int main()
{
int matrix_width = WIDTH,
matrix_size = WIDTH * WIDTH;
void *params = malloc(sizeof(int) +
3 * matrix_size * sizeof(float));
*((int *) params) = matrix_width;
float *A = (float *) (((char *) params) + sizeof(int));
float *B = A + matrix_size;
float *C = B + matrix_size;
init_mtx(A, matrix_size);
init_mtx(B, matrix_size);
printf("Both Matrices A and B look like this:\n");
print_mtx(A, matrix_width);
printf("\n");
optimized_matrix_multiplication(params);
printf("Result looks like this:\n");
print_mtx(C, matrix_width);
free(params);
return 0;
}
main(int ac, char **av)
{
FILE *fp;
matrix prox=NULL,prepare_prox_mtx(),first_guess=NULL;
matrix TMP=NULL,RES=NULL,mds2res();
int dim ;
if (ac != 3) {
printf("USAGE: %s <data_file> <dimenssion> \n",av[0]);
exit(1);
}
fp = fopen(av[1],"r");
dim = atoi(av[2]);
prox = prepare_prox_mtx(fp,dim);
first_guess = generate_first_guess(dim,pnts,50);
if (dim >1 ) TMP = metric_mds(dim,pnts,first_guess,prox,J_ee,DJ_ee);
else TMP = metric_mds(dim,pnts,first_guess,prox,OneJ_ee,DOneJ_ee);
RES = mds2res(TMP,dim);
compare_results(RES,prox,dim);
free_all_mtx(dim,&TMP,&prox,&first_guess);
printf("Points coordinates are: \n \n");
print_mtx(RES);
}
int main(int argc, char *argv[])
{
scif_epd_t epd;
int bytes_sent, bytes_received, count;
/* message state related variables */
int *id, *type, *threads;
size_t message_size, request, request_size;
void *message, *params;
/* do the standard open, bind, connect in SCIF */
epd = scif_obc();
/* Create request */
request_size = sizeof(size_t);
request = 1;
printf("= About to send %zd bytes\n", request_size);
/* Send message */
bytes_sent = scif_send(epd, &request, request_size, 1);
printf("= Sent %d bytes\n= Waiting for reply ...\n", bytes_sent);
/* Receive size of the reply */
bytes_received = scif_recv(epd, &message_size, sizeof(size_t), 1);
printf("= Received %d bytes. Expecting a message of size %zu bytes\n",
bytes_received, message_size);
/* Receive the actual reply */
message = malloc(message_size);
count = scif_recv(epd, message, message_size, 1);
bytes_received += count;
printf("= Received %d bytes. Total bytes received: %d bytes\n",
count, bytes_received);
/*
* Extract the message received:
* --------------------------------------------------
* | type (int) | threads (int) | id (int) | params |
* --------------------------------------------------
*/
type = (int *) message;
threads = type + 1;
id = threads + 1;
params = (void *) ((char *) id + sizeof(int));
printf("= Content size: %zu bytes - Type: %d - Threads: %d - ID: %d\n",
message_size, *type, *threads, *id);
/* output results */
int matrix_width, matrix_size;
float *result;
switch(*type) {
case 1:
printf("= Sleep duration left: %u\n", *((unsigned int *) params));
break;
case 2:
matrix_width = *((int *) params);
result = (float *) (((char *) params) + sizeof(int));
print_mtx(result, matrix_width);
break;
case 3: case 4: case 6:
matrix_width = *((int *) params);
matrix_size = matrix_width * matrix_width;
result = (float *) (((char *) params) + sizeof(int)
+ 2 * matrix_size * sizeof(float));
print_mtx(result, matrix_width);
break;
case 5:
matrix_width = *((int *) params + 1);
matrix_size = matrix_width * matrix_width;
result = (float *) (((char *) params) + 2 * sizeof(int)
+ 2 * matrix_size * sizeof(float));
print_mtx(result, matrix_width);
break;
default:
printf("= Dat shit cray!\n");
break;
}
free(message);
if (scif_close(epd) != 0) {
fprintf(stderr, "scif_close failed with error %d\n", errno);
exit(EXIT_FAILURE);
