BS baseline: Add memcpy variant

4 files changed, 153 insertions, 42 deletions

diff --git a/BS/baselines/cpu/Makefile b/BS/baselines/cpu/Makefile
index bed2561..b67602f 100644
--- a/BS/baselines/cpu/Makefile
+++ b/BS/baselines/cpu/Makefile
@@ -1,4 +1,5 @@
 NUMA ?= 0
+NUMA_MEMCPY ?= 0
 FLAGS =
 
 ifeq (${NUMA}, 1)
@@ -9,7 +10,7 @@ endif
 all: bs_omp
 
 bs_omp: bs_omp.c
-	gcc -Wall -Wextra -pedantic -march=native -O2 -DNUMA=${NUMA} bs_omp.c -o bs_omp -fopenmp ${FLAGS}
+	gcc -Wall -Wextra -pedantic -march=native -O2 -DNUMA=${NUMA} -DNUMA_MEMCPY=${NUMA_MEMCPY} bs_omp.c -o bs_omp -fopenmp ${FLAGS}
 
 bs_omp_O0: bs_omp.c
 	gcc bs_omp.c -o bs_omp_O0 -fopenmp
diff --git a/BS/baselines/cpu/bs_omp.c b/BS/baselines/cpu/bs_omp.c
index 1b9947c..874299b 100644
--- a/BS/baselines/cpu/bs_omp.c
+++ b/BS/baselines/cpu/bs_omp.c
@@ -21,6 +21,15 @@ int numa_node_in = -1;
 int numa_node_cpu = -1;
 #endif
 
+
+#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
+
+
 #define DTYPE uint64_t
 /*
 * @brief creates a "test file" by filling a bufferwith values
@@ -90,6 +99,10 @@ uint64_t binarySearch(DTYPE * input, uint64_t input_size, DTYPE* querys, unsigne
     bitmask_in = numa_parse_nodestring(argv[3]);
     numa_node_cpu = atoi(argv[4]);
 #endif
+#if NUMA_MEMCPY
+    bitmask_cpu = numa_parse_nodestring(argv[5]);
+    numa_node_cpu_memcpy = atoi(argv[6]);
+#endif
 
     printf("Vector size: %lu, num searches: %lu\n", input_size, n_querys);
 
@@ -106,10 +119,17 @@ uint64_t binarySearch(DTYPE * input, uint64_t input_size, DTYPE* querys, unsigne
 #endif
 
 #if NUMA
+#if NUMA_MEMCPY
+    if (bitmask_cpu) {
+        numa_set_membind(bitmask_cpu);
+        numa_free_nodemask(bitmask_cpu);
+    }
+#else
     struct bitmask *bitmask_all = numa_allocate_nodemask();
     numa_bitmask_setall(bitmask_all);
     numa_set_membind(bitmask_all);
     numa_free_nodemask(bitmask_all);
+#endif // NUMA_MEMCPY
 #endif
 
     DTYPE result_host = -1;
@@ -137,10 +157,71 @@ uint64_t binarySearch(DTYPE * input, uint64_t input_size, DTYPE* querys, unsigne
     }
 #endif
 
+#if NUMA_MEMCPY
+    numa_node_in_is_local = ((numa_node_cpu == numa_node_in) || (numa_node_cpu + 8 == numa_node_in)) * 1;
+#endif
+
+#if NUMA_MEMCPY
+    DTYPE *input_local = input;
+    DTYPE *querys_local = querys;
+    start(&timer, 1, 0);
+    if (!numa_node_in_is_local) {
+        input_local = numa_alloc((input_size) * sizeof(DTYPE));
+        querys_local = numa_alloc((n_querys) * sizeof(DTYPE));
+    }
+    stop(&timer, 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) {
+        memcpy(input_local, input, input_size * sizeof(DTYPE));
+        memcpy(querys_local, querys, n_querys * sizeof(DTYPE));
+    } else {
+        input_local = input;
+        querys_local = querys;
+    }
+    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] = input_local;
+    if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
+        perror("move_pages(input_local)");
+    }
+    else if (mp_status[0] < 0) {
+        printf("move_pages error: %d", mp_status[0]);
+    }
+    else {
+        numa_node_local = mp_status[0];
+    }
+#endif
+
     start(&timer, 0, 0);
+#if NUMA_MEMCPY
+    result_host = binarySearch(input_local, input_size - 1, querys_local, n_querys);
+#else
     result_host = binarySearch(input, input_size - 1, querys, n_querys);
+#endif
     stop(&timer, 0);
 
