SCAN-RSS: NUMA support

5 files changed, 231 insertions, 16 deletions

diff --git a/SCAN-RSS/baselines/cpu/Makefile b/SCAN-RSS/baselines/cpu/Makefile
index 0bdedf6..88ba57d 100644
--- a/SCAN-RSS/baselines/cpu/Makefile
+++ b/SCAN-RSS/baselines/cpu/Makefile
@@ -1,9 +1,16 @@
+NUMA ?= 0
+FLAGS =
+
+ifeq (${NUMA}, 1)
+	FLAGS += -lnuma
+endif
+
 TYPE ?= UINT64
 
 all: scan
 
 scan: app_baseline.cpp
-	g++ -O2 app_baseline.cpp -fopenmp -DTHRUST_HOST_SYSTEM=THRUST_HOST_SYSTEM_CPP -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_OMP -lgomp -I/usr/local/cuda-8.0/include -lm -o scan -D${TYPE}
+	g++ -Wall -Wextra -pedantic -march=native -O2 app_baseline.cpp -fopenmp -DTHRUST_HOST_SYSTEM=THRUST_HOST_SYSTEM_CPP -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_OMP -lgomp -I/usr/local/cuda-8.0/include -lm -o scan -DNUMA=${NUMA} -D${TYPE} ${FLAGS}
 
 run: scan
 	./scan -i 1258291200 -t 4
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;
 }
diff --git a/SCAN-RSS/dimes-hetsim-hbm.sh b/SCAN-RSS/dimes-hetsim-hbm.sh
new file mode 100755
index 0000000..0a198f5
--- /dev/null
+++ b/SCAN-RSS/dimes-hetsim-hbm.sh
@@ -0,0 +1,33 @@
+#!/bin/sh
+
+cd baselines/cpu
+make -B NUMA=1
+
+mkdir -p log/$(hostname)
+fn=log/$(hostname)/$(date +%Y%m%d)
+
+# upstream uses 251658240 * INT64 == 1.875 GiB
+
+(
+
+echo "single-node execution (1/2)" >&2
+
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
+	./scan -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 5 -x 1 \
+	::: nr_threads 1 2 4 8 12 16 \
+	::: cpu $(seq 0 7) \
+	::: ram $(seq 0 15) \
+	::: input_size 251658240
+
+echo "multi-node execution (2/2)" >&2
+
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
+	./scan -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 -x 1 \
+	::: nr_threads 32 48 64 96 128 \
+	::: cpu -1 \
+	::: ram $(seq 0 15) \
+	::: input_size 251658240
+
+) > ${fn}.txt
+
+xz -f -v -9 -M 800M ${fn}.txt
diff --git a/SCAN-RSS/dimes-hetsim-nmc.sh b/SCAN-RSS/dimes-hetsim-nmc.sh
new file mode 100755
index 0000000..820ca4d
--- /dev/null
+++ b/SCAN-RSS/dimes-hetsim-nmc.sh
@@ -0,0 +1,66 @@
+#!/bin/bash
+
+mkdir -p log/$(hostname) baselines/cpu/log/$(hostname)
+fn=log/$(hostname)/$(date +%Y%m%d)
+
+# upstream uses 251658240 * INT64 == 1.875 GiB
+
+run_benchmark_nmc() {
+	local "$@"
+	sudo limit_ranks_to_numa_node ${numa_rank}
+	if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1; then
+		bin/host_code -w 0 -e 100 -i ${input_size} -x 1
+	fi
+	return $?
+}
+
+export -f run_benchmark_nmc
+
+(
+
+echo "NMC single-node operation (1/2)" >&2
+
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
+	run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 input_size={input_size} numa_rank={numa_rank} \
+	::: numa_rank 0 1 \
+	::: nr_dpus 64 128 256 512 768 1024 \
+	::: input_size 251658240
+
+echo "NMC multi-node operation (2/2)" >&2
+
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
+	run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 input_size={input_size} numa_rank={numa_rank} \
+	::: numa_rank any \
+	::: nr_dpus 1536 2048 2304 \
+	::: input_size 251658240
+
+) > ${fn}.txt
+
+xz -f -v -9 -M 800M ${fn}.txt
+
+cd baselines/cpu
+make -B NUMA=1
+
+(
+
+echo "CPU single-node operation (1/2)" >&2
+
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
+	./scan -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 -x 1 \
+	::: ram 0 1 \
+	::: cpu 0 1 \
+	::: nr_threads 1 2 4 8 12 16 32 \
+	::: input_size 251658240
+
+echo "CPU multi-node operation (1/2)" >&2
+
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
+	./scan -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 -x 1 \
+	::: ram 0-1 \
+	::: cpu -1 \
+	::: nr_threads 48 64 \
+	::: input_size 251658240
+
+) > ${fn}.txt
+
+xz -f -v -9 -M 800M ${fn}.txt
diff --git a/SCAN-RSS/host/app.c b/SCAN-RSS/host/app.c
index 93c2c47..6771207 100644
--- a/SCAN-RSS/host/app.c
+++ b/SCAN-RSS/host/app.c
@@ -29,10 +29,12 @@
 #include <dpu_probe.h>
 #endif
 
+#include <dpu_management.h>
+#include <dpu_target_macros.h>
+
 // Pointer declaration
 static T* A;
 static T* C;
-static T* C2;
 
 // Create input arrays
 static void read_input(T* A, unsigned int nr_elements, unsigned int nr_elements_round) {
@@ -71,6 +73,8 @@ 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));
@@ -98,9 +102,8 @@ int main(int argc, char **argv) {
     // Input/output allocation
     A = malloc(input_size_dpu_round * NR_DPUS * sizeof(T));
     C = malloc(input_size_dpu_round * NR_DPUS * sizeof(T));
-    C2 = malloc(input_size_dpu_round * NR_DPUS * sizeof(T));
     T *bufferA = A;
-    T *bufferC = C2;
+    T *bufferC = C;
 
     // Create an input file with arbitrary data
     read_input(A, input_size, input_size_dpu_round * NR_DPUS);
@@ -132,6 +135,23 @@ 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;
+            }
+            */
+        }
+
         // Compute output on CPU (performance comparison and verification purposes)
         if(rep >= p.n_warmup) {
             start(&timer, 2, 0);
@@ -314,8 +334,8 @@ int main(int argc, char **argv) {
             printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
             printf("[::] SCAN-RSS UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d b_unroll=%d n_elements=%d",
                 NR_DPUS, nr_of_ranks, NR_TASKLETS, XSTR(T), BLOCK_SIZE, UNROLL, input_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_sync_us=%f latency_read_us=%f latency_free_us=%f",
                 timer.time[0],
                 timer.time[1],
@@ -359,7 +379,6 @@ int main(int argc, char **argv) {
     // Deallocation
     free(A);
     free(C);
-    free(C2);
 
 #if !WITH_ALLOC_OVERHEAD
     DPU_ASSERT(dpu_free(dpu_set));