int main(int argc, char *argv[])
{
gaspi_rank_t nprocs, n;
gaspi_number_t max_groups, gsize;
gaspi_rank_t *partners;
TSUITE_INIT(argc, argv);
gaspi_group_max(&max_groups);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT(gaspi_proc_num(&nprocs));
ASSERT(gaspi_group_size(GASPI_GROUP_ALL, &gsize));
partners = malloc(gsize * sizeof(gaspi_rank_t));
ASSERT(gaspi_group_ranks(GASPI_GROUP_ALL, partners));
for(n = 0; n < gsize; n++)
{
assert(partners[n] == n);
}
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
return EXIT_SUCCESS;
}
int main(int argc, char* argv[])
{
int i, j;
gaspi_number_t gsize;
int comm_state = WORKING;
int num_failures = 0;
int timesteps = 0;
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT (gaspi_proc_rank(&myrank));
ASSERT (gaspi_proc_num(&numprocs));
read_params(argc, argv, ×teps, &numprocs_idle);
numprocs_working = numprocs - numprocs_idle;
numprocs_working_and_idle = numprocs_working + numprocs_idle;
gaspi_rank_t *comm_main_ranks = malloc( numprocs_idle * sizeof(gaspi_rank_t));
init_array_2(comm_main_ranks, numprocs_working);
/* contains info of all processes:
which are working(0), broken(1) and idle(2).
keeps updated all the time(iterations) */
int * status_processes = (int *) malloc(numprocs * sizeof(int));
init_array_3(status_processes, numprocs, WORKING);
for(i = numprocs-1, j=0; j < numprocs_idle;--i,++j)
{
status_processes[i] = IDLE; // putting last processes to IDLE
}
// ===== GASPI group creation =====
if(status_processes[myrank]==WORKING)
{
ASSERT(gaspi_group_create(&COMM_MAIN));
gaspi_number_t i;
for(i=0; i<numprocs; i++)
{
if(status_processes[i]==WORKING)
{
ASSERT(gaspi_group_add(COMM_MAIN, i));
ASSERT(gaspi_group_size(COMM_MAIN, &gsize));
}
}
ASSERT(gaspi_group_ranks (COMM_MAIN, comm_main_ranks));
ASSERT(gaspi_group_commit (COMM_MAIN, GASPI_BLOCK));
}
/* ====== Init a SYNC FLAGS Segment ====== */
/* used to communicate the WORKING, BROKEN, or FINISHED_WORK status between the working and idle processes. */
gaspi_size_t SYNC_global_mem_size;
SYNC_global_mem_size = numprocs * sizeof(int);
gaspi_pointer_t gm_ptr_sync=NULL;
ASSERT(init_segment (gm_seg_sync_flags_id, SYNC_global_mem_size));
ASSERT(gaspi_segment_ptr (gm_seg_sync_flags_id, &gm_ptr_sync));
int * sync_flags = (int *) gm_ptr_sync;
init_array_3(sync_flags, numprocs, WORKING);
/* ====== Init a health check write FLAGS Segment ====== */
/* This array is used to send the gaspi_write message write before health_chk routine,
which will then update the gaspi internal health vector */
gaspi_size_t health_chk_global_mem_size;
health_chk_global_mem_size = numprocs*sizeof(int);
gaspi_pointer_t gm_ptr_health_chk=NULL;
ASSERT(init_segment (gm_seg_health_chk_array_id, health_chk_global_mem_size));
ASSERT(gaspi_segment_ptr (gm_seg_health_chk_array_id, &gm_ptr_health_chk));
gaspi_state_vector_t health_vec = (gaspi_state_vector_t) malloc(numprocs);
ASSERT(gaspi_state_vec_get(health_vec));
gaspi_rank_t * avoid_list= (gaspi_rank_t *) malloc(numprocs * sizeof(gaspi_rank_t));
for(i = 0;i < numprocs; ++i)
avoid_list[i] = (gaspi_rank_t) 0;
gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK);
/* ===== TIME-STEP LOOP ===== */
if(status_processes[myrank]==IDLE)
{
/* IDLE processes remain in this loop */
while(1)
{
gaspi_printf("%d.", myrank);
if(sync_flags[0] == WORKING)
{
/* NO FAILURE REPORTED */
usleep(1000000);
}
if(sync_flags[0] == BROKEN)
{
/* FAILURE REPORTED */
gaspi_printf("myrank: %d Broken reported\n", myrank);
comm_state=BROKEN;
break;
}
if(sync_flags[0] == WORKFINISHED)
{
/* WORKFINISHED REPORTED */
gaspi_printf("myrank: %d WorkFinished reported\n", myrank);
comm_state = WORKFINISHED;
break;
}
}
}
int time_step;
for(time_step=1; time_step <= timesteps && comm_state!=WORKFINISHED; time_step++)
{
gaspi_printf("== time_step: %d ==\n", time_step);
if(comm_state==WORKING && status_processes[myrank]==WORKING)
{
gaspi_barrier(COMM_MAIN, GASPI_TIMEOUT_TIME);
sleep(1); // NOTE: this is the work section.
