/**
* \brief Initialize the PAMI message layer.
*
* This gets the task ID, number of tasks, number of contexts, and
* checks that each process has the same max number of contexts. It
* then creates the contexts initializes the dispatch tables.
*/
static int init()
{
pami_configuration_t query;
pami_dispatch_hint_t options;
//bzero(&options, sizeof(pami_dispatch_hint_t));
memset(&options, '\0', sizeof(pami_dispatch_hint_t));
options.consistency = PAMI_HINT_ENABLE;
PAMI_Client_create("TEST", &client, NULL, 0);
query.name = PAMI_CLIENT_TASK_ID;
PAMI_Client_query (client, &query,1);
task = query.value.intval;
query.name = PAMI_CLIENT_NUM_TASKS;
PAMI_Client_query (client, &query,1);
size = query.value.intval;
if(size <= 1) return 0;
query.name = PAMI_CLIENT_NUM_CONTEXTS;
PAMI_Client_query(client, &query,1);
num_contexts = query.value.intval;
//assert(num_contexts <= MAX_CONTEXTS);
//assert((num_contexts&(num_contexts-1)) == 0);
//query.name = PAMI_CLIENT_CONST_CONTEXTS;
//PAMI_Client_query (client, &query,1);
//assert(query.value.intval);
query.value.intval = 1;
PAMI_Context_createv(client, &query, 1, contexts, num_contexts);
pami_dispatch_callback_function RecvShortFN;
RecvShortFN.p2p = RecvShortCB;
pami_dispatch_callback_function RecvLongFN;
RecvLongFN.p2p = RecvLongCB;
/*
* The context index is used for the dispatch cookie. This allows
* the call-back to identify the context index in use.
*/
size_t i;
for (i=0; i<num_contexts; ++i) {
PAMI_Dispatch_set(contexts[i],
SHORT_DISPATCH,
RecvShortFN,
(void*)i,
options);
PAMI_Dispatch_set(contexts[i],
LONG_DISPATCH,
RecvLongFN,
(void*)i,
options);
}
printf("Task=%zu Size=%zu <PAMI Initialized> thread-level=%d\n", task, size, 13);
return 1;
}
int main ()
{
pami_client_t client;
pami_context_t *context;
pami_result_t result;
pami_configuration_t configuration;
PAMI_Client_create ("TEST", &client, NULL, 0);
configuration.name = PAMI_CLIENT_NUM_CONTEXTS;
result = PAMI_Client_query(client, &configuration, 1);
size_t num = configuration.value.intval;
context = (pami_context_t*) malloc (num*sizeof(pami_context_t));
/* Create four contexts - every task creates the same number */
PAMI_Context_createv (client, NULL, 0, context, num);
createEndpointTable (client, num);
pami_dispatch_callback_function fn;
fn.p2p = test_dispatch;
pami_dispatch_hint_t options = {0};
pami_send_hint_t hints = {0};
volatile size_t expect = 0;
size_t i;
for (i=0; i<num; i++)
{
PAMI_Context_lock (context[i]);
result = PAMI_Dispatch_set (context[i], 0, fn, (void *)&expect, options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami dispatch. result = %d\n", result);
return 1;
}
}
configuration.name = PAMI_CLIENT_TASK_ID;
result = PAMI_Client_query(client, &configuration, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
pami_task_t task_id = configuration.value.intval;
fprintf (stderr, "My task id = %d\n", task_id);
configuration.name = PAMI_CLIENT_NUM_TASKS;
result = PAMI_Client_query(client, &configuration, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
size_t num_tasks = configuration.value.intval;
fprintf (stderr, "Number of tasks = %zu\n", num_tasks);
uint8_t header[16];
uint8_t data[1024];
volatile size_t active = 1;
pami_send_t parameters;
parameters.send.dispatch = 0;
parameters.send.header.iov_base = header;
parameters.send.header.iov_len = 16;
parameters.send.data.iov_base = data;
parameters.send.data.iov_len = 1024;
parameters.send.hints = hints;
parameters.events.cookie = (void *) &active;
parameters.events.local_fn = decrement;
parameters.events.remote_fn = NULL;
/* Send a message to endpoint "num_tasks * num_contexts - 1" */
pami_task_t target_task = (pami_task_t) -1;
size_t target_offset = (size_t) -1;
PAMI_Endpoint_query (_endpoint[num*num_tasks-1], &target_task, &target_offset);
if (task_id == target_task) expect += num_tasks;
send_endpoint (context[0], num*num_tasks-1, ¶meters);
fprintf (stdout, "before advance, active = %zu, expect = %zu\n", active, expect);
while ((active + expect) > 0) PAMI_Context_advancev (context, num, 100);
for (i=0; i<num; i++) PAMI_Context_unlock (context[i]);
result = PAMI_Context_destroyv (context, num);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami context. result = %d\n", result);
return 1;
}
result = PAMI_Client_destroy (&client);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to finalize pami client. result = %d\n", result);
return 1;
}
fprintf (stdout, "Success (%d)\n", task_id);
return 0;
};
int main(int argc, char **argv)
{
MPI_Init(&argc, &argv);
optiq_pami_transport_init();
struct optiq_pami_transport *pami_transport = optiq_pami_transport_get();
int rank = pami_transport->rank;
int local_rank = 1;
int imm_bytes = 128;
void *local_buf = malloc (imm_bytes);
for (int i = 0; i < imm_bytes; i++) {
((char*)local_buf)[i] = i % 128;
}
int remote_rank = 3;
int iters = 30;
int cookie = iters;
pami_dispatch_callback_function fn;
pami_dispatch_hint_t options = {};
/*Receive memory request*/
fn.p2p = optiq_recv_test_fn;
pami_result_t result = PAMI_Dispatch_set (pami_transport->context,
OPTIQ_TEST_PAMI_IMM,
fn,
(void *) &cookie,
options);
assert(result == PAMI_SUCCESS);
if (result != PAMI_SUCCESS) {
return 0;
}
for (int nbytes = 1; nbytes <= imm_bytes; nbytes++)
{
MPI_Barrier(MPI_COMM_WORLD);
uint64_t t0 = GetTimeBase();
if (rank == local_rank)
{
for (int i = 0; i < iters; i++) {
optiq_pami_send_immediate(pami_transport->context, OPTIQ_TEST_PAMI_IMM, NULL, 0, local_buf, nbytes, pami_transport->endpoints[remote_rank]);
}
while (cookie > 0) {
PAMI_Context_advance (pami_transport->context, 100);
}
}
if (rank == remote_rank)
{
while (cookie > 0) {
PAMI_Context_advance (pami_transport->context, 100);
}
for (int i = 0; i < iters; i++) {
optiq_pami_send_immediate(pami_transport->context, OPTIQ_TEST_PAMI_IMM, NULL, 0, local_buf, nbytes, pami_transport->endpoints[local_rank]);
}
}
cookie = iters;
uint64_t t1 = GetTimeBase();
double max_t, t = (double)(t1 - t0)/1.6e3/iters;
MPI_Reduce(&t, &max_t, 1, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
if (pami_transport->rank == 0) {
max_t = max_t/2;
double bw = (double)nbytes/1024/1024/max_t*1e6;