+#if NUMA_MEMCPY
+    start(&timer, 3, 0);
+    if (!numa_node_in_is_local) {
+        numa_free(input_local, input_size * sizeof(DTYPE));
+        numa_free(querys_local, n_querys * sizeof(DTYPE));
+    }
+    stop(&timer, 3);
+#endif
+
     unsigned int nr_threads = 0;
 #pragma omp parallel
 #pragma omp atomic
@@ -148,6 +229,17 @@ uint64_t binarySearch(DTYPE * input, uint64_t input_size, DTYPE* querys, unsigne
 
     int status = (result_host);
     if (status) {
+#if NUMA_MEMCPY
+        printf("[::] BS-CPU-MEMCPY | n_threads=%d e_type=%s n_elements=%lu"
+            " numa_node_in=%d numa_node_local=%d numa_node_cpu=%d numa_node_cpu_memcpy=%d numa_distance_in_cpu=%d"
+            " | throughput_MBps=%f throughput_MOpps=%f"
+            " latency_kernel_us=%f latency_alloc_us=%f latency_memcpy_us=%f latency_free_us=%f latency_total_us=%f\n",
+            nr_threads, "uint64_t", input_size,
+            numa_node_in, numa_node_local, numa_node_cpu, numa_node_cpu_memcpy, numa_distance(numa_node_in, numa_node_cpu),
+            n_querys * sizeof(DTYPE) / timer.time[0], n_querys / timer.time[0],
+            timer.time[0], timer.time[1], timer.time[2], timer.time[3],
+            timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3]);
+#else
         printf("[::] BS-CPU | n_threads=%d e_type=%s n_elements=%lu"
 #if NUMA
             " numa_node_in=%d numa_node_cpu=%d numa_distance_in_cpu=%d"
@@ -158,9 +250,9 @@ uint64_t binarySearch(DTYPE * input, uint64_t input_size, DTYPE* querys, unsigne
             numa_node_in, numa_node_cpu, numa_distance(numa_node_in, numa_node_cpu),
 #endif
             n_querys * sizeof(DTYPE) / timer.time[0]);
-        printf(" throughput_MOpps=%f",
-            n_querys / timer.time[0]);
-        printall(&timer, 0);
+        printf(" throughput_MOpps=%f latency_us=%f\n",
+            n_querys / timer.time[0], timer.time[0]);
+#endif
     } else {
         printf("[ERROR]\n");
     }
diff --git a/BS/dimes-hetsim-hbm.sh b/BS/dimes-hetsim-hbm.sh
index 2aabe15..4e1500d 100755
--- a/BS/dimes-hetsim-hbm.sh
+++ b/BS/dimes-hetsim-hbm.sh
@@ -1,12 +1,10 @@
 #!/bin/bash
 
-echo PrIM BS
-
 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 2048576 * uint64 ≈ 16 MiB (DPU max: 64 MiB)
 # upstream DPU version uses 2 queries
@@ -32,7 +30,7 @@ export -f run_benchmark
 
 echo "single-node execution, DPU ref (1/4)" >&2
 
-parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
 	run_benchmark i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
 	input_size=${input_size_dpu} num_queries=${num_queries_dpu} \
 	::: i $(seq 1 5) \
@@ -42,7 +40,7 @@ parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
 
 echo "single-node execution, HBM ref (2/4)" >&2
 
-parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
 	run_benchmark i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
 	input_size=${input_size_hbm} num_queries=${num_queries_hbm} \
 	::: i $(seq 1 5) \
@@ -52,7 +50,7 @@ parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
 
 echo "multi-node execution, DPU ref (3/4)" >&2
 
-parallel -j1 --eta --joblog ${fn}.3.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
 	run_benchmark i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
 	input_size=${input_size_dpu} num_queries=${num_queries_dpu} \
 	::: i $(seq 1 40) \
@@ -62,7 +60,7 @@ parallel -j1 --eta --joblog ${fn}.3.joblog --header : \
 
 echo "multi-node execution, HBM ref (4/4)" >&2
 
-parallel -j1 --eta --joblog ${fn}.4.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \
 	run_benchmark i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
 	input_size=${input_size_hbm} num_queries=${num_queries_hbm} \
 	::: i $(seq 1 40) \
@@ -70,6 +68,4 @@ parallel -j1 --eta --joblog ${fn}.4.joblog --header : \
 	::: cpu -1 \
 	::: ram $(seq 0 15)
 
-) > ${fn}.txt
-
-xz -f -v -9 -M 800M ${fn}.txt
+) >> ${fn}.txt
diff --git a/BS/dimes-hetsim-nmc.sh b/BS/dimes-hetsim-nmc.sh
index 2313985..195334b 100755
--- a/BS/dimes-hetsim-nmc.sh
+++ b/BS/dimes-hetsim-nmc.sh
@@ -1,9 +1,7 @@
 #!/bin/bash
 
-echo PrIM BS
-
 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 2048576 * uint64 ≈ 16 MiB (DPU max: 64 MiB)
 # upstream DPU version uses 2 queries
@@ -15,19 +13,18 @@ num_queries_dpu=1048576
 
 run_benchmark_nmc() {
 	local "$@"
+	set -e
 	sudo limit_ranks_to_numa_node ${numa_rank}
-	if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1 INPUT_SIZE=${input_size} PROBLEM_SIZE=${num_queries}; then
-		bin/bs_host -w 0 -e 100 2>&1
-	fi
-	return $?
+	make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1 INPUT_SIZE=${input_size} PROBLEM_SIZE=${num_queries}
+	bin/bs_host -w 0 -e 100 2>&1
 }
 
