test/mpi: add p2p benchmarks in test/mpi/bench
Add point-to-point benchmark code in MyDef. The tests have automatic
warm-ups and adjusts number of iterations for measurement accuracy.
It produces latency measurements with standard deviations and equivalent
bandwidths.
MYDEF_BOOT=[topsrc_dir]/modules/mydef_boot
export PATH=$MYDEF_BOOT/bin:$PATH
export PERL5LIB=$MYDEF_BOOT/lib/perl5
export MYDEFLIB=$MYDEF_BOOT/lib/MyDef
To run:
mydef_page p2p_latency.def # -> p2p_latency.c
mpicc p2p_latency.c && mpi_run -n 2 ./a.out
Alternatively use mydef_run (uses settings from config):
mydef_run p2p_latency.def
Next commit will add "make testing".
Esse commit está contido em:
Hui Zhou
@@ -0,0 +1,3 @@
|
||||
module: c
|
||||
CC: mpicc
|
||||
run: mpirun -n 2
|
||||
@@ -0,0 +1,93 @@
|
||||
/*
|
||||
* bench_frame : boilerplate for mpi program
|
||||
* measure(iter) : measures `tf_dur` for $(iter) iterations
|
||||
* run_stat(N, var) : run N measurements and obtain (avg, std) in sum1, sum2
|
||||
* warm_up(iter, dur): repeat until measurements (iter, dur) stabilize
|
||||
* report_latency(msgsize, MULTIPLICITY) : print a line of latency result
|
||||
*/
|
||||
|
||||
subcode: bench_frame
|
||||
$include stdio
|
||||
$include stdlib
|
||||
$include mpi
|
||||
|
||||
$function main
|
||||
MPI_Init(NULL, NULL);
|
||||
|
||||
$my grank, gsize: int
|
||||
MPI_Comm_rank(MPI_COMM_WORLD, &grank);
|
||||
MPI_Comm_size(MPI_COMM_WORLD, &gsize);
|
||||
$(if:MIN_PROCS)
|
||||
$if gsize < $(MIN_PROCS)
|
||||
printf("! Test $(_pagename) requires $(MIN_PROCS) processes !\n");
|
||||
return 1
|
||||
|
||||
MPI_Comm comm = MPI_COMM_WORLD;
|
||||
char *buf = malloc(MAX_BUFSIZE)
|
||||
$if !buf
|
||||
printf("! Failed to allocate buffer (size=%d)\n", MAX_BUFSIZE)
|
||||
return 1
|
||||
|
||||
$if grank == 0
|
||||
printf("TEST $(_pagename):\n")
|
||||
$call @report_header
|
||||
$call main
|
||||
$if grank == 0
|
||||
printf("\n")
|
||||
|
||||
MPI_Finalize();
|
||||
|
||||
macros:
|
||||
use_double: 1
|
||||
|
||||
#----------------------------------------
|
||||
subcode: _autoload
|
||||
$register_prefix(comm) MPI_Comm
|
||||
|
||||
subcode: foreach_size
|
||||
$for int size = 0; size < $(MAX_MSG); size = (size==0)?1:size*2
|
||||
$(set:MSG_SIZE=size)
|
||||
BLOCK
|
||||
|
||||
subcode: measure(iter)
|
||||
tf_start = MPI_Wtime()
|
||||
$for 0:$(iter)
|
||||
BLOCK
|
||||
tf_dur = MPI_Wtime() - tf_start
|
||||
|
||||
subcode: run_stat(N, var)
|
||||
$my double sum1=0, double sum2=0
|
||||
$for 0:$(N)
|
||||
BLOCK
|
||||
sum1 += $(var)
|
||||
sum2 += $(var) * $(var)
|
||||
sum1 /= $(N)
|
||||
sum2 /= $(N)
|
||||
sum2 = sqrt(sum2 - sum1 * sum1)
|
||||
|
||||
subcode: warm_up(iter, dur)
|
||||
$(set:MIN_ITER=(int) ($(iter) * 0.001 / $(dur)))
|
||||
$(iter) = 2
|
||||
$my double last_dur = 1.0
|
||||
$my int num_best = 0
|
||||
$while num_best < 10
|
||||
BLOCK
|
||||
$if $(iter) < $(MIN_ITER)
|
||||
$(iter) = $(MIN_ITER)
|
||||
num_best = 0
|
||||
continue
|
||||
# check that t_dur is no longer monotonically decreasing
|
||||
$if $(dur) > last_dur
|
||||
num_best++
|
||||
last_dur = $(dur)
|
||||
|
||||
subcode: header_latency
|
||||
printf("%12s %10s(us) %6s(us) %12s(MB/s)\n", "msgsize", "latency", "sigma", "bandwidth")
|
||||
|
||||
subcode: report_latency(MSGSIZE, MULTIPLICITY)
|
||||
$my tf_latency, tf_sigma, tf_bw
|
||||
tf_latency = sum1 / ($(MULTIPLICITY)) * 1e6
|
||||
tf_sigma = sum2 / ($(MULTIPLICITY)) * 1e6
|
||||
tf_bw = $(MSGSIZE) / tf_latency
|
||||
printf("%12d %10.3f %6.3f %12.3f\n", $(MSGSIZE), tf_latency, tf_sigma, tf_bw)
|
||||
|
||||
@@ -0,0 +1,79 @@
|
||||
/*
|
||||
* Defines following functions:
|
||||
* bench_p2p
|
||||
* bench_send, bench_warmup
|
||||
* bench_recv
|
||||
*
|
||||
* For each measurement -
|
||||
* First sender tells receiver the `iter` parameter. `iter = 0` means to quit.