printf("nbytes = %d t = %8.4f(us) bw = %8.4f(MB/s)\n", nbytes, max_t, bw);
}
}
free(local_buf);
optiq_pami_transport_finalize();
return 0;
}
int main (int argc, char ** argv)
{
/*volatile size_t send_active = 2; */
volatile size_t send_active = 1;
volatile size_t recv_active = 1;
pami_client_t client;
pami_context_t context;
char cl_string[] = "TEST";
pami_result_t result = PAMI_ERROR;
fprintf (stderr, "Before Client initialize\n");
result = PAMI_Client_create (cl_string, &client, NULL, 0);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to initialize pami client. result = %d\n", result);
return 1;
}
fprintf (stderr, "After Client initialize\n");
fprintf (stderr, "before context createv\n");
{ result = PAMI_Context_createv(client, NULL, 0, &context, 1); }
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami context. result = %d\n", result);
return 1;
}
fprintf (stderr, "after context createv\n");
pami_configuration_t configuration;
configuration.name = PAMI_CLIENT_TASK_ID;
result = PAMI_Client_query(client, &configuration,1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
size_t task_id = configuration.value.intval;
fprintf (stderr, "My task id = %zu\n", task_id);
configuration.name = PAMI_CLIENT_NUM_TASKS;
result = PAMI_Client_query(client, &configuration,1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
/*size_t num_tasks = configuration.value.intval; */
size_t dispatch = 0;
pami_dispatch_callback_function fn;
fn.p2p = test_dispatch;
pami_dispatch_hint_t options={};
fprintf (stderr, "Before PAMI_Dispatch_set() .. &recv_active = %p, recv_active = %zu\n", &recv_active, recv_active);
result = PAMI_Dispatch_set (context,
dispatch,
fn,
(void *)&recv_active,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami dispatch. result = %d\n", result);
return 1;
}
pami_send_t parameters;
parameters.send.dispatch = dispatch;
parameters.send.header.iov_base = (void *)&dispatch; /* send *something* */
parameters.send.header.iov_len = sizeof(size_t);
parameters.send.data.iov_base = (void *)&dispatch; /* send *something* */
parameters.send.data.iov_len = sizeof(size_t);
parameters.events.cookie = (void *) &send_active;
parameters.events.local_fn = send_done_local;
/*parameters.events.remote_fn = send_done_remote; */
parameters.events.remote_fn = NULL;
#if 1
int iter;
for (iter=0; iter < 100; iter++)
{
fprintf (stderr, "before send ...\n");
PAMI_Endpoint_create (client, task_id, 0, ¶meters.send.dest);
result = PAMI_Send (context, ¶meters);
fprintf (stderr, "... after send.\n");
fprintf (stderr, "before send-recv advance loop (send_active = %zu, recv_active = %zu) ...\n", send_active, recv_active);
while (send_active || recv_active)
{
result = PAMI_Context_advance (context, 100);
if ( (result != PAMI_SUCCESS) && (result != PAMI_EAGAIN) )
{
fprintf (stderr, "Error. Unable to advance pami context. result = %d\n", result);
return 1;
}
}
fprintf (stderr, "... after send-recv advance loop\n");
send_active = recv_active = 1;
}
#endif
result = PAMI_Context_destroyv(&context, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami context. result = %d\n", result);
return 1;
}
result = PAMI_Client_destroy(&client);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to finalize pami client. result = %d\n", result);
return 1;
}
return 0;
};
int main (int argc, char ** argv)
{
volatile size_t _rts_active = 1;
volatile size_t _ack_active = 1;
memset(&null_send_hint, 0, sizeof(null_send_hint));
pami_client_t client;
pami_context_t context[2];
char cl_string[] = "TEST";
pami_result_t result = PAMI_ERROR;
result = PAMI_Client_create (cl_string, &client, NULL, 0);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami client. result = %d\n", result);
return 1;
}
#ifdef TEST_CROSSTALK
size_t num = 2;
#else
size_t num = 1;
#endif
result = PAMI_Context_createv(client, NULL, 0, context, num);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami context(s). result = %d\n", result);
return 1;
}
pami_configuration_t configuration;
configuration.name = PAMI_CLIENT_TASK_ID;
result = PAMI_Client_query(client, &configuration,1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
pami_task_t task_id = configuration.value.intval;
fprintf (stderr, "My task id = %d\n", task_id);
configuration.name = PAMI_CLIENT_NUM_TASKS;
result = PAMI_Client_query(client, &configuration,1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
size_t num_tasks = configuration.value.intval;
fprintf (stderr, "Number of tasks = %zu\n", num_tasks);
if (num_tasks < 2)
{
fprintf (stderr, "Error. This test requires at least 2 tasks. Number of tasks in this job: %zu\n", num_tasks);
return 1;
}
pami_dispatch_hint_t options={};
#ifdef USE_SHMEM_OPTION
options.use_shmem = PAMI_HINT_ENABLE;
fprintf (stderr, "##########################################\n");
fprintf (stderr, "shared memory optimizations forced ON\n");
fprintf (stderr, "##########################################\n");
#elif defined(NO_SHMEM_OPTION)
options.use_shmem = PAMI_HINT_DISABLE;
fprintf (stderr, "##########################################\n");
fprintf (stderr, "shared memory optimizations forced OFF\n");
fprintf (stderr, "##########################################\n");
#endif
size_t i = 0;
#ifdef TEST_CROSSTALK
for (i=0; i<2; i++)
#endif
{
pami_dispatch_callback_function fn;
fprintf (stderr, "Before PAMI_Dispatch_set(%d) .. &_rts_active = %p, _rts_active = %zu\n", DISPATCH_ID_RTS, &_rts_active, _rts_active);
fn.p2p = dispatch_rts;
result = PAMI_Dispatch_set (context[i],
DISPATCH_ID_RTS,
fn,
(void *)&_rts_active,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami dispatch. result = %d\n", result);
return 1;
}
fprintf (stderr, "Before PAMI_Dispatch_set(%d) .. &_ack_active = %p, _ack_active = %zu\n", DISPATCH_ID_ACK, &_ack_active, _ack_active);
fn.p2p = dispatch_ack;
result = PAMI_Dispatch_set (context[i],
DISPATCH_ID_ACK,
fn,
(void *)&_ack_active,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami dispatch. result = %d\n", result);
return 1;
}
}
if (task_id == 0)
{
pami_send_immediate_t parameters;
#ifdef TEST_CROSSTALK
fprintf (stdout, "PAMI_Rget('simple') functional test [crosstalk]\n");
fprintf (stdout, "\n");
PAMI_Endpoint_create (client, num_tasks-1, 1, ¶meters.dest);