}
printf("= scif_close success\n");
return EXIT_SUCCESS;
}
int main(int argc, char *argv[])
{
int node_id = 0;
int arrival_lambda = 10;
int thread_cpu_map[N_THREADS];
int i,j,k;
int n_threads;
int n_left;
int n_right;
int next_index_left = 3;
int next_index_right = 7;
float local_square = 0.0, remote_square = 0.0;
/***************** make sure #args is correct and get the n_threads, n_left and n_right */
if(argc < 4)
{
printf("Usage: ./test_numa_comb n_of_threads n_of_threads_on_node0 n_of_threads_on_node1\n");
exit(-1);
}
n_threads = atoi(argv[1]);
n_left = atoi(argv[2]);
n_right = atoi(argv[3]);
/******************* Set the thread_cpu_map according to the n_left and n_right */
printf("n_threads: %d, n_left: %d, n_right: %d\n",n_threads,n_left,n_right);
for(i = 0; i < n_left; i++)
{
thread_cpu_map[i] = next_index_left;
next_index_left--;
}
for(i = n_left; i < n_threads; i++)
{
thread_cpu_map[i] = next_index_right;
next_index_right--;
}
for(i = 0; i < n_threads; i++)
{
printf("Thread %d is on cpu %d\n",i,thread_cpu_map[i]);
}
thread_params para[n_threads]; //The parameters to pass to the threads
//printf("The return value of numa_get_run_node_mask(void) is %d\n",numa_get_run_node_mask());
//printf("The return value of numa_max_node(void) is %d\n",numa_max_node());
//numa_tonode_memory((void *)spinlock_ptr,sizeof(pthread_spinlock_t),node_id); //This doesn't work
//initilize the spinlock pointer and put it on a specific node
pthread_spinlock_t *spinlock_ptr = numa_alloc_onnode(sizeof(pthread_spinlock_t),node_id);
if(spinlock_ptr == NULL) //error handling of the allocating of a spinlock pointer on a specific node
{
printf("alloc of spinlock on a node failed.\n");
exit(-1);
}
/* initialise syncs */
pthread_barrier_init(&fin_barrier, NULL, n_threads);
pthread_spin_init(spinlock_ptr,0);
int rc;
//create the threads
for(i = 0; i < n_threads; i++)
{
para[i].thread_id = i;
para[i].arrival_lambda = arrival_lambda;
para[i].spinlock_ptr = spinlock_ptr;
CPU_ZERO(&cpuset[i]);
CPU_SET(thread_cpu_map[i],&cpuset[i]);
rc = pthread_create(&threads[i],NULL,work,(void*)¶[i]);
E (rc);
}
start_work_flag = 1;
/* wait here */
for(i = 0; i < n_threads; i++)
pthread_join(threads[i],NULL);
pthread_barrier_destroy(&fin_barrier);
/*
for(i = 0; i < n_threads; i++)
{
printf("The time to get one lock for thread %d is : %.9f\n",i,time_in_cs[i]/num_access_each_thread[i]);
printf("The number of lock accesses for thread %d is : %d\n",i,num_access_each_thread[i]);
}
*/
qsort((void*)g_tss,(size_t)access_count,(size_t)sizeof(timestamp),cmp_timestamp);
/*
for (i = 0; i < access_count; i++)
printf("%lu with id %d\n",g_tss[i].ts,g_tss[i].id);
*/
/* for (i = 0; i < access_count; i++)
* {
* printf ("%lu %d\n", g_tss[i].ts, g_tss[i].id);
* } */
/* */
int cs_order[access_count/2];
for(i = 0; i < access_count/2; i++)
{
cs_order[i] = g_tss[i*2].id;
//printf("%d in cs\n",cs_order[i]);
}
int cs_matrix[n_threads][n_threads];
uint64_t delay_matrix[n_threads][n_threads];
float prob_matrix[n_threads][n_threads];
float rate_matrix[n_threads][n_threads];
// zero out all the matrices
memset(&cs_matrix, '\0', n_threads*n_threads*sizeof(int));
memset(&delay_matrix, '\0', n_threads*n_threads*sizeof(uint64_t));
memset(&prob_matrix, '\0', n_threads*n_threads*sizeof(float));
int local_count2 = 0, remote_count2 = 0;
uint64_t diff;
for(i = 0; i < n_threads; i++)
for(j = 0; j < n_threads; j++)
for(k = 0; k < access_count/2 -1 ; k++)
{
if(cs_order[k] == i && cs_order[k+1] == j)
{
cs_matrix[i][j]++;
