Performance Optimization and Modeling of Fine-Grained Irregular Communication in UPC

Listing 6

A UPC microbenchmark for measuring the latency of individual remote-memory transfers.

	int nblks = n/BLOCKSIZE + (n% BLOCKSIZE) ?1 : 0;
	int mythread_nblks = nblks/THREADS + (MYTHREAD < (nblks% THREADS) ?1 : 0);
	shared [BLOCKSIZE] double^∗ = upc_all_alloc (nblks, BLOCKSIZE ∗ sizeof(double));
	double tmp;
	int^∗mythread_indices = (int^∗)malloc(mythread_nblks ∗ BLOCKSIZE ∗ sizeof(int));
	/∗ let array “mythread_indices” contain random global indices with affinity to “remote threads” ∗/
	randomize (mythread_indices, mythread_nblks ∗ BLOCKSIZE);
	/∗ start timing … ∗/
	for (int mb = 0, i = 0; mb < mythread_nblks; mb++)
	for (int k = 0; k < BLOCKSIZE; k++, i++)
	tmp = [mythread_indices[i]];
	/∗ stop timing … ∗/