diff options
| -rw-r--r-- | BS/Makefile | 7 | ||||
| -rw-r--r-- | BS/host/app.c | 148 | ||||
| -rwxr-xr-x | BS/run-paper-strong-rank.sh | 3 | ||||
| -rwxr-xr-x | BS/run-paper-weak.sh | 6 | ||||
| -rwxr-xr-x | BS/run.sh | 20 | ||||
| -rwxr-xr-x | BS/support/timer.h | 6 |
6 files changed, 144 insertions, 46 deletions
diff --git a/BS/Makefile b/BS/Makefile index fae815c..90896a8 100644 --- a/BS/Makefile +++ b/BS/Makefile @@ -1,13 +1,18 @@ NR_DPUS ?= 1 NR_TASKLETS ?= 16 PROBLEM_SIZE ?= 8192 # original 2, wtf +WITH_ALLOC_OVERHEAD ?= 0 +WITH_LOAD_OVERHEAD ?= 0 +WITH_FREE_OVERHEAD ?= 0 +WITH_DPUINFO ?= 0 +SDK_SINGLETHREADED ?= 0 COMMON_INCLUDES := support HOST_SOURCES := $(wildcard host/*.c) DPU_SOURCES := $(wildcard dpu/*.c) COMMON_FLAGS := -Wall -Wextra -g -I${COMMON_INCLUDES} -HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DPROBLEM_SIZE=${PROBLEM_SIZE} +HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DPROBLEM_SIZE=${PROBLEM_SIZE} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD} -DWITH_DPUINFO=${WITH_DPUINFO} -DSDK_SINGLETHREADED=${SDK_SINGLETHREADED} DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} QUIET = @ diff --git a/BS/host/app.c b/BS/host/app.c index 1160163..bcb22c5 100644 --- a/BS/host/app.c +++ b/BS/host/app.c @@ -18,6 +18,17 @@ #include <dpu_probe.h> #endif +#if WITH_DPUINFO +#include <dpu_management.h> +#include <dpu_target_macros.h> +#endif + +#if SDK_SINGLETHREADED +#define DPU_ALLOC_PROFILE "nrThreadsPerRank=0" +#else +#define DPU_ALLOC_PROFILE NULL +#endif + #define XSTR(x) STR(x) #define STR(x) #x @@ -75,18 +86,30 @@ int main(int argc, char **argv) { struct Params p = input_params(argc, argv); struct dpu_set_t dpu_set, dpu; uint32_t nr_of_dpus; + uint32_t nr_of_ranks; uint64_t input_size = INPUT_SIZE; uint64_t num_querys = p.num_querys; DTYPE result_host = -1; DTYPE result_dpu = -1; - // Create the timer - Timer timer; + // Timer declaration + Timer timer; - // Allocate DPUs and load binary - DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set)); - DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL)); - DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus)); + // Allocate DPUs and load binary +#if !WITH_ALLOC_OVERHEAD + DPU_ASSERT(dpu_alloc(NR_DPUS, DPU_ALLOC_PROFILE, &dpu_set)); + timer.time[0] = 0; // alloc +#endif +#if !WITH_LOAD_OVERHEAD + DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL)); + DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus)); + DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks)); + assert(nr_of_dpus == NR_DPUS); + timer.time[1] = 0; // load +#endif +#if !WITH_FREE_OVERHEAD + timer.time[6] = 0; // free +#endif #if ENERGY struct dpu_probe_t probe; @@ -94,10 +117,10 @@ int main(int argc, char **argv) { #endif // Query number adjustement for proper partitioning - if(num_querys % (nr_of_dpus * NR_TASKLETS)) - num_querys = num_querys + (nr_of_dpus * NR_TASKLETS - num_querys % (nr_of_dpus * NR_TASKLETS)); + if(num_querys % (NR_DPUS * NR_TASKLETS)) + num_querys = num_querys + (NR_DPUS * NR_TASKLETS - num_querys % (NR_DPUS * NR_TASKLETS)); - assert(num_querys % (nr_of_dpus * NR_TASKLETS) == 0 && "Input dimension"); // Allocate input and querys vectors + assert(num_querys % (NR_DPUS * NR_TASKLETS) == 0 && "Input dimension"); // Allocate input and querys vectors DTYPE * input = malloc((input_size) * sizeof(DTYPE)); DTYPE * querys = malloc((num_querys) * sizeof(DTYPE)); @@ -106,20 +129,56 @@ int main(int argc, char **argv) { create_test_file(input, querys, input_size, num_querys); // Create kernel arguments - uint64_t slice_per_dpu = num_querys / nr_of_dpus; + uint64_t slice_per_dpu = num_querys / NR_DPUS; dpu_arguments_t