diff options
author | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-07-05 11:34:18 +0200 |
---|---|---|
committer | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-07-05 11:34:18 +0200 |
commit | f5377a304ba5390400ebff633f162212d3c7ab5d (patch) | |
tree | 41de28e6c0b5cf0d3393ae4adcc4c40572a3d45f | |
parent | 48978a7ed71b7c537015214f3ac08fb5b9d182f2 (diff) |
GEMV: log rank NUMA node
-rwxr-xr-x | GEMV/dimes-hetsim.sh | 30 | ||||
-rw-r--r-- | GEMV/host/app.c | 26 |
2 files changed, 24 insertions, 32 deletions
diff --git a/GEMV/dimes-hetsim.sh b/GEMV/dimes-hetsim.sh deleted file mode 100755 index d7163d5..0000000 --- a/GEMV/dimes-hetsim.sh +++ /dev/null @@ -1,30 +0,0 @@ -#!/bin/sh - -cd baselines/cpu -make -B NUMA=1 - -mkdir -p log/$(hostname) -fn=log/$(hostname)/$(date +%Y%m%d)-baseline.txt - -# gemv hardcodes 167772160 double elements → 1.25 GiB of data - -( - -for i in `seq 1 20`; do - for nr_threads in 1 2 4 8 12 16; do - for cpu in 0 1 2 3 4 5 6 7; do - for ram in 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15; do - OMP_NUM_THREADS=$nr_threads ./gemv $ram $ram $cpu - done - done - done - for nr_threads in 32 48 64 96 128; do - for ram in 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15; do - OMP_NUM_THREADS=$nr_threads ./gemv $ram $ram -1 - done - done -done - -) | tee $fn - -xz -f -v -9 -M 800M $fn diff --git a/GEMV/host/app.c b/GEMV/host/app.c index d96315a..ef0f2e5 100644 --- a/GEMV/host/app.c +++ b/GEMV/host/app.c @@ -18,6 +18,9 @@ #include <dpu_probe.h> #endif +#include <dpu_management.h> +#include <dpu_target_macros.h> + #define XSTR(x) STR(x) #define STR(x) #x @@ -77,6 +80,8 @@ int main(int argc, char **argv) { // Timer 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)); @@ -179,6 +184,23 @@ 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; + } + */ + } + if(rep >= p.n_warmup) { start(&timer, 2, 0); } @@ -286,8 +308,8 @@ int main(int argc, char **argv) { if (rep >= p.n_warmup) { printf("[::] GEMV UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d", NR_DPUS, nr_of_ranks, NR_TASKLETS, XSTR(T), BLOCK_SIZE, n_size * m_size); - 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], |