diff = g_tss[2*k+2].ts - g_tss[2*k+1].ts;
delay_matrix[i][j] += diff;
if(is_on_same_node(i, j, n_threads, n_left, n_right))
{
dprintf("local_delay: %lu\n", diff);
local_square += sqr(diff);
local_count2++;
}
else
{
dprintf("remote_delay: %lu\n", diff);
remote_square += sqr(diff);
remote_count2++;
}
}
}
int num_access[n_threads];
for(i = 0; i < access_count/2 -1; i++)
for(j = 0; j < n_threads; j++)
{
if (cs_order[i] == j) num_access[j]++;
}
for(i = 0; i < n_threads; i++)
printf("num_access[%d]:%d\n",i,num_access[i]);
for(i = 0; i < n_threads; i++)
for(j = 0; j < n_threads ; j++)
{
prob_matrix[i][j] = (float)cs_matrix[i][j]/(float)num_access[i];
rate_matrix[i][j] = 1.0/((delay_matrix[i][j]/(float)cs_matrix[i][j])/CPU_FREQ);
}
printf ("\n***************** PROBS *******************\n");
printf ("Lock is on LP, [L, R] is [%d, %d]:\n", n_left - 1, n_right);
// tl
printf ("L -> L\n");
print_mtx (n_threads, n_threads, prob_matrix,
0, 0, n_left, n_left, 0);
// tr
printf ("L -> R\n");
print_mtx (n_threads, n_threads, prob_matrix,
n_left, 0, n_threads, n_left, 0);
printf ("Lock is on RP, [L, R] is [%d, %d]:\n", n_left, n_right - 1);
// br
printf ("R -> R\n");
print_mtx (n_threads, n_threads, prob_matrix,
n_left, n_left, n_threads, n_threads, 0);
// bl
printf ("R -> L\n");
print_mtx (n_threads, n_threads, prob_matrix,
0, n_left, n_left, n_threads, 0);
printf ("\n***************** RATES *******************\n");
printf ("Lock is on LP, [L, R] is [%d, %d]:\n", n_left - 1, n_right);
// tl
printf ("L -> L\n");
print_mtx (n_threads, n_threads, rate_matrix,
0, 0, n_left, n_left, 1);
// tr
printf ("L -> R\n");
print_mtx (n_threads, n_threads, rate_matrix,
n_left, 0, n_threads, n_left, 1);
printf ("Lock is on RP, [L, R] is [%d, %d]:\n", n_left, n_right - 1);
// br
printf ("R -> R\n");
print_mtx (n_threads, n_threads, rate_matrix,
n_left, n_left, n_threads, n_threads, 1);
// bl
printf ("R -> \n");
print_mtx (n_threads, n_threads, rate_matrix,
0, n_left, n_left, n_threads, 1);
//print the intra-core and inter-core delay
//thread 0 - n_left -1 are on the left core, n_left to n_threads are on the right core
uint64_t local_delay = 0, remote_delay = 0;
int local_count = 0, remote_count = 0;
float local_prob = 0.0, remote_prob = 0.0;
for(i = 0; i < n_threads; i++)
for(j = 0; j < n_threads; j++)
{
if (j == i)
continue;
if(is_on_same_node(i, j, n_threads, n_left, n_right))
{
//printf("%d and %d on the same node\n",i,j);
local_delay += delay_matrix[i][j];
local_count += cs_matrix[i][j];
local_prob += prob_matrix[j][i];
}
else
{
//printf("%d and %d not the same node\n",i,j);
remote_delay += delay_matrix[i][j];
remote_count += cs_matrix[i][j];
remote_prob += prob_matrix[j][i];
}
}
float local = (float)local_delay/(local_count);
float remote = (float)remote_delay/(remote_count);
printf("\n\n**************************** Aggregates ***************************\n");
printf("local delay: %f, remote_delay: %f, local_count: %d, remote_count: %d\n",(float)local_delay/(local_count),(float)remote_delay/(remote_count),local_count,remote_count);
printf("local prob:%f, remote prob: %f\n",local_prob/n_threads, remote_prob/n_threads);
printf("local delay variance:%f, remote delay variance: %f\n",local_square/local_count - local*local, remote_square/remote_count - remote*remote);
printf("local count2: %d, remote_count2:%d\n",local_count2, remote_count2);
pthread_spin_destroy(spinlock_ptr);
numa_free((void *)spinlock_ptr,sizeof(pthread_spinlock_t));
pthread_exit(NULL);
return 0;
}