 export -f run_benchmark_nmc
 
 run_benchmark_baseline() {
 	local "$@"
-	OMP_NUM_THREADS=${nr_threads} ./bs_omp ${input_size} ${num_queries} ${ram} ${cpu} 2>&1
-	return $?
+	set -e
+	OMP_NUM_THREADS=${nr_threads} ./bs_omp ${input_size} ${num_queries} ${ram} ${cpu} ${ram_local} ${cpu_memcpy} 2>&1
 }
 
 export -f run_benchmark_baseline
@@ -36,7 +33,7 @@ export -f run_benchmark_baseline
 
 echo "NMC single-node upstream-ref (1/4)" >&2
 
-parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
 	run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
 	num_queries=${num_queries_upstream} input_size=${input_size_upstream} \
 	::: numa_rank 0 1 \
@@ -44,7 +41,7 @@ parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
 
 echo "NMC multi-node upstream-ref (2/4)" >&2
 
-parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
 	run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
 	num_queries=${num_queries_upstream} input_size=${input_size_upstream} \
 	::: numa_rank -1 \
@@ -52,7 +49,7 @@ parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
 
 echo "NMC single-node DPU-ref (3/4)" >&2
 
-parallel -j1 --eta --joblog ${fn}.3.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
 	run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
 	num_queries=${num_queries_dpu} input_size=${input_size_dpu} \
 	::: numa_rank 0 1 \
@@ -60,24 +57,51 @@ parallel -j1 --eta --joblog ${fn}.3.joblog --header : \
 
 echo "NMC multi-node DPU-ref (4/4)" >&2
 
-parallel -j1 --eta --joblog ${fn}.4.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \
 	run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
 	num_queries=${num_queries_dpu} input_size=${input_size_dpu} \
 	::: 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
 
 (
 
-echo "CPU single-node upstream-ref (1/4)" >&2
+make -B NUMA=1 NUMA_MEMCPY=1
 
-parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
+echo "CPU single-node upstream-ref with memcpy, copy node == input node (1/6)" >&2
+
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
+	run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
+	ram_local={ram_local} cpu_memcpy={cpu_memcpy} \
+	num_queries=${num_queries_upstream} input_size=${input_size_upstream} \
+	::: i $(seq 1 20) \
+	:::         ram 0 1 \
+	:::+ cpu_memcpy 0 1 \
+	::: ram_local 0 1 \
+	:::+      cpu 0 1 \
+	::: nr_threads 1 2 4 8 12 16
+
+echo "CPU single-node dpu-ref with memcpy, copy node == input node (2/6)" >&2
+
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
+	run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
+	ram_local={ram_local} cpu_memcpy={cpu_memcpy} \
+	num_queries=${num_queries_dpu} input_size=${input_size_dpu} \
+	::: i $(seq 1 20) \
+	:::         ram 0 1 \
+	:::+ cpu_memcpy 0 1 \
+	::: ram_local 0 1 \
+	:::+      cpu 0 1 \
+	::: nr_threads 1 2 4 8 12 16
+
+make -B NUMA=1
+
+echo "CPU single-node upstream-ref (3/6)" >&2
+
+parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
 	run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
 	num_queries=${num_queries_upstream} input_size=${input_size_upstream} \
 	::: i $(seq 1 20) \
@@ -85,9 +109,9 @@ parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
 	::: ram 0 1 \
 	::: nr_threads 1 2 4 8 12 16
 
-echo "CPU single-node DPU-ref (2/4)" >&2
+echo "CPU single-node DPU-ref (4/6)" >&2
 
-parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \
 	run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
 	num_queries=${num_queries_dpu} input_size=${input_size_dpu} \
 	::: i $(seq 1 20) \
@@ -95,9 +119,9 @@ parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
 	::: ram 0 1 \
 	::: nr_threads 1 2 4 8 12 16
 
-echo "CPU multi-node upstream-ref (3/4)" >&2
+echo "CPU multi-node upstream-ref (5/6)" >&2
 
-parallel -j1 --eta --joblog ${fn}.3.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.5.joblog --resume --header : \
 	run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
 	num_queries=${num_queries_upstream} input_size=${input_size_upstream} \
 	::: i $(seq 1 20) \
@@ -105,9 +129,9 @@ parallel -j1 --eta --joblog ${fn}.3.joblog --header : \
 	::: ram 0 1 \
 	::: nr_threads 24 32
 
-echo "CPU multi-node DPU-ref (4/4)" >&2
+echo "CPU multi-node DPU-ref (6/6)" >&2
 
-parallel -j1 --eta --joblog ${fn}.4.joblog --header : \
+parallel -j1 --eta --joblog ${fn}.6.joblog --resume --header : \
 	run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
 	num_queries=${num_queries_dpu} input_size=${input_size_dpu} \
 	::: i $(seq 1 20) \
@@ -115,6 +139,4 @@ parallel -j1 --eta --joblog ${fn}.4.joblog --header : \
 	::: ram 0 1 \
 	::: nr_threads 24 32
 
-) > ${fn}.txt
-
-xz -f -v -9 -M 800M ${fn}.txt
+) >> ${fn}.txt