#else
fprintf (stdout, "PAMI_Rget('simple') functional test\n");
fprintf (stdout, "\n");
PAMI_Endpoint_create (client, num_tasks-1, 0, ¶meters.dest);
#endif
/* Allocate some memory from the heap. */
void * send_buffer = malloc (BUFFERSIZE);
/* Initialize the memory for validation. */
initialize_data ((uint32_t *)send_buffer, BUFFERSIZE, 0);
print_data (send_buffer, BUFFERSIZE);
/* Send an 'rts' message to the target task and provide the memory region */
rts_info_t rts_info;
PAMI_Endpoint_create (client, 0, 0, &rts_info.origin);
rts_info.bytes = BUFFERSIZE;
/* Create a memory region for this memoru buffer */
size_t bytes = 0;
pami_result_t pami_rc = PAMI_Memregion_create (context[0], send_buffer, BUFFERSIZE, &bytes, &(rts_info.memregion));
if (PAMI_SUCCESS != pami_rc) {
fprintf (stderr, "PAMI_Memregion_create failed with rc = %d\n", pami_rc) ;
exit(1);
}
parameters.dispatch = DISPATCH_ID_RTS;
parameters.header.iov_base = &rts_info;
parameters.header.iov_len = sizeof(rts_info_t);
parameters.data.iov_base = NULL;
parameters.data.iov_len = 0;
fprintf (stderr, "Before PAMI_Send_immediate()\n");
PAMI_Send_immediate (context[0], ¶meters);
/* wait for the 'ack' */
fprintf (stderr, "Wait for 'ack', _ack_active = %zu\n", _ack_active);
while (_ack_active != 0)
{
result = PAMI_Context_advance (context[0], 100);
if (result != PAMI_SUCCESS && result != PAMI_EAGAIN)
{
fprintf (stderr, "Error. Unable to advance pami context. result = %d\n", result);
return 1;
}
}
/* Destroy the local memory region */
PAMI_Memregion_destroy (context[0], &(rts_info.memregion));
free (send_buffer);
switch (_ack_status)
{
case 0:
fprintf (stdout, "Test PASSED\n");
break;
case 1:
fprintf (stdout, "Test FAILED (rget error)\n");
break;
case 2:
fprintf (stdout, "Test FAILED (data error)\n");
break;
default:
fprintf (stdout, "Test FAILED (unknown error)\n");
break;
}
}
else if (task_id == num_tasks-1)
{
#ifdef TEST_CROSSTALK
size_t contextid = 1;
#else
size_t contextid = 0;
#endif
/* wait for the 'rts' */
fprintf (stderr, "Wait for 'rts', _rts_active = %zu, contextid = %zu\n", _rts_active, contextid);
while (_rts_active != 0)
{
result = PAMI_Context_advance (context[contextid], 100);
if (result != PAMI_SUCCESS && result != PAMI_EAGAIN)
{
fprintf (stderr, "Error. Unable to advance pami context. result = %d\n", result);
return 1;
}
}
}
fprintf (stderr, "Test completed .. cleanup\n");
result = PAMI_Context_destroyv (context, num);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami context. result = %d\n", result);
return 1;
}
result = PAMI_Client_destroy(&client);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami client. result = %d\n", result);
return 1;
}
/*fprintf (stdout, "Success (%d)\n", task_id); */
return 0;
};
void test_fn (int argc, char * argv[], pami_client_t client, pami_context_t context[])
{
int num_doubles = 16;
if (argc > 1) num_doubles = atoi (argv[1]);
size_t num_tasks = size (client);
pami_task_t my_task_id = task (client);
pami_task_t target_task_id = num_tasks - 1;
pami_task_t origin_task_id = 0;
/*
* Allocate a 'window' of memory region information, one for each task in the
* client. Only the 'local' memory region information for this task will
* contain a valid data buffer. The memory region information is marked
* 'active' when the memory regions are received from each remote task.
*/
memregion_information_t * mr_info =
(memregion_information_t *) malloc (sizeof(memregion_information_t) * num_tasks);
unsigned i;
for (i = 0; i < num_tasks; i++)
{
mr_info[i].data.iov_len = 0;
mr_info[i].data.iov_base = NULL;
mr_info[i].active = 0;
}
/*
* Create a local memregion for each context.
*
* Note that both memregions will describe the same memory location. This is
* necessary when writing portable, platform independent code as the physical
* hardware underlying the contexts may, or may not, require separate memory
* pinning.
*/
size_t actual_memregion_bytes = 0;
mr_info[my_task_id].data.iov_base = malloc (sizeof(double) * num_doubles);
mr_info[my_task_id].data.iov_len = sizeof(double) * num_doubles;
PAMI_Memregion_create (context[0],
mr_info[my_task_id].data.iov_base,
mr_info[my_task_id].data.iov_len,
& actual_memregion_bytes,
& mr_info[my_task_id].memregion[0]);
PAMI_Memregion_create (context[1],
mr_info[my_task_id].data.iov_base,
mr_info[my_task_id].data.iov_len,
& actual_memregion_bytes,
& mr_info[my_task_id].memregion[1]);
mr_info[my_task_id].active = 1;
/*
* Register the memory region exchange dispatch; only needed on the
* first context of each task.
*/
pami_dispatch_hint_t mr_hint = {0};
pami_dispatch_callback_function mr_dispatch;
mr_dispatch.p2p = exchange_memregion_recv_cb;
PAMI_Dispatch_set (context[0], MEMREGION_EXCHANGE_DISPATCH_ID, mr_dispatch, (void *) mr_info, mr_hint);
accumulate_test_information_t test_info;
test_info.data_buffer.iov_base = malloc (sizeof(double) * num_doubles);
test_info.data_buffer.iov_len = sizeof(double) * num_doubles;
test_info.scalar = 1.2;
test_info.data_fn[ACCUMULATE_TEST_SCALAR_SUM] = accumulate_scalar_sum_data_function;
test_info.data_cookie[ACCUMULATE_TEST_SCALAR_SUM] = (void *) & test_info.scalar;
test_info.data_fn[ACCUMULATE_TEST_VECTOR_SUM] = accumulate_vector_sum_data_function;
test_info.data_cookie[ACCUMULATE_TEST_VECTOR_SUM] = malloc (sizeof(double) * num_doubles);
test_info.data_fn[ACCUMULATE_TEST_SCALAR_SUBTRACT] = accumulate_scalar_subtract_data_function;
test_info.data_cookie[ACCUMULATE_TEST_SCALAR_SUBTRACT] = (void *) & test_info.scalar;
test_info.data_fn[ACCUMULATE_TEST_VECTOR_SUBTRACT] = accumulate_vector_subtract_data_function;
test_info.data_cookie[ACCUMULATE_TEST_VECTOR_SUBTRACT] = malloc (sizeof(double) * num_doubles);
test_info.data_fn[ACCUMULATE_TEST_VECTOR_MAX_SUM] = accumulate_vector_max_sum_data_function;
test_info.data_cookie[ACCUMULATE_TEST_VECTOR_MAX_SUM] = malloc (sizeof(double) * num_doubles);
test_info.data_fn[ACCUMULATE_TEST_VECTOR_MIN_SUM] = accumulate_vector_min_sum_data_function;
test_info.data_cookie[ACCUMULATE_TEST_VECTOR_MIN_SUM] = malloc (sizeof(double) * num_doubles);
/*
* Register the accumulate dispatch; needed on both
* contexts to enable "crosstalk".
