Brandon Potter
106 行
3.2 KiB
Makefile
106 行
3.2 KiB
Makefile
HIPCC=hipcc
|
|
BUILD=./build
|
|
SRC=./src
|
|
RESULTS=./results
|
|
|
|
#rocshmem_DIR=${HOME}/rocshmem
|
|
#MPI_HOME=${HOME}/mpich/install
|
|
NCCL_HOME=${HOME}/rccl/build
|
|
|
|
MPI_FLAGS=-lmpi -lhsa-runtime64 -lrt -L${MPI_HOME}/lib -fgpu-rdc
|
|
SHMEM_FLAGS=${MPI_FLAGS} -lmlx5 -libverbs
|
|
RCCL_FLAGS=${MPI_FLAGS} -Wl,-rpath,$(NCCL_HOME) -L${NCCL_HOME} -lrccl
|
|
|
|
.SILENT: run_scan extract_scan run_sort run_sort_shmem run_sort_rccl extract_sort
|
|
|
|
all: ${BUILD}/sort_shmem ${BUILD}/sort_rccl ${BUILD}/sort_mpi
|
|
|
|
${BUILD}/sort_shmem: ${BUILD}/sort_shmem.o ${rocshmem_DIR}/lib/librocshmem.a
|
|
${HIPCC} $^ ${SHMEM_FLAGS} -o $@
|
|
|
|
${BUILD}/sort_shmem.o: ${SRC}/sort_shmem.cu
|
|
${HIPCC} $^ -I${rocshmem_DIR}/include -I${MPI_HOME}/include -fgpu-rdc -o $@ -c
|
|
|
|
${BUILD}/sort_rccl: ${BUILD}/sort_rccl.o
|
|
${HIPCC} $^ ${RCCL_FLAGS} -o $@
|
|
|
|
${BUILD}/sort_rccl.o: ${SRC}/sort_rccl.cu
|
|
${HIPCC} $^ -I$(NCCL_HOME)/include/rccl -I${MPI_HOME}/include -fgpu-rdc -o $@ -c
|
|
|
|
${BUILD}/sort_mpi: ${BUILD}/sort_mpi.o
|
|
${HIPCC} $^ ${MPI_FLAGS} -o $@
|
|
|
|
${BUILD}/sort_mpi.o: ${SRC}/sort_mpi.cu
|
|
${HIPCC} $^ -I${MPI_HOME}/include -fgpu-rdc -o $@ -c
|
|
|
|
RO_FLAGS=ROC_SHMEM_RO=1 RO_NET_CPU_QUEUE=1
|
|
ITERS?=0 1 2 3 4 5 6 7 8 9
|
|
TIMEOUT=1m
|
|
HOSTS=sv-pdp-0,sv-pdp-1,sv-pdp-2,sv-pdp-3
|
|
SCAN_SIZE=1024
|
|
PES=2 4 8 12 16
|
|
PES_RCCL=2 4 8
|
|
|
|
TYPE ?= Naive
|
|
LABEL ?= naive
|
|
PARAM ?= 0
|
|
NUM_PES ?= 2
|
|
|
|
run_sort_shmem: ${BUILD}/sort_shmem
|
|
printf "${TYPE} ";\
|
|
echo "" > ${RESULTS}/sort_${LABEL}_${NUM_PES}.out; \
|
|
for j in ${ITERS}; do \
|
|
${RO_FLAGS} timeout ${TIMEOUT} mpirun -np ${NUM_PES} -hosts ${HOSTS} ${BUILD}/sort_shmem ${PARAM} >> ${RESULTS}/sort_${LABEL}_${NUM_PES}.out;\
|
|
done;
|
|
|
|
run_sort_rccl: ${BUILD}/sort_rccl
|
|
printf "RCCL "; \
|
|
echo "" > ${RESULTS}/sort_rccl_${NUM_PES}.out; \
|
|
for j in ${ITERS}; do \
|
|
timeout ${TIMEOUT} mpirun -np ${NUM_PES} -hosts ${HOSTS} ${BUILD}/sort_rccl >> ${RESULTS}/sort_rccl_${NUM_PES}.out;\
|
|
done;
|
|
|
|
run_sort_mpi: ${BUILD}/sort_rccl
|
|
printf "MPI2 "; \
|
|
echo "" > ${RESULTS}/sort_mpi2_${NUM_PES}.out; \
|
|
for j in ${ITERS}; do \
|
|
timeout ${TIMEOUT} mpirun -np ${NUM_PES} -hosts ${HOSTS} ${BUILD}/sort_mpi >> ${RESULTS}/sort_mpi2_${NUM_PES}.out;\
|
|
done;
|
|
|
|
run_sort: ${BUILD}/sort_shmem ${BUILD}/sort_rccl
|
|
for i in ${PES}; do \
|
|
printf "%d " $$i; \
|
|
$(MAKE) --no-print-directory run_sort_shmem TYPE=NAIVE LABEL=naive PARAM=0 NUM_PES=$${i}; \
|
|
$(MAKE) --no-print-directory run_sort_shmem TYPE=MPI LABEL=mpi PARAM=1 NUM_PES=$${i}; \
|
|
$(MAKE) --no-print-directory run_sort_shmem TYPE=GCEN LABEL=gcen PARAM=2 NUM_PES=$${i}; \
|
|
$(MAKE) --no-print-directory run_sort_shmem TYPE=GCEN2 LABEL=gcen2 PARAM=3 NUM_PES=$${i}; \
|
|
$(MAKE) --no-print-directory run_sort_mpi NUM_PES=$${i}; \
|
|
printf "\n";\
|
|
done
|
|
for i in ${PES_RCCL}; do \
|
|
$(MAKE) --no-print-directory run_sort_rccl NUM_PES=$${i}; \
|
|
printf "%d " $$i; \
|
|
done
|
|
|
|
$(MAKE) extract_sort
|
|
|
|
|
|
extract_sort:
|
|
printf "Sort latency\n"
|
|
printf "PROCS\tType\tRuns"
|
|
for i in ${PES}; do \
|
|
for type in mpi mpi2 rccl naive gcen gcen2; do\
|
|
printf "\n%d\t$${type}\t" $$i; \
|
|
file=${RESULTS}/sort_$${type}_$${i}.out;\
|
|
latency=$$(grep -E "Avg time" $${file}); \
|
|
grep -E "Avg time" $${file} | while read -r j; do\
|
|
val=$$(echo $$j | grep -oE -m1 "[0-9]+\.[0-9]+");\
|
|
printf "%s\t" $${val};\
|
|
done; \
|
|
done;\
|
|
done
|
|
printf "\n"
|
|
|
|
clean:
|
|
rm build/*;
|