|
||||
* For each iteration runs `send_side` and `recv_side` assuming the measurement on sender side represents a latency measurement.
|
||||
*
|
||||
* Caller page defines -
|
||||
* subcode: sender_side, recv_side
|
||||
* macro:
|
||||
* MULTIPLICITY: divisor for each measurement
|
||||
*/
|
||||
|
||||
macros:
|
||||
MIN_PROCS: 2
|
||||
MAX_BUFSIZE: 5000000 # 5 MB
|
||||
|
||||
subcode: _autoload
|
||||
$register_name(src) int
|
||||
$register_name(dst) int
|
||||
$register_name(buf) void *
|
||||
$register_name(size) int
|
||||
$define TAG 0
|
||||
$define SYNC_TAG 100
|
||||
$define MAX_BUFSIZE 5000000
|
||||
$define NUM_REPEAT 20
|
||||
|
||||
subcode: report_header
|
||||
$call header_latency
|
||||
|
||||
fncode: bench_p2p(comm, src, dst, buf, size)
|
||||
int rank;
|
||||
MPI_Comm_rank(comm, &rank)
|
||||
|
||||
$(if:!MULTIPLICITY)
|
||||
$(set:MULTIPLICITY=1)
|
||||
|
||||
$if rank == src
|
||||
iter = bench_warmup(comm, dst, buf, size)
|
||||
&call run_stat, NUM_REPEAT, tf_latency
|
||||
tf_latency = bench_send(iter, comm, dst, buf, size)
|
||||
tf_latency /= iter
|
||||
$call report_latency, size, $(MULTIPLICITY)
|
||||
$call send_stop
|
||||
$elif rank == dst
|
||||
bench_recv(comm, src, buf, size)
|
||||
|
||||
subcode: send_stop
|
||||
iter = 0;
|
||||
MPI_Send(&iter, 1, MPI_INT, dst, SYNC_TAG, comm)
|
||||
|
||||
#----------------------------------------
|
||||
fncode: bench_send(int iter, comm, dst, buf, size)
|
||||
# synchronize with receiver
|
||||
MPI_Send(&iter, 1, MPI_INT, dst, SYNC_TAG, comm);
|
||||
|
||||
&call measure, iter
|
||||
$call @send_side
|
||||
|
||||
return tf_dur
|
||||
|
||||
fncode: bench_recv(comm, src, buf, size)
|
||||
$while 1
|
||||
int iter;
|
||||
# synchronize with sender */
|
||||
MPI_Recv(&iter, 1, MPI_INT, src, SYNC_TAG, comm, MPI_STATUS_IGNORE);
|
||||
$if iter == 0
|
||||
# time to quit
|
||||
break
|
||||
$for i=0:iter
|
||||
$call @recv_side
|
||||
|
||||
fncode: bench_warmup(comm, dst, buf, size): int
|
||||
&call warm_up, iter, tf_dur
|
||||
tf_dur = bench_send(iter, comm, dst, buf, size)
|
||||
return iter
|
||||
@@ -0,0 +1,26 @@
|
||||
include: macros/bench_frame.def
|
||||
include: macros/bench_p2p.def
|
||||
|
||||
subcode: _autoload
|
||||
$define WINDOW_SIZE 64
|
||||
|
||||
page: p2p_bw, bench_frame
|
||||
MULTIPLICITY: WINDOW_SIZE
|
||||
data: buf, size, MPI_CHAR
|
||||
|
||||
$for int size = 1; size < MAX_BUFSIZE; size *= 2
|
||||
bench_p2p(comm, 0, 1, buf, size)
|
||||
|
||||
subcode: send_side
|
||||
$my MPI_Request reqs[WINDOW_SIZE]
|
||||
$for j=0:WINDOW_SIZE
|
||||
MPI_Isend($(data), dst, TAG, comm, &reqs[j])
|
||||
MPI_Waitall(WINDOW_SIZE, reqs, MPI_STATUSES_IGNORE)
|
||||
MPI_Recv(NULL, 0, MPI_DATATYPE_NULL, dst, TAG, comm, MPI_STATUS_IGNORE)
|
||||
|
||||
subcode: recv_side
|
||||
$my MPI_Request reqs[WINDOW_SIZE]
|
||||
$for j=0:WINDOW_SIZE
|
||||
MPI_Irecv($(data), src, TAG, comm, &reqs[j])
|
||||
MPI_Waitall(WINDOW_SIZE, reqs, MPI_STATUSES_IGNORE)
|
||||
MPI_Send(NULL, 0, MPI_DATATYPE_NULL, src, TAG, comm)
|
||||
@@ -0,0 +1,18 @@
|
||||
include: macros/bench_frame.def
|
||||
include: macros/bench_p2p.def
|
||||
|
||||
page: p2p_latency, bench_frame
|
||||
MULTIPLICITY: 2
|
||||
data: buf, size, MPI_CHAR
|
||||
|
||||
bench_p2p(comm, 0, 1, buf, 0)
|
||||
$for int size = 1; size < MAX_BUFSIZE; size *= 2
|
||||
bench_p2p(comm, 0, 1, buf, size)
|
||||
|
||||
subcode: send_side
|
||||
MPI_Send($(data), dst, TAG, comm);
|
||||
MPI_Recv($(data), dst, TAG, comm, MPI_STATUS_IGNORE);
|
||||
|
||||
subcode: recv_side
|
||||
MPI_Recv($(data), src, TAG, comm, MPI_STATUS_IGNORE);
|
||||
MPI_Send($(data), src, TAG, comm);
|
||||
Referência em uma Nova Issue
Bloquear um usuário