_state
 225     *pts;
 226
 227   LFDS710_MISC_BARRIER_LOAD;
 228
 229   LFDS710_PAL_ASSERT( libbenchmark_threadset_per_thread_state != NULL );
 230
 231   pts = (struct libbenchmark_threadset_per_thread_state *) libbenchmark_threadset_per_thread_state;
 232
 233   ptbs = LIBBENCHMARK_THREADSET_PER_THREAD_STATE_GET_USERS_PER_THREAD_STATE( *pts );
 234   fs = LIBBENCHMARK_THREADSET_PER_THREAD_STATE_GET_USERS_OVERALL_STATE( *pts );
 235
 236   LIBBENCHMARK_PAL_TIME_UNITS_PER_SECOND( &time_units_per_second );
 237
 238   libbenchmark_threadset_thread_ready_and_wait( pts );
 239
 240   LIBBENCHMARK_PAL_GET_HIGHRES_TIME( &current_time );
 241
 242   end_time = current_time + time_units_per_second * libbenchmark_globals_benchmark_duration_in_seconds;
 243
 244   while( current_time < end_time )
 245   {
 246     libbenchmark_datastructure_freelist_pthread_mutex_pop( fs, &fe );
 247     libbenchmark_datastructure_freelist_pthread_mutex_push( fs, fe );
 248     operation_count++;
 249
 250     if( time_loop++ == TIME_LOOP_COUNT )
 251     {
 252       time_loop = 0;
 253       LIBBENCHMARK_PAL_GET_HIGHRES_TIME( &current_time );
 254     }
 255   }
 256
 257   ptbs->operation_count = operation_count;
 258
 259   LFDS710_MISC_BARRIER_STORE;
 260
 261   lfds710_misc_force_store();
 262
 263   return LIBSHARED_PAL_THREAD_RETURN_CAST(RETURN_SUCCESS);
 264 }
 265
 266
 267
 268
 269
 270 /****************************************************************************/
 271 void libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_cleanup( struct lfds710_list_aso_state *logical_processor_set,
 272                                                                        enum libbenchmark_topology_numa_mode numa_mode,
 273                                                                        struct libbenchmark_results_state *rs,
 274                                                                        struct libbenchmark_threadset_state *tsets )
 275 {
 276   struct libbenchmark_datastructure_freelist_pthread_mutex_state
 277     *fs;
 278
 279   struct lfds710_list_asu_element
 280     *lasue = NULL;
 281
 282   struct libbenchmark_benchmark_freelist_pthread_mutex_push1_pop1_per_thread_benchmark_state
 283     *ptbs;
 284
 285   struct libbenchmark_threadset_per_thread_state
 286     *pts;
 287
 288   LFDS710_PAL_ASSERT( logical_processor_set != NULL );
 289   // TRD : numa_mode can be any value in its range
 290   LFDS710_PAL_ASSERT( rs != NULL );
 291   LFDS710_PAL_ASSERT( tsets != NULL );
 292
 293   while( LFDS710_LIST_ASU_GET_START_AND_THEN_NEXT(tsets->list_of_per_thread_states,lasue) )
 294   {
 295     pts = LFDS710_LIST_ASU_GET_VALUE_FROM_ELEMENT( *lasue );
 296
 297     ptbs = LIBBENCHMARK_THREADSET_PER_THREAD_STATE_GET_USERS_PER_THREAD_STATE( *pts );
 298
 299     libbenchmark_results_put_result( rs,
 300                                      LIBBENCHMARK_DATASTRUCTURE_ID_FREELIST,
 301                                      LIBBENCHMARK_BENCHMARK_ID_PUSH1_THEN_POP1,
 302                                      LIBBENCHMARK_LOCK_ID_PTHREAD_MUTEX,
 303                                      numa_mode,
 304                                      logical_processor_set,
 305                                      LIBBENCHMARK_TOPOLOGY_NODE_GET_LOGICAL_PROCESSOR_NUMBER( *pts->tns_lp ),
 306                                      LIBBENCHMARK_TOPOLOGY_NODE_GET_WINDOWS_GROUP_NUMBER( *pts->tns_lp ),
 307                                      ptbs->operation_count );
 308   }
 309
 310   fs = tsets->users_threadset_state;
 311
 312   libbenchmark_datastructure_freelist_pthread_mutex_cleanup( fs, NULL );
 313
 314   return;
 315 }
 316