input_arguments = {input_size, slice_per_dpu, 0}; for (unsigned int rep = 0; rep < p.n_warmup + p.n_reps; rep++) { // Perform input transfers uint64_t i = 0; +#if WITH_ALLOC_OVERHEAD + if(rep >= p.n_warmup) { + start(&timer, 0, 0); + } + DPU_ASSERT(dpu_alloc(NR_DPUS, DPU_ALLOC_PROFILE, &dpu_set)); + if(rep >= p.n_warmup) { + stop(&timer, 0); + } +#endif +#if WITH_DPUINFO + printf("DPUs:"); + DPU_FOREACH (dpu_set, dpu) { + int rank = dpu_get_rank_id(dpu_get_rank(dpu_from_set(dpu))) & DPU_TARGET_MASK; + int slice = dpu_get_slice_id(dpu_from_set(dpu)); + int member = dpu_get_member_id(dpu_from_set(dpu)); + printf(" %d(%d.%d)", rank, slice, member); + } + printf("\n"); +#endif +#if WITH_LOAD_OVERHEAD + if(rep >= p.n_warmup) { + start(&timer, 1, 0); + } + DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL)); + if(rep >= p.n_warmup) { + stop(&timer, 1); + } + DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus)); + DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks)); + assert(nr_of_dpus == NR_DPUS); +#endif + // Compute host solution - start(&timer, 0, 0); + if(rep >= p.n_warmup) { + start(&timer, 2, 0); + } result_host = binarySearch(input, querys, input_size - 1, num_querys); - stop(&timer, 0); + if(rep >= p.n_warmup) { + stop(&timer, 2); + } if (rep >= p.n_warmup) { - start(&timer, 1, 0); + start(&timer, 3, 0); } DPU_FOREACH(dpu_set, dpu, i) @@ -148,13 +207,13 @@ int main(int argc, char **argv) { DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, input_size * sizeof(DTYPE), slice_per_dpu * sizeof(DTYPE), DPU_XFER_DEFAULT)); if (rep >= p.n_warmup) { - stop(&timer, 1); + stop(&timer, 3); } // Run kernel on DPUs if (rep >= p.n_warmup) { - start(&timer, 2, 0); + start(&timer, 4, 0); #if ENERGY DPU_ASSERT(dpu_probe_start(&probe)); #endif @@ -164,7 +223,7 @@ int main(int argc, char **argv) { if (rep >= p.n_warmup) { - stop(&timer, 2); + stop(&timer, 4); #if ENERGY DPU_ASSERT(dpu_probe_stop(&probe)); #endif @@ -183,9 +242,9 @@ int main(int argc, char **argv) { // Retrieve results if (rep >= p.n_warmup) { - start(&timer, 3, 0); + start(&timer, 5, 0); } - dpu_results_t* results_retrieve[nr_of_dpus]; + dpu_results_t* results_retrieve[NR_DPUS]; i = 0; DPU_FOREACH(dpu_set, dpu, i) { @@ -207,24 +266,53 @@ int main(int argc, char **argv) { free(results_retrieve[i]); } if(rep >= p.n_warmup) { - stop(&timer, 3); + stop(&timer, 5); + } + +#if WITH_ALLOC_OVERHEAD +#if WITH_FREE_OVERHEAD + if(rep >= p.n_warmup) { + start(&timer, 6, 0); } +#endif + DPU_ASSERT(dpu_free(dpu_set)); +#if WITH_FREE_OVERHEAD + if(rep >= p.n_warmup) { + stop(&timer, 6); + } +#endif +#endif int status = (result_dpu == result_host); if (status) { printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] results are equal\n"); if (rep >= p.n_warmup) { - printf("[::] BS NMC | n_dpus=%d n_tasklets=%d e_type=%s n_elements=%lu " - "| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f", - nr_of_dpus, NR_TASKLETS, XSTR(DTYPE), input_size, - num_querys * sizeof(DTYPE) / timer.time[0], - num_querys * sizeof(DTYPE) / timer.time[2], - num_querys * sizeof(DTYPE) / (timer.time[1] + timer.time[2] + timer.time[3])); - printf(" throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f", - num_querys / timer.time[0], - num_querys / timer.time[2], - num_querys / (timer.time[1] + timer.time[2] + timer.time[3])); - printall(&timer, 3); + printf("[::] BS UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%lu b_sdk_singlethreaded=%d ", + NR_DPUS, nr_of_ranks, NR_TASKLETS, XSTR(DTYPE), BLOCK_SIZE, input_size, SDK_SINGLETHREADED); + 