if(time_step == 5 && myrank== 1)
{
exit (-1);
}
}
if(time_step<timesteps )
{
send_global_msg_to_check_state(health_vec, avoid_list);
num_failures = check_comm_health(status_processes, health_vec);
gaspi_printf("%d NUM_FAILURES at timestep %d = %d\n", myrank, time_step, num_failures);
if( num_failures != 0 )
{
rescue_process = numprocs_working;
if(myrank==0)
{
// message the IDLE process
sync_flags[0]=BROKEN;
for(i = 0 ; i < num_failures ; ++i)
{
/* TODO: multiple failures at the same time. */
gaspi_printf("messaging rescue_process: %d\n", rescue_process);
ASSERT(gaspi_write(gm_seg_sync_flags_id, 0, rescue_process, gm_seg_sync_flags_id, 0, sizeof(int), 0, GASPI_BLOCK));
rescue_process++;
}
}
if(myrank==0 || myrank==rescue_process)
gaspi_printf("%d REPAIRING COMM_MAIN FLAG 1\n", myrank);
update_status_processes_array(status_processes, health_vec);
numprocs_working_and_idle = refresh_numprocs_working_and_idle(status_processes);
if(myrank != rescue_process)
{
ASSERT(gaspi_group_delete(COMM_MAIN));
ASSERT(recover());
}
ASSERT(gaspi_group_create(&COMM_MAIN_NEW));
for(i = 0; i < numprocs; i++)
{
if(status_processes[i]==WORKING)
{
ASSERT(gaspi_group_add(COMM_MAIN_NEW, i));
ASSERT(gaspi_group_size(COMM_MAIN_NEW, &gsize));
if(gsize == numprocs_working)
break;
}
}
gaspi_printf("%d: COMM_MAIN_NEW size is: %hi\n", myrank, gsize);
ASSERT(gaspi_group_commit (COMM_MAIN_NEW, GASPI_BLOCK));
init_array_2(comm_main_ranks, numprocs_working);
ASSERT(gaspi_group_ranks (COMM_MAIN_NEW, comm_main_ranks));
gaspi_printf("printing group_ranks_main: \n");
gaspi_printf_array(comm_main_ranks, numprocs_working);
comm_state = WORKING;
gaspi_printf("%d REPAIRING COMM_MAIN_NEW FLAG 2\n", myrank);
if(status_processes[myrank] == WORKING)
{
ASSERT(gaspi_barrier(COMM_MAIN_NEW, GASPI_BLOCK));
ASSERT(gaspi_barrier(COMM_MAIN_NEW, GASPI_BLOCK));
}
/* set things to work again */
COMM_MAIN = COMM_MAIN_NEW;
time_step = 5;
}
}
}
if(myrank == 0)
{
gaspi_printf("finished successfully\n");
}
gaspi_proc_term(10000);
return EXIT_SUCCESS;
}