*/
pami_dispatch_hint_t acc_hint = {0};
acc_hint.recv_immediate = PAMI_HINT_DISABLE;
pami_dispatch_callback_function acc_dispatch;
acc_dispatch.p2p = accumulate_test_recv_cb;
PAMI_Dispatch_set (context[0], ACCUMULATE_TEST_DISPATCH_ID, acc_dispatch, (void *) & test_info, acc_hint);
PAMI_Dispatch_set (context[1], ACCUMULATE_TEST_DISPATCH_ID, acc_dispatch, (void *) & test_info, acc_hint);
simple_barrier(client, context[0]);
/*
* Exchange the memory regions
*/
volatile unsigned mr_exchange_active = 0;
pami_send_t mr_exchange_parameters = {0};
mr_exchange_parameters.send.dispatch = MEMREGION_EXCHANGE_DISPATCH_ID;
mr_exchange_parameters.send.header.iov_base = NULL;
mr_exchange_parameters.send.header.iov_len = 0;
mr_exchange_parameters.send.data.iov_base = (void *) mr_info[my_task_id].memregion;
mr_exchange_parameters.send.data.iov_len = sizeof(pami_memregion_t) * 2;
mr_exchange_parameters.events.cookie = (void *) & mr_exchange_active;
mr_exchange_parameters.events.local_fn = decrement;
for (i = 0; i < num_tasks; i++)
{
if (i == my_task_id) continue;
PAMI_Endpoint_create (client, i, 0, & mr_exchange_parameters.send.dest);
mr_exchange_active++;
PAMI_Send (context[0], & mr_exchange_parameters);
}
/*
* Advance until local memory regions have been sent and
* all memory regions have been received.
*/
unsigned num_memregions_active;
do
{
num_memregions_active = 0;
for (i = 0; i < num_tasks; i++)
num_memregions_active += mr_info[i].active;
PAMI_Context_advance (context[0], 1);
}
while (num_memregions_active < num_tasks);
while (mr_exchange_active > 0)
PAMI_Context_advance (context[0], 1);
#ifdef ASYNC_PROGRESS
async_progress_t async_progress;
async_progress_open (client, &async_progress);
async_progress_enable (&async_progress, context[1]);
#endif
if (my_task_id == target_task_id)
{
/*
* This is the "passive target" task.
*/
#ifdef ASYNC_PROGRESS
/*
* Do "something" besides communication for a little bit.
*/
sleep(1);
#else
/*
* Advance the second context for a little bit.
*/
fprintf (stdout, "(%03d) spoofing async progress\n", __LINE__);
for (i=0; i<10; i++)
{
fprintf (stdout, "(%03d) 'async progress context' advancing\n", __LINE__);
PAMI_Context_advance (context[1], 100000);
fprintf (stdout, "(%03d) 'async progress context' sleeping\n", __LINE__);
sleep(1);
}
#endif
}
else if (my_task_id == origin_task_id)
{
/*
* This is the "active origin" task.
*/
{
/*
* Use rdma put to initialize the remote buffer with the local data.
*/
volatile unsigned rput_active = 1;
pami_rput_simple_t rput_parameters = {0};
PAMI_Endpoint_create (client, target_task_id, 1, & rput_parameters.rma.dest);
rput_parameters.rma.bytes = num_doubles * sizeof(double);
rput_parameters.rdma.local.mr = mr_info[origin_task_id].memregion;
rput_parameters.rdma.local.offset = 0;
rput_parameters.rdma.remote.mr = mr_info[target_task_id].memregion;
rput_parameters.rdma.remote.offset = 0;
rput_parameters.put.rdone_fn = decrement;
rput_parameters.rma.cookie = (void *) & rput_active;
PAMI_Rput (context[0], & rput_parameters);
while (rput_active > 0)
PAMI_Context_advance (context[0], 1);
}
{
volatile unsigned send_active = 0;
accumulate_test_t test_id = ACCUMULATE_TEST_SCALAR_SUM;
pami_send_t send_parameters = {0};
PAMI_Endpoint_create (client, target_task_id, 1, & send_parameters.send.dest);
send_parameters.send.dispatch = ACCUMULATE_TEST_DISPATCH_ID;
send_parameters.send.header.iov_len = sizeof (accumulate_test_t);
send_parameters.send.header.iov_base = (void *) & test_id;
send_parameters.send.data.iov_base = test_info.data_buffer.iov_base;
send_parameters.send.data.iov_len = test_info.data_buffer.iov_len;
send_parameters.events.remote_fn = decrement;
send_parameters.events.cookie = (void *) & send_active;
for (test_id = ACCUMULATE_TEST_SCALAR_SUM; test_id < ACCUMULATE_TEST_COUNT; test_id++)
{
send_active = 1;
fprintf (stdout, "(%03d) sending data buffer for accumulate test \"%s\"\n", __LINE__, accumulate_test_name[test_id]);
PAMI_Send (context[0], & send_parameters);
fprintf (stdout, "(%03d) waiting for remote completion of data buffer sent for accumulate test \"%s\"\n", __LINE__, accumulate_test_name[test_id]);
while (send_active > 0)
PAMI_Context_advance (context[0], 1);
fprintf (stdout, "(%03d) data buffer received on remote for accumulate test \"%s\"\n", __LINE__, accumulate_test_name[test_id]);
}
}
{
/*
* Use rdma get to retrieve the remote buffer and compare results.
*/
volatile unsigned rget_active = 1;
pami_rget_simple_t rget_parameters = {0};
PAMI_Endpoint_create (client, target_task_id, 1, & rget_parameters.rma.dest);
rget_parameters.rma.done_fn = decrement;
rget_parameters.rma.cookie = (void *) & rget_active;
rget_parameters.rma.bytes = sizeof(double) * num_doubles;
rget_parameters.rdma.local.mr = mr_info[origin_task_id].memregion;
rget_parameters.rdma.local.offset = 0;
rget_parameters.rdma.remote.mr = mr_info[target_task_id].memregion;
rget_parameters.rdma.remote.offset = 0;
PAMI_Rget (context[0], & rget_parameters);
while (rget_active > 0)
PAMI_Context_advance (context[0], 1);
}
}
else
{
/*
* All other tasks, if any, do nothing and simply enter the barrier.
*/
}
simple_barrier (client, context[0]);
#ifdef ASYNC_PROGRESS
async_progress_disable (&async_progress, context[1]);
async_progress_close (&async_progress);
#endif
/*
* Do cleanup ?