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], + timer.time[2], + timer.time[3], + timer.time[4], + timer.time[5], + timer.time[6]); + printf(" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f", + input_size * sizeof(DTYPE) / timer.time[2], + input_size * sizeof(DTYPE) / (timer.time[4]), + input_size * sizeof(DTYPE) / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6])); + printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f", + input_size * sizeof(DTYPE) / (timer.time[3] + timer.time[4] + timer.time[5]), + input_size * sizeof(DTYPE) / (timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]), + input_size * sizeof(DTYPE) / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5])); + printf(" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f", + input_size / timer.time[2], + input_size / (timer.time[4]), + input_size / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6])); + printf(" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f\n", + input_size / (timer.time[3] + timer.time[4] + timer.time[5]), + input_size / (timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]), + input_size / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5])); } } else { printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] results differ!\n"); diff --git a/BS/run-paper-strong-rank.sh b/BS/run-paper-strong-rank.sh index 824b279..c2d4f36 100755 --- a/BS/run-paper-strong-rank.sh +++ b/BS/run-paper-strong-rank.sh @@ -14,8 +14,7 @@ echo "prim-benchmarks BS strong-rank (dfatool edition)" echo "Started at $(date)" echo "Revision $(git describe --always)" -# >64 are not part of upstream -for nr_dpus in 128 1 4 16 64; do +for nr_dpus in 1 4 16 64; do for nr_tasklets in 1 2 4 8 16; do echo if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 verbose=1; then diff --git a/BS/run-paper-weak.sh b/BS/run-paper-weak.sh index f845e30..a27c547 100755 --- a/BS/run-paper-weak.sh +++ b/BS/run-paper-weak.sh @@ -15,11 +15,11 @@ echo "prim-benchmarks BS weak (dfatool edition)" echo "Started at $(date)" echo "Revision $(git describe --always)" -# 256 and 512 are not part of upstream -for nr_dpus in 256 512 1 4 16 64; do +for nr_dpus in 1 4 16 64; do for nr_tasklets in 1 2 4 8 16; do echo - if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 verbose=1; then + # original Makefile sets PROBLEM_SIZE=2, for some reason. + if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 verbose=1 PROBLEM_SIZE=2; then i=$(( nr_dpus * 262144 )) timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 -i $i || true fi @@ -8,15 +8,21 @@ set -e # -e: number of timed iterations # -i; ignored, always uses 262144 elements -echo "prim-benchmarks VA (dfatool edition)" +( + +echo "prim-benchmarks BS (dfatool edition)" echo "Started at $(date)" echo "Revision $(git describe --always)" -for nr_dpus in 1 2 4 8 16 32 64 128 256 512; do - for nr_tasklets in 1 2 3 4 6 8 10 12 16 20 24; do - echo - if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then - timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 || true - fi +for i in 262144 16777216; do + for nr_dpus in 1 4 8 16 32 64 128 256 512 768 1024 1536 2048 2304 2542; do + for nr_tasklets in 8 12 16; do + echo + if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then + timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 -i $i || true + fi + done done done +echo "Completed at $(date)" +) | tee "log-$(hostname)-ndpus.txt" diff --git a/BS/support/timer.h b/BS/support/timer.h index 08027af..ff1ae1b 100755 --- a/BS/support/timer.h +++ b/BS/support/timer.h @@ -37,9 +37,9 @@ typedef struct Timer{
- struct timeval startTime[4];
- struct timeval stopTime[4];
- double time[4];
+ struct timeval startTime[7];
+ struct timeval stopTime[7];
+ double time[7];
}Timer;
|