GEMV: specify memcpy cpu node

3 files changed, 87 insertions, 45 deletions

diff --git a/GEMV/baselines/cpu/gemv_openmp.c b/GEMV/baselines/cpu/gemv_openmp.c
index 1a3eed1..21e24cb 100644
--- a/GEMV/baselines/cpu/gemv_openmp.c
+++ b/GEMV/baselines/cpu/gemv_openmp.c
@@ -23,6 +23,7 @@ int numa_node_cpu = -1;
 
 #if NUMA_MEMCPY
 struct bitmask* bitmask_cpu;
+int numa_node_cpu_memcpy = -1;
 int numa_node_local = -1;
 int numa_node_in_is_local = 0;
 #endif
@@ -54,6 +55,7 @@ int main(int argc, char *argv[])
     numa_node_cpu = atoi(argv[3]);
 #if NUMA_MEMCPY
     bitmask_cpu   = numa_parse_nodestring(argv[4]);
+    numa_node_cpu_memcpy = atoi(argv[5]);
 #endif // NUMA_MEMCPY
 #else
     (void) argv;
@@ -142,7 +144,7 @@ int main(int argc, char *argv[])
 #endif
 
     Timer timer;
-    for (int i = 0; i < 40; i++) {
+    for (int i = 0; i < 20; i++) {
 
 #pragma omp parallel
         {
@@ -172,6 +174,15 @@ int main(int argc, char *argv[])
             return 1;
         }
 
+        if (!numa_node_in_is_local) {
+            if (numa_node_cpu_memcpy != -1) {
+                if (numa_run_on_node(numa_node_cpu_memcpy) == -1) {
+                    perror("numa_run_on_node");
+                    numa_node_cpu_memcpy = -1;
+                }
+            }
+        }
+
         start(&timer, 2, 0);
         if (!numa_node_in_is_local) {
             //for (size_t i=0; i < rows; i++ ) {
@@ -185,6 +196,13 @@ int main(int argc, char *argv[])
         }
         stop(&timer, 2);
 
+        if (numa_node_cpu != -1) {
+            if (numa_run_on_node(numa_node_cpu) == -1) {
+                perror("numa_run_on_node");
+                numa_node_cpu = -1;
+            }
+        }
+
         mp_pages[0] = A_local;
         if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
             perror("move_pages(A_local)");
@@ -218,10 +236,10 @@ int main(int argc, char *argv[])
 
 #if NUMA_MEMCPY
         printf("[::] GEMV-CPU-MEMCPY | n_threads=%d e_type=%s n_elements=%ld"
-            " numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_node_local=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
+            " numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_node_local=%d numa_node_cpu_memcpy=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
             " | throughput_MBps=%f throughput_MOpps=%f",
             nr_threads, XSTR(T), rows * cols,
-            numa_node_in, numa_node_out, numa_node_cpu, numa_node_local, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
+            numa_node_in, numa_node_out, numa_node_cpu, numa_node_local, numa_node_cpu_memcpy, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
             rows * cols * sizeof(T) / timer.time[0],
             rows * cols / timer.time[0]);
         printf(" latency_kernel_us=%f latency_alloc_us=%f latency_memcpy_us=%f latency_free_us=%f latency_total_us=%f\n",
diff --git a/GEMV/dimes-hetsim-hbm.sh b/GEMV/dimes-hetsim-hbm.sh
index ad6895e..bbba9e0 100755
--- a/GEMV/dimes-hetsim-hbm.sh
+++ b/GEMV/dimes-hetsim-hbm.sh
@@ -1,10 +1,9 @@
 #!/bin/bash
 
 cd baselines/cpu
-make -B NUMA=1
 
 mkdir -p log/$(hostname)
-fn=log/$(hostname)/$(date +%Y%m%d)
+fn=log/$(hostname)/dimes-hetsim-hbm
 
 # upstream DPU version uses -m 163840 -n 4096
 # → (163840 * 4096 + 4096) * uint32 ≈ 2.5 GiB
@@ -14,7 +13,7 @@ fn=log/$(hostname)/$(date +%Y%m%d)
 
 run_benchmark() {
 	local "$@"
-	OMP_NUM_THREADS=$nr_threads ./gemv $ram $ram $cpu
+	OMP_NUM_THREADS=$nr_threads ./gemv $ram_in $ram_out $cpu $ram_local $cpu_memcpy
 	return $?
 }
 
@@ -22,22 +21,36 @@ export -f run_benchmark
 
 (
 
-echo "CPU single-node (1/2)" >&2
+make -B NUMA=1 NUMA_MEMCPY=1
 
-parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
-	run_benchmark nr_threads={nr_threads} ram={ram} cpu={cpu} \
-	::: cpu $(seq 0 7) \
-	::: ram $(seq 0 15) \
-	::: nr_threads 1 2 4 8 12 16
+echo "CPU single-node operation with setup cost, memcpy node == input node, cpu node == output node (1/3)" >&2
 
-echo "CPU multi-node (2/2)" >&2
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
+	run_benchmark nr_threads={nr_threads} ram_in={ram_in} ram_out={ram_out} ram_local={ram_local} cpu={cpu} cpu_memcpy={cpu_memcpy} \
+	::: nr_threads 1 2 4 8 12 16 \
+	:::      ram_in $(seq 0 15) \
+	:::+ cpu_memcpy $(seq 0 7) $(seq 0 7) \
+	:::   ram_local $(seq 0 15) \
+	:::+        cpu $(seq 0 7) $(seq 0 7) \
+	:::+    ram_out $(seq 0 15)
 
-parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
-	run_benchmark nr_threads={nr_threads} ram={ram} cpu={cpu} \
-	::: cpu -1 \
-	::: ram $(seq 0 15) \
-	::: nr_threads 32 48 64 96 128
+make -B NUMA=1
 
-) > ${fn}.txt
+echo "single-node execution, cpu/out on same node (2/3)" >&2
+
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
+	run_benchmark nr_threads={nr_threads} ram_in={ram_in} ram_out={ram_out} cpu={cpu} \
+	::: nr_threads 1 2 4 8 12 16 \
+	:::   ram_in $(seq 0 15) \
+	:::      cpu $(seq 0 7) $(seq 0 7) \
+	:::+ ram_out $(seq 0 15)
+
+echo "multi-node execution (3/3)" >&2
+parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
+	run_benchmark nr_threads={nr_threads} ram_in={ram_in} ram_out={ram_out} cpu={cpu} \
+	::: nr_threads 32 48 64 96 128 \
+	::: cpu -1 \
+	::: ram_in $(seq 0 15) \
+	::: ram_out $(seq 0 15)
 
-xz -f -v -9 -M 800M ${fn}.txt
+) >> ${fn}.txt
diff --git a/GEMV/dimes-hetsim-nmc.sh b/GEMV/dimes-hetsim-nmc.sh
index ed01cdc..416665d 100755
--- a/GEMV/dimes-hetsim-nmc.sh
+++ b/GEMV/dimes-hetsim-nmc.sh
@@ -1,7 +1,7 @@
 #!/bin/bash
 
 mkdir -p log/$(hostname) baselines/cpu/log/$(hostname)
-fn=log/$(hostname)/$(date +%Y%m%d)
+fn=log/$(hostname)/dimes-hetsim-nmc
 
 # upstream DPU version uses -m 163840 -n 4096
 # → (163840 * 4096 + 4096) * uint32 ≈ 2.5 GiB
@@ -10,23 +10,22 @@ fn=log/$(hostname)/$(date +%Y%m%d)
 # → ≈ 1.25 GiB
 
 # Note: Upstream uses int32_t in DPU version and double in baseline.
-# Here, we use int32_t for both.
+# Here, we use (163840 * 4096 + 4096 ) * int32_t for both.
 
 run_benchmark_nmc() {
 	local "$@"
+	set -e
 	sudo limit_ranks_to_numa_node ${numa_rank}
-	if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1; then
-		bin/gemv_host -w 0 -e 100 -m 163840 -n 4096
-	fi
-	return $?
+	make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1
+	bin/gemv_host -w 0 -e 100 -m 163840 -n 4096
 }
 
 export -f run_benchmark_nmc
 
 run_benchmark_baseline() {
 	local "$@"
-	OMP_NUM_THREADS=$nr_threads ./gemv $ram $ram $cpu
-	return $?
+	set -e
+	OMP_NUM_THREADS=$nr_threads ./gemv $ram_in $ram_out $cpu $ram_local $cpu_memcpy
 }
 
 export -f run_benchmark_baseline
@@ -35,43 +34,55 @@ export -f run_benchmark_baseline
 
 echo "NMC single-node (1/2)" >&2
 
-parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
 	run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
 	::: numa_rank 0 1 \
 	::: nr_dpus 64 128 256 512 768 1024
 
 echo "NMC multi-node (2/2)" >&2
 
-parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
 	run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
 	::: numa_rank -1 \
 	::: nr_dpus 1536 2048 2304
 
-) > ${fn}.txt
-
-xz -f -v -9 -M 800M ${fn}.txt
+) >> ${fn}.txt
 
 cd baselines/cpu
-make -B NUMA=1 TYPE=int32_t
 
 (
 
-echo "CPU single-node (1/2)" >&2
+make -B NUMA=1 TYPE=int32_t NUMA_MEMCPY=1
+
+echo "CPU single-node operation with setup cost, memcpy/in and cpu/out on same node (1/3)" >&2
+
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
+	run_benchmark_baseline nr_threads={nr_threads} ram_in={ram_in} ram_out={ram_out} cpu={cpu} ram_local={ram_local} cpu_memcpy={cpu_memcpy} \
+	:::      ram_in 0 1 \
+	:::+ cpu_memcpy 0 1 \
+	:::        cpu 0 1 \
+	:::+   ram_out 0 1 \
+	:::+ ram_local 0 1 \
+	::: nr_threads 1 2 4 8 12 16
 
-parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
-	run_benchmark_baseline nr_threads={nr_threads} ram={ram} cpu={cpu} \
+make -B NUMA=1 TYPE=int32_t
+
+echo "CPU single-node (2/3)" >&2
+
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
+	run_benchmark_baseline nr_threads={nr_threads} ram_in={ram_in} ram_out={ram_out} cpu={cpu} \
 	::: cpu 0 1 \
-	::: ram 0 1 \
+	::: ram_in 0 1 \
+	::: ram_out 0 1 \
 	::: nr_threads 1 2 4 8 12 16
 
-echo "CPU multi-node (2/2)" >&2
+echo "CPU multi-node (3/3)" >&2
 
-parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
-	run_benchmark_baseline nr_threads={nr_threads} ram={ram} cpu={cpu} \
+parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
+	run_benchmark_baseline nr_threads={nr_threads} ram_in={ram_in} ram_out={ram_out} cpu={cpu} \
 	::: cpu -1 \
-	::: ram 0 1 \
+	::: ram_in 0 1 \
+	::: ram_out 0 1 \
 	::: nr_threads 24 32
 
-) > ${fn}.txt
-
-xz -f -v -9 -M 800M ${fn}.txt
+) >> ${fn}.txt