*/
return;
}
int main (int argc, char ** argv)
{
pami_client_t client;
pami_context_t context;
pami_configuration_t * configuration = NULL;
char cl_string[] = "TEST";
pami_result_t result = PAMI_ERROR;
result = PAMI_Client_create (cl_string, &client, NULL, 0);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to initialize pami client. result = %d\n", result);
return 1;
}
pami_task_t task = PAMIX_Client_task (client);
size_t size = PAMIX_Client_size (client);
result = PAMI_Context_createv (client, configuration, 0, &context, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami context. result = %d\n", result);
return 1;
}
/* Attempt to send using a dispatch id that has not been registered. */
char metadata[1024];
char buffer[1024];
volatile unsigned recv_active = 1;
volatile unsigned send_active = 1;
pami_send_t parameters;
parameters.send.dispatch = 10;
parameters.send.header.iov_base = metadata;
parameters.send.header.iov_len = 8;
parameters.send.data.iov_base = buffer;
parameters.send.data.iov_len = 8;
parameters.events.cookie = (void *) & send_active;
parameters.events.local_fn = decrement;
parameters.events.remote_fn = NULL;
PAMI_Endpoint_create (client, size-1, 0, ¶meters.send.dest);
if (task == 0)
{
result = PAMI_Send (context, ¶meters);
if (result == PAMI_SUCCESS)
{
fprintf (stderr, "Test failure. Expected error when using an unregistered dispatch id.\n");
return 1;
}
}
size_t dispatch = 10;
pami_dispatch_callback_function fn;
fn.p2p = test_dispatch;
pami_dispatch_hint_t options = {};
result = PAMI_Dispatch_set (context, dispatch, fn, (void *) & recv_active, options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami dispatch. result = %d\n", result);
return 1;
}
if (task == 0)
{
PAMI_Endpoint_create (client, size-1, 0, ¶meters.send.dest);
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to send using a registered dispatch id.\n");
return 1;
}
while (send_active)
{
result = PAMI_Context_advance (context, 1000);
if ( (result != PAMI_SUCCESS) && (result != PAMI_EAGAIN) )
{
fprintf (stderr, "Error. Unable to advance pami context. result = %d\n", result);
return 1;
}
}
while (recv_active)
{
result = PAMI_Context_advance (context, 1000);
if ( (result != PAMI_SUCCESS) && (result != PAMI_EAGAIN) )
{
fprintf (stderr, "Error. Unable to advance pami context. result = %d\n", result);
return 1;
}
}
}
else if (task == (size-1))
{
PAMI_Endpoint_create (client, 0, 0, ¶meters.send.dest);
while (recv_active)
{
result = PAMI_Context_advance (context, 1000);
if ( (result != PAMI_SUCCESS) && (result != PAMI_EAGAIN) )
{
fprintf (stderr, "Error. Unable to advance pami context. result = %d\n", result);
return 1;
}
}
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to send using a registered dispatch id.\n");
return 1;
}
while (send_active)
{
result = PAMI_Context_advance (context, 1000);
if ( (result != PAMI_SUCCESS) && (result != PAMI_EAGAIN) )
{
fprintf (stderr, "Error. Unable to advance pami context. result = %d\n", result);
return 1;
}
}
}
result = PAMI_Context_destroyv(&context, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami context. result = %d\n", result);
return 1;
}
result = PAMI_Client_destroy(&client);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to finalize pami client. result = %d\n", result);
return 1;
}
return 0;
};
void optiq_pami_transport_init(struct optiq_transport *self)
{
#ifdef __bgq__
const char client_name[] = "OPTIQ";
struct optiq_pami_transport *pami_transport;
pami_result_t result;
pami_configuration_t query_configurations[3];
size_t contexts;
int configuration_count = 0;
pami_configuration_t *configurations = NULL;
pami_transport = (struct optiq_pami_transport *) optiq_transport_get_concrete_transport(self);
optiq_pami_data_init(pami_transport);
pami_transport->num_contexts = 1;
/*
* Create client
*/
result = PAMI_Client_create(client_name, &pami_transport->client, configurations, configuration_count);
assert(result == PAMI_SUCCESS);
if (result != PAMI_SUCCESS) {
return;
}
/*
* Create context
*/
result = PAMI_Context_createv(pami_transport->client, configurations, configuration_count, &pami_transport->context, pami_transport->num_contexts);
assert(result == PAMI_SUCCESS);
if (result != PAMI_SUCCESS) {
return;
}
query_configurations[0].name = PAMI_CLIENT_NUM_TASKS;
query_configurations[1].name = PAMI_CLIENT_TASK_ID;
query_configurations[2].name = PAMI_CLIENT_NUM_CONTEXTS;
result = PAMI_Client_query(pami_transport->client, query_configurations, 3);
self->size = query_configurations[0].value.intval;
self->rank = query_configurations[1].value.intval;
contexts = query_configurations[2].value.intval;
pami_transport->rank = self->rank;
pami_transport->size = self->size;
pami_transport->node_id = self->rank;
assert(contexts >= 1);
/*Create endpoint for communication*/
pami_transport->endpoints = (pami_endpoint_t *)core_memory_alloc(sizeof(pami_endpoint_t) * self->size, "endpoints", "pami_init");
for (int i = 0; i < self->size; i++) {
PAMI_Endpoint_create(pami_transport->client, i, 0, &pami_transport->endpoints[i]);
}
/*
* Register dispatch IDs
*/
pami_dispatch_callback_function fn;
pami_dispatch_hint_t options = {};
/*Message has come notification*/
fn.p2p = optiq_recv_message_fn;
result = PAMI_Dispatch_set (pami_transport->context,
RECV_MESSAGE_DISPATCH_ID,
fn,
(void *) pami_transport,
options);
assert(result == PAMI_SUCCESS);
if (result != PAMI_SUCCESS) {
return;
}
/*Job done notification*/
fn.p2p = optiq_recv_job_done_notification_fn;
result = PAMI_Dispatch_set (pami_transport->context,
JOB_DONE_NOTIFICATION_DISPATCH_ID,
fn,
(void *) pami_transport,
options);
assert(result == PAMI_SUCCESS);
if (result != PAMI_SUCCESS) {
return;
}
/*Other initialization*/
pami_transport->avail_recv_messages = &self->avail_recv_messages;
pami_transport->in_use_recv_messages = &self->in_use_recv_messages;
pami_transport->avail_send_messages = &self->avail_send_messages;
#endif
}
int main (int argc, char ** argv)
{
pami_client_t client;
pami_context_t context;
pami_task_t task;
size_t size;
pami_dispatch_callback_function fn;
pami_dispatch_hint_t options;
pami_result_t result = PAMI_ERROR;
/* ====== INITIALIZE ====== */
result = PAMI_Client_create ("TEST", &client, NULL, 0);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to initialize pami client. result = %d\n", result);
return 1;
}
task = client_task (client);
size = client_size (client);
result = PAMI_Context_createv (client, NULL, 0, &context, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami context. result = %d\n", result);
return 1;
}
fn.p2p = dispatch_fn;
options.recv_immediate = PAMI_HINT_DEFAULT;
result = PAMI_Dispatch_set (context,
DISPATCH_ID_DEFAULT_EXPECT_IMMEDIATE,
fn,
(void *) EXPECT_IMMEDIATE,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to register DISPATCH_ID_DEFAULT_EXPECT_IMMEDIATE. result = %d\n", result);
return 1;
}
options.recv_immediate = PAMI_HINT_DEFAULT;
result = PAMI_Dispatch_set (context,
DISPATCH_ID_DEFAULT_EXPECT_ASYNC,
fn,
(void *) EXPECT_ASYNC,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to register DISPATCH_ID_DEFAULT_EXPECT_ASYNC. result = %d\n", result);
return 1;
}
options.recv_immediate = PAMI_HINT_ENABLE;
result = PAMI_Dispatch_set (context,
DISPATCH_ID_ENABLE,
fn,
(void *) EXPECT_IMMEDIATE,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to register DISPATCH_ID_ENABLE. result = %d\n", result);
return 1;
}
options.recv_immediate = PAMI_HINT_DISABLE;
result = PAMI_Dispatch_set (context,
DISPATCH_ID_DISABLE,
fn,
(void *) EXPECT_ASYNC,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to register DISPATCH_ID_DISABLE. result = %d\n", result);
return 1;
}
/* ====== START TEST ====== */
__test_errors = 0;
__test_recvs = 0;
size_t test_count = 0;
volatile size_t send_active = 0;
pami_send_t parameters;
parameters.send.header.iov_base = __junk;
parameters.send.header.iov_len = 0;
parameters.send.data.iov_base = __junk;
parameters.send.data.iov_len = 0;
parameters.send.dispatch = 0;
parameters.events.cookie = (void *) & send_active;
parameters.events.local_fn = decrement;
parameters.events.remote_fn = NULL;
result = PAMI_Endpoint_create (client, 0, 0, ¶meters.send.dest);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error: PAMI_Endpoint_create() failed for task 0, context 0 with %d.\n", result);
return 1;
}
/* ===================================================================
* 'recv_immediate' default
*
* (header+data) > recv_immediate_max MUST be an asynchronous receive
*
* A zero-byte send will \b always result in an immediate receive.
* \see pami_dispatch_p2p_function
*
* Data sizes to test:
* - recv_immediate_max + 1
* - 0
*/
parameters.send.data.iov_len = recv_immediate_max (context, DISPATCH_ID_DEFAULT_EXPECT_ASYNC) + 1;
parameters.send.dispatch = DISPATCH_ID_DEFAULT_EXPECT_ASYNC;
test_count++;
if (task == 1)
{
send_active++;
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error: Unable to send to 0x%08x using dispatch %zu with %d.\n", parameters.send.dest, parameters.send.dispatch, result);
return 1;
}
}
parameters.send.data.iov_len = 0;
parameters.send.dispatch = DISPATCH_ID_DEFAULT_EXPECT_IMMEDIATE;
test_count++;
if (task == 1)
{
send_active++;
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error: Unable to send to 0x%08x using dispatch %zu with %d.\n", parameters.send.dest, parameters.send.dispatch, result);
return 1;
}
}
/* ===================================================================
* 'recv_immediate' enabled
*
* All receives are 'immediate'. (header+data) > recv_immediate_max is
* invalid, but may not neccesarily return an error.
*
* Data sizes to test:
* - 0
* - recv_immediate_max
* - recv_immediate_max + 1 ...... ?
*/
parameters.send.data.iov_len = 0;
parameters.send.dispatch = DISPATCH_ID_ENABLE;
test_count++;
if (task == 1)
{
send_active++;
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error: Unable to send to 0x%08x using dispatch %zu with %d.\n", parameters.send.dest, parameters.send.dispatch, result);
return 1;
}
}
parameters.send.data.iov_len = recv_immediate_max (context, DISPATCH_ID_DEFAULT_EXPECT_ASYNC);
parameters.send.dispatch = DISPATCH_ID_ENABLE;
test_count++;
if (task == 1)
{
send_active++;
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error: Unable to send to 0x%08x using dispatch %zu with %d.\n", parameters.send.dest, parameters.send.dispatch, result);
return 1;
}
}
#if 0
parameters.send.data.iov_len = recv_immediate_max (context, DISPATCH_ID_DEFAULT_EXPECT_ASYNC) + 1;
parameters.send.dispatch = DISPATCH_ID_ENABLE;
test_count++;
if (task == 1)
{
send_active++;
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error: Unable to send to 0x%08x using dispatch %d with %d.\n", parameters.send.dest, parameters.send.dispatch, result);
return 1;
}
}
#endif
/* ===================================================================
* 'recv_immediate' disabled
*
* All receives are NOT 'immediate' - even "zero byte data"
*
* Data sizes to test:
* - 0
* - recv_immediate_max
* - recv_immediate_max + 1
*/
parameters.send.data.iov_len = 0;
parameters.send.dispatch = DISPATCH_ID_DISABLE;
test_count++;
if (task == 1)
{
send_active++;
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error: Unable to send to 0x%08x using dispatch %zu with %d.\n", parameters.send.dest, parameters.send.dispatch, result);
return 1;
}
}
parameters.send.data.iov_len = recv_immediate_max (context, DISPATCH_ID_DEFAULT_EXPECT_ASYNC);
parameters.send.dispatch = DISPATCH_ID_DISABLE;
test_count++;
if (task == 1)
{
send_active++;
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error: Unable to send to 0x%08x using dispatch %zu with %d.\n", parameters.send.dest, parameters.send.dispatch, result);
return 1;
}
}
parameters.send.data.iov_len = recv_immediate_max (context, DISPATCH_ID_DEFAULT_EXPECT_ASYNC) + 1;
parameters.send.dispatch = DISPATCH_ID_DISABLE;
test_count++;
if (task == 1)
{
send_active++;
result = PAMI_Send (context, ¶meters);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error: Unable to send to 0x%08x using dispatch %zu with %d.\n", parameters.send.dest, parameters.send.dispatch, result);
return 1;
}
}
/* ====== WAIT FOR COMMUNICATION COMPLETION ====== */
if (task == 0)
{
while (__test_recvs < test_count)
PAMI_Context_advance (context, 1000);
}
else if (task == 1)
{
while (send_active)
PAMI_Context_advance (context, 1000);
}
/* ====== CLEANUP ====== */
result = PAMI_Context_destroyv (&context, 1);
if (result != PAMI_SUCCESS) {
fprintf (stderr, "Error. Unable to destroy context, result = %d\n", result);
return 1;
}
result = PAMI_Client_destroy (&client);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami client. result = %d\n", result);
return 1;
}
/* ====== REPORT ERRORS ====== */
if (__test_errors > 0)
{
fprintf (stderr, "Error. Non-compliant PAMI receive immediate implementation! error count = %zu\n", __test_errors);
return 1;
}
return 0;
}
int main(int argc, char* argv[])
{
pami_result_t result = PAMI_ERROR;
/* initialize the second client */
char * clientname = "";
pami_client_t client;
result = PAMI_Client_create(clientname, &client, NULL, 0);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_create");
/* query properties of the client */
pami_configuration_t config[3];
size_t num_contexts;
config[0].name = PAMI_CLIENT_NUM_TASKS;
config[1].name = PAMI_CLIENT_TASK_ID;
config[2].name = PAMI_CLIENT_NUM_CONTEXTS;
result = PAMI_Client_query(client, config, 3);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_query");
world_size = config[0].value.intval;
world_rank = config[1].value.intval;
num_contexts = config[2].value.intval;
TEST_ASSERT(num_contexts>1,"num_contexts>1");
if (world_rank==0)
{
printf("hello world from rank %ld of %ld \n", world_rank, world_size );
fflush(stdout);
}
/* initialize the contexts */
contexts = (pami_context_t *) safemalloc( num_contexts * sizeof(pami_context_t) );
result = PAMI_Context_createv( client, NULL, 0, contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_createv");
/* setup the world geometry */
pami_geometry_t world_geometry;
result = PAMI_Geometry_world(client, &world_geometry );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Geometry_world");
int status = pthread_create(&Progress_thread, NULL, &Progress_function, NULL);
TEST_ASSERT(status==0, "pthread_create");
/************************************************************************/
int n = (argc>1 ? atoi(argv[1]) : 1000);
size_t bytes = n * sizeof(int);
int * local = (int *) safemalloc(bytes);
for (int i=0; i<n; i++)
local[i] = world_rank;
/* create the endpoint */
int target = (world_rank>0 ? world_rank-1 : world_size-1);
pami_endpoint_t target_ep;
result = PAMI_Endpoint_create(client, (pami_task_t) target, 1, &target_ep);
//result = PAMI_Endpoint_create(client, (pami_task_t) target, 0, &target_ep);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Endpoint_create");
/* register the dispatch function */
pami_dispatch_callback_function dispatch_cb;
size_t dispatch_id = 37;
dispatch_cb.p2p = dispatch_recv_cb;
pami_dispatch_hint_t dispatch_hint = {0};
int dispatch_cookie = 1000000+world_rank;
//dispatch_hint.recv_immediate = PAMI_HINT_DISABLE;
result = PAMI_Dispatch_set(contexts[0], dispatch_id, dispatch_cb, &dispatch_cookie, dispatch_hint);
result = PAMI_Dispatch_set(contexts[1], dispatch_id, dispatch_cb, &dispatch_cookie, dispatch_hint);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Dispatch_set");
/* make sure everything is ready */
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
// The iovec structure describes a buffer. It contains two fields:
// void *iov_base - Contains the address of a buffer.
