coNCePTuaL # coNCePTuaL version of the OSU MPI latency test
# Code written by Scott Pakin <pakin@lanl.gov>
# Ensure that future changes to the coNCePTuaL language don't break this
# program.
Require language version "0.5.3b".
# For the user's convenience we convert the C version's hardwired
# constants to command-line arguments.
message_alignment is "Message buffer alignment (bytes)" and comes from
"--align" or "-a" with default 64.
max_msg_size
is "Maximum message size (bytes)" and comes from "--max-size"
or "-m" with default 4M.
skip
is "Number of small-message warmup iterations" and comes from "--skip" or
"-k" with default 1000.
loop is "Number of small-message iterations" and comes from "--loop" or "-l" with
default 10000.
skip_large is "Number of large-message warmup iterations" and comes from
"--skip-large" or "-j" with default 10.
loop_large is "Number of large-message iterations" and comes from "--loop-large" or
"-i" with default 100.
large_message_size is "Large message size (bytes)" and comes from "--large-size"
or "-s" with default 8K.
# Heres an extra bonus, courtesy of coNCePTuaL.
Assert that "The latency test requires at least two nodes" with num_tasks>=2.
Let s_buf be 0 and r_buf be 1 while {
for each size in {0}, {1, 2, 4, ..., max_msg_size} {
all tasks touch all message buffers then
let loop be loop_large if size > large_message_size otherwise loop
and skip be skip_large if size > large_message_size otherwise skip while {
all tasks synchronize then
task 0 resets its counters then
for loop repetitions plus skip warmup repetitions {
task 0 sends a size-byte message_alignment-byte-aligned message
from buffer s_buf to task 1 who receives it into buffer r_buf then
task 1 sends a size-byte message_alignment-byte-aligned message
from buffer s_buf to task 0 who receives it into buffer r_buf
} then
task 0 logs size as "Size" and elapsed_usecs/total_msgs as "Latency (us)"
}
}
}
/*
* Copyright (C) 2002-2004 the Network-Based Computing Laboratory
* (NBCL), The Ohio State University.
*/
#include "mpi.h"
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <unistd.h>
#define MESSAGE_ALIGNMENT 64
#define MAX_MSG_SIZE (1<<22)
#define MYBUFSIZE (MAX_MSG_SIZE + MESSAGE_ALIGNMENT)
char s_buf_original[MYBUFSIZE];
char r_buf_original[MYBUFSIZE];
int skip = 1000;
int loop = 10000;
int skip_large = 10;
int loop_large = 100;
int large_message_size = 8192;
int main(int argc, char *argv[])
{
int myid, numprocs, i;
int size;
MPI_Status reqstat;
char *s_buf, *r_buf;
int align_size;
double t_start = 0.0, t_end = 0.0;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &myid);
align_size = MESSAGE_ALIGNMENT;
s_buf =
(char *) (((unsigned long) s_buf_original + (align_size - 1)) /
align_size * align_size);
r_buf =
(char *) (((unsigned long) r_buf_original + (align_size - 1)) /
align_size * align_size);
if (myid == 0) {
fprintf(stdout, "# OSU MPI Latency Test (Version 2.0)\n");
fprintf(stdout, "# Size\t\tLatency (us) \n");
}
for (size = 0; size <= MAX_MSG_SIZE;
size = (size ? size * 2 : size + 1)) {
/* touch the data */
for (i = 0; i < size; i++) {
s_buf[i] = 'a';
r_buf[i] = 'b';
}
if (size > large_message_size) {
loop = loop_large;
skip = skip_large;
}
MPI_Barrier(MPI_COMM_WORLD);
if (myid == 0) {
for (i = 0; i < loop + skip; i++) {
if (i == skip)
t_start = MPI_Wtime();
MPI_Send(s_buf, size, MPI_CHAR, 1, 1, MPI_COMM_WORLD);
MPI_Recv(r_buf, size, MPI_CHAR, 1, 1, MPI_COMM_WORLD,
&reqstat);
}
t_end = MPI_Wtime();
} else if (myