diff options
Diffstat (limited to 'SCAN-RSS/baselines/cpu/app_baseline.cpp')
| -rw-r--r-- | SCAN-RSS/baselines/cpu/app_baseline.cpp | 108 |
1 files changed, 99 insertions, 9 deletions
diff --git a/SCAN-RSS/baselines/cpu/app_baseline.cpp b/SCAN-RSS/baselines/cpu/app_baseline.cpp index f54f123..b0ab600 100644 --- a/SCAN-RSS/baselines/cpu/app_baseline.cpp +++ b/SCAN-RSS/baselines/cpu/app_baseline.cpp @@ -34,6 +34,18 @@ #include "../../support/common.h" #include "../../support/timer.h" +#if NUMA +#include <numaif.h> +#include <numa.h> + +void* mp_pages[1]; +int mp_status[1]; +int mp_nodes[1]; +int numa_node_in = -1; +int numa_node_out = -1; +int numa_node_cpu = -1; +#endif + #define XSTR(x) STR(x) #define STR(x) #x @@ -44,7 +56,6 @@ // Pointer declaration static T* A; static T* C; -static T* C2; /** * @brief creates input arrays @@ -76,6 +87,11 @@ typedef struct Params { int n_reps; int n_threads; int exp; +#if NUMA + struct bitmask* bitmask_in; + struct bitmask* bitmask_out; + int numa_node_cpu; +#endif }Params; void usage() { @@ -101,9 +117,14 @@ struct Params input_params(int argc, char **argv) { p.n_reps = 3; p.exp = 0; p.n_threads = 8; +#if NUMA + p.bitmask_in = NULL; + p.bitmask_out = NULL; + p.numa_node_cpu = -1; +#endif int opt; - while((opt = getopt(argc, argv, "hi:w:e:x:t:")) >= 0) { + while((opt = getopt(argc, argv, "hi:w:e:x:t:a:b:c:")) >= 0) { switch(opt) { case 'h': usage(); @@ -114,6 +135,11 @@ struct Params input_params(int argc, char **argv) { case 'e': p.n_reps = atoi(optarg); break; case 'x': p.exp = atoi(optarg); break; case 't': p.n_threads = atoi(optarg); break; +#if NUMA + case 'a': p.bitmask_in = numa_parse_nodestring(optarg); break; + case 'b': p.bitmask_out = numa_parse_nodestring(optarg); break; + case 'c': p.numa_node_cpu = atoi(optarg); break; +#endif default: fprintf(stderr, "\nUnrecognized option!\n"); usage(); @@ -132,23 +158,76 @@ int main(int argc, char **argv) { struct Params p = input_params(argc, argv); - unsigned int nr_of_dpus = 1; - unsigned int i = 0; const unsigned int input_size = p.exp == 0 ? p.input_size * p.n_threads : p.input_size; assert(input_size % (p.n_threads) == 0 && "Input size!"); // Input/output allocation + +#if NUMA + if (p.bitmask_in) { + numa_set_membind(p.bitmask_in); + numa_free_nodemask(p.bitmask_in); + } + A = (T*) numa_alloc(input_size * sizeof(T)); +#else A = (T*)malloc(input_size * sizeof(T)); - C = (T*)malloc(input_size * sizeof(T)); - T *bufferA = A; +#endif + +#if NUMA + if (p.bitmask_out) { + numa_set_membind(p.bitmask_out); + numa_free_nodemask(p.bitmask_out); + } + C = (T*) numa_alloc(input_size * sizeof(T)); +#else + C = (T*) malloc(input_size * sizeof(T)); +#endif // Create an input file with arbitrary data. read_input(A, input_size); +#if NUMA + struct bitmask *bitmask_all = numa_allocate_nodemask(); + numa_bitmask_setall(bitmask_all); + numa_set_membind(bitmask_all); + numa_free_nodemask(bitmask_all); +#endif + +#if NUMA + mp_pages[0] = A; + if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) { + perror("move_pages(A)"); + } + else if (mp_status[0] < 0) { + printf("move_pages error: %d", mp_status[0]); + } + else { + numa_node_in = mp_status[0]; + } + + mp_pages[0] = C; + if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) { + perror("move_pages(C)"); + } + else if (mp_status[0] < 0) { + printf("move_pages error: %d", mp_status[0]); + } + else { + numa_node_out = mp_status[0]; + } + + numa_node_cpu = p.numa_node_cpu; + if (numa_node_cpu != -1) { + if (numa_run_on_node(numa_node_cpu) == -1) { + perror("numa_run_on_node"); + numa_node_cpu = -1; + } + } +#endif + // Timer declaration Timer timer; - float time_gpu = 0; thrust::omp::vector<T> h_output(input_size); @@ -189,9 +268,15 @@ int main(int argc, char **argv) { printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n"); if(rep >= p.n_warmup) { - printf("[::] SCAN-RSS CPU | n_threads=%d e_type=%s n_elements=%d " - "| throughput_cpu_ref_MBps=%f throughput_MBps=%f", + printf("[::] SCAN-RSS-CPU | n_threads=%d e_type=%s n_elements=%d" +#if NUMA + " numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d" +#endif + " | throughput_cpu_ref_MBps=%f throughput_MBps=%f", nr_threads, XSTR(T), input_size, +#if NUMA + numa_node_in, numa_node_out, numa_node_cpu, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out), +#endif input_size * sizeof(T) / timer.time[0], input_size * sizeof(T) / timer.time[1]); printf(" throughput_cpu_ref_MOpps=%f throughput_MOpps=%f", @@ -213,8 +298,13 @@ int main(int argc, char **argv) { // Deallocation +#if NUMA + numa_free(A, input_size * sizeof(T)); + numa_free(C, input_size * sizeof(T)); +#else free(A); free(C); +#endif return 0; } |