// size_t iov_len - Contains the length of the buffer.
int header = 37373;
int active = 1;
pami_send_t parameters;
parameters.send.header.iov_base = &header;
parameters.send.header.iov_len = sizeof(int);
parameters.send.data.iov_base = local;
parameters.send.data.iov_len = bytes;
parameters.send.dispatch = dispatch_id;
//parameters.send.hints = ;
parameters.send.dest = target_ep;
parameters.events.cookie = &active;
parameters.events.local_fn = cb_done;
parameters.events.remote_fn = NULL;//cb_done;
uint64_t t0 = GetTimeBase();
result = PAMI_Send(contexts[0], ¶meters);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Send");
while (active)
{
//result = PAMI_Context_advance( contexts[0], 100);
//TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_advance");
result = PAMI_Context_trylock_advancev(&(contexts[0]), 1, 1000);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_trylock_advancev");
}
uint64_t t1 = GetTimeBase();
uint64_t dt = t1-t0;
/* barrier on non-progressing context to make sure CHT does its job */
barrier(world_geometry, contexts[0]);
printf("%ld: PAMI_Send of %ld bytes achieves %lf MB/s \n", (long)world_rank, bytes, 1.6e9*1e-6*(double)bytes/(double)dt );
fflush(stdout);
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
free(local);
/************************************************************************/
void * rv;
status = pthread_cancel(Progress_thread);
TEST_ASSERT(status==0, "pthread_cancel");
status = pthread_join(Progress_thread, &rv);
TEST_ASSERT(status==0, "pthread_join");
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
/* finalize the contexts */
result = PAMI_Context_destroyv( contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_destroyv");
free(contexts);
/* finalize the client */
result = PAMI_Client_destroy( &client );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_destroy");
if (world_rank==0)
printf("%ld: end of test \n", world_rank );
fflush(stdout);
return 0;
}
int main (int argc, char ** argv)
{
pami_client_t client;
pami_context_t context[2];
pami_task_t task_id;
size_t num_tasks = 0;
size_t ncontexts = 0;
size_t errors = 0;
pami_result_t result = PAMI_ERROR;
pami_type_t subtype;
pami_type_t compound_type;
pami_type_t simple_type;
info_t exchange[MAX_TASKS];
double data[BUFFERSIZE];
volatile unsigned ready;
{ /* init */
ready = 0;
unsigned i;
for (i = 0; i < MAX_TASKS; i++)
exchange[i].active = 0;
for (i = 0; i < BUFFERSIZE; i++)
data[i] = E;
result = PAMI_Client_create ("TEST", &client, NULL, 0);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami client. result = %d\n", result);
return 1;
}
pami_configuration_t configuration;
configuration.name = PAMI_CLIENT_TASK_ID;
result = PAMI_Client_query(client, &configuration, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
task_id = configuration.value.intval;
/*fprintf (stderr, "My task id = %d\n", task_id);*/
configuration.name = PAMI_CLIENT_NUM_TASKS;
result = PAMI_Client_query(client, &configuration, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
num_tasks = configuration.value.intval;
/*if (task_id == 0)
fprintf (stderr, "Number of tasks = %zu\n", num_tasks);*/
if ((num_tasks < 2) || (num_tasks > MAX_TASKS))
{
fprintf (stderr, "Error. This test requires 2-%d tasks. Number of tasks in this job: %zu\n", MAX_TASKS, num_tasks);
return 1;
}
if (task_id == num_tasks - 1)
{
for (i = 0; i < 320; i++)
data[i] = PI;
}
configuration.name = PAMI_CLIENT_NUM_CONTEXTS;
result = PAMI_Client_query(client, &configuration, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
ncontexts = (configuration.value.intval < 2) ? 1 : 2;
/*if (task_id == 0)
fprintf (stderr, "maximum contexts = %zu, number of contexts used in this test = %zu\n", configuration.value.intval, ncontexts);*/
result = PAMI_Context_createv(client, NULL, 0, context, ncontexts);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami context(s). result = %d\n", result);
return 1;
}
pami_dispatch_hint_t options = {};
for (i = 0; i < ncontexts; i++)
{
pami_dispatch_callback_function fn;
fn.p2p = dispatch_exchange;
result = PAMI_Dispatch_set (context[i],
EXCHANGE_DISPATCH_ID,
fn,
(void *) exchange,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami 'exchange' dispatch. result = %d\n", result);
return 1;
}
fn.p2p = dispatch_notify;
result = PAMI_Dispatch_set (context[i],
NOTIFY_DISPATCH_ID,
fn,
(void *) & ready,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami 'notify' dispatch. result = %d\n", result);
return 1;
}
}
/* ***********************************
* Create the pami types
************************************/
/* This compound noncontiguous type is composed of one double, skips a double,
* two doubles, skips a double, three doubles, skips a double, five doubles,
* skips a double, six doubles, skips a double, seven doubles, skips a double,
* eight doubles, then skips two doubles.
*
* This results in a type with 32 doubles that is 40 doubles
* 'wide'.
*/
PAMI_Type_create (&subtype);
PAMI_Type_add_simple (subtype,
sizeof(double), /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 2); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 2, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 3); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 3, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 4); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 5, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 6); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 6, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 7); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 7, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 8);/* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 8, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 10);/* stride */
PAMI_Type_complete (subtype, sizeof(double));
/* This noncontiguous type is composed of the above compound type, repeated
* ten times with no stride.
*
* This results in a type with 320 doubles that is 400 doubles
* 'wide'.
*/
PAMI_Type_create (&compound_type);
PAMI_Type_add_typed (compound_type,
subtype, /* subtype */
0, /* offset */
10, /* count */
sizeof(double) * 32); /* stride */
PAMI_Type_complete (compound_type, sizeof(double));
/* This simple noncontiguous type is composed of eight contiguous doubles,
* then skips a _single_ double, repeated 40 times.
*
* This results in a type with 320 doubles that is 360 doubles 'wide'.
