diff options
| author | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-07-05 09:28:49 +0200 |
|---|---|---|
| committer | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-07-05 09:29:20 +0200 |
| commit | 3cd92eca2c472211a4974fb477f11843fdb74393 (patch) | |
| tree | 96c43efcee9995d57a3c87cab0fd0be335ec9ce3 /VA/host | |
| parent | 8bf22d92554c172fbd18db9fb1266d8e72678312 (diff) | |
VA: log NUMA node of allocated rank(s)
Diffstat (limited to 'VA/host')
| -rw-r--r-- | VA/host/app.c | 26 |
1 files changed, 22 insertions, 4 deletions
diff --git a/VA/host/app.c b/VA/host/app.c index d363514..5fe3f61 100644 --- a/VA/host/app.c +++ b/VA/host/app.c @@ -29,10 +29,8 @@ #include <dpu_probe.h> #endif -#if WITH_DPUINFO #include <dpu_management.h> #include <dpu_target_macros.h> -#endif // Pointer declaration static T* A; @@ -73,6 +71,8 @@ int main(int argc, char **argv) { // Timer declaration Timer timer; + int numa_node_rank = -2; + // Allocate DPUs and load binary #if !WITH_ALLOC_OVERHEAD DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set)); @@ -144,6 +144,24 @@ int main(int argc, char **argv) { assert(nr_of_dpus == NR_DPUS); #endif + // int prev_rank_id = -1; + int rank_id = -1; + DPU_FOREACH (dpu_set, dpu) { + rank_id = dpu_get_rank_id(dpu_get_rank(dpu_from_set(dpu))) & DPU_TARGET_MASK; + if ((numa_node_rank != -2) && numa_node_rank != dpu_get_rank_numa_node(dpu_get_rank(dpu_from_set(dpu)))) { + numa_node_rank = -1; + } else { + numa_node_rank = dpu_get_rank_numa_node(dpu_get_rank(dpu_from_set(dpu))); + } + /* + if (rank_id != prev_rank_id) { + printf("/dev/dpu_rank%d @ NUMA node %d\n", rank_id, numa_node_rank); + prev_rank_id = rank_id; + } + */ + } + + // Compute output on CPU (performance comparison and verification purposes) if(rep >= p.n_warmup) { start(&timer, 2, 0); @@ -257,8 +275,8 @@ int main(int argc, char **argv) { if (rep >= p.n_warmup) { printf("[::] VA-UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d n_elements_per_dpu=%d", nr_of_dpus, nr_of_ranks, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size, input_size / NR_DPUS); - printf(" b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d ", - WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD); + printf(" b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d numa_node_rank=%d ", + WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD, numa_node_rank); printf("| latency_alloc_us=%f latency_load_us=%f latency_cpu_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f latency_free_us=%f", timer.time[0], timer.time[1], |