*/
PAMI_Type_create (&simple_type);
PAMI_Type_add_simple (simple_type,
sizeof(double) * 8, /* bytes */
0, /* offset */
40, /* count */
sizeof(double) * 9); /* stride */
PAMI_Type_complete (simple_type, sizeof(double));
/* Create a memory region for the local data buffer. */
size_t bytes;
result = PAMI_Memregion_create (context[0], (void *) data, BUFFERSIZE,
&bytes, &exchange[task_id].mr);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create memory region. result = %d\n", result);
return 1;
}
else if (bytes < BUFFERSIZE)
{
fprintf (stderr, "Error. Unable to create memory region of a large enough size. result = %d\n", result);
return 1;
}
/* Broadcast the memory region to all tasks - including self. */
for (i = 0; i < num_tasks; i++)
{
pami_send_immediate_t parameters;
parameters.dispatch = EXCHANGE_DISPATCH_ID;
parameters.header.iov_base = (void *) & bytes;
parameters.header.iov_len = sizeof(size_t);
parameters.data.iov_base = (void *) & exchange[task_id].mr;
parameters.data.iov_len = sizeof(pami_memregion_t);
PAMI_Endpoint_create (client, i, 0, ¶meters.dest);
result = PAMI_Send_immediate (context[0], ¶meters);
}
/* Advance until all memory regions have been received. */
for (i = 0; i < num_tasks; i++)
{
while (exchange[i].active == 0)
PAMI_Context_advance (context[0], 100);
}
} /* init done */
pami_send_immediate_t notify;
notify.dispatch = NOTIFY_DISPATCH_ID;
notify.header.iov_base = NULL;
notify.header.iov_len = 0;
notify.data.iov_base = NULL;
notify.data.iov_len = 0;
volatile size_t active = 1;
pami_rget_typed_t parameters;
parameters.rma.hints = (pami_send_hint_t) {0};
parameters.rma.cookie = (void *) & active;
parameters.rma.done_fn = decrement;
parameters.rma.bytes = 320 * sizeof(double);
if (task_id == 0)
{
fprintf (stdout, "PAMI_Rget('typed') functional test %s\n", (ncontexts < 2) ? "" : "[crosstalk]");
fprintf (stdout, "\n");
parameters.rdma.local.mr = &exchange[0].mr;
parameters.rdma.remote.mr = &exchange[num_tasks - 1].mr;
PAMI_Endpoint_create (client, num_tasks - 1, ncontexts - 1, ¶meters.rma.dest);
PAMI_Endpoint_create (client, num_tasks - 1, ncontexts - 1, ¬ify.dest);
}
else
{
parameters.rdma.local.mr = &exchange[num_tasks - 1].mr;
parameters.rdma.remote.mr = &exchange[0].mr;
PAMI_Endpoint_create (client, 0, 0, ¶meters.rma.dest);
PAMI_Endpoint_create (client, 0, 0, ¬ify.dest);
}
/* ******************************************************************** */
/* contiguous -> contiguous transfer test */
/* ******************************************************************** */
if (task_id == 0)
{
parameters.rdma.local.offset = 0;
parameters.rdma.remote.offset = 0;
parameters.type.local = PAMI_TYPE_DOUBLE;
parameters.type.remote = PAMI_TYPE_DOUBLE;
active = 1;
PAMI_Rget_typed (context[0], ¶meters);
while (active > 0)
PAMI_Context_advance (context[0], 100);
/* Notify the remote task that the data has been transfered. */
PAMI_Send_immediate (context[0], ¬ify);
}
else if (task_id == num_tasks - 1)
{
/* Wait for notification that the data has been transfered. */
while (ready == 0)
PAMI_Context_advance (context[ncontexts - 1], 100);
ready = 0;
}
/* ******************************************************************** */
/* contiguous -> non-contiguous transfer test */
/* ******************************************************************** */
if (task_id == num_tasks - 1)
{
parameters.rdma.local.offset = 4 * 1024;
parameters.rdma.remote.offset = 0;
parameters.type.local = simple_type;
parameters.type.remote = PAMI_TYPE_DOUBLE;
active = 1;
PAMI_Rget_typed (context[ncontexts - 1], ¶meters);
while (active > 0)
PAMI_Context_advance (context[ncontexts - 1], 100);
/* Notify the remote task that the data has been transfered. */
PAMI_Send_immediate (context[ncontexts - 1], ¬ify);
}
else if (task_id == 0)
{
/* Wait for notification that the data has been transfered. */
while (ready == 0)
PAMI_Context_advance (context[0], 100);
ready = 0;
}
/* ******************************************************************** */
/* non-contiguous -> non-contiguous transfer test */
/* ******************************************************************** */
if (task_id == 0)
{
parameters.rdma.local.offset = 4 * 1024;
parameters.rdma.remote.offset = 4 * 1024;
parameters.type.local = compound_type;
parameters.type.remote = simple_type;
active = 1;
PAMI_Rget_typed (context[0], ¶meters);
while (active > 0)
PAMI_Context_advance (context[0], 100);
/* Notify the remote task that the data has been transfered. */
PAMI_Send_immediate (context[0], ¬ify);
}
else if (task_id == num_tasks - 1)
{
/* Wait for notification that the data has been transfered. */
while (ready == 0)
PAMI_Context_advance (context[ncontexts - 1], 100);
ready = 0;
}
/* ******************************************************************** */
/* non-contiguous -> contiguous transfer test */
/* ******************************************************************** */
if (task_id == num_tasks - 1)
{
parameters.rdma.local.offset = 8 * 1024;
parameters.rdma.remote.offset = 4 * 1024;
parameters.type.local = PAMI_TYPE_DOUBLE;
parameters.type.remote = compound_type;
active = 1;
PAMI_Rget_typed (context[ncontexts - 1], ¶meters);
while (active > 0)
PAMI_Context_advance (context[ncontexts - 1], 100);
/* Notify the remote task that the data has been transfered. */
PAMI_Send_immediate (context[ncontexts - 1], ¬ify);
}
else if (task_id == 0)
{
/* Wait for notification that the data has been transfered. */
while (ready == 0)
PAMI_Context_advance (context[0], 100);
ready = 0;
}
/* ******************************************************************** */
/* VERIFY data buffers */
/* ******************************************************************** */
if (task_id == num_tasks - 1)
{
if (task_id == 0)
{
unsigned i = 0;
for (; i < 320; i++)
{
if (data[i] != PI)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, PI, data[i]);
}
}
for (; i < 512; i++)
{
if (data[i] != E)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, E, data[i]);
}
}
unsigned j = 0;
for (; j < 40; j++)
{
unsigned n = 0;
for (; n < 8; n++)
{
if (data[i] != PI)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, PI, data[i]);
}
i++;
}
if (data[i] != E)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, E, data[i]);
}
i++;
}
for (; i < 1024; i++)
{
if (data[i] != E)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, E, data[i]);
}
}
for (; i < 1024 + 320; i++)
{
if (data[i] != PI)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, PI, data[i]);
}
}
for (; i < BUFFERSIZE; i++)
{
if (data[i] != E)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, E, data[i]);
}
}
}
}
{ /* cleanup */
result = PAMI_Context_destroyv(context, ncontexts);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami context. result = %d\n", result);
return 1;
}
result = PAMI_Client_destroy(&client);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami client. result = %d\n", result);
return 1;
}
if (task_id == num_tasks - 1)
{
if (errors)
fprintf (stdout, "Test completed with errors (%zu)\n", errors);
else
fprintf (stdout, "Test completed with success\n");
}
} /* cleanup done */
return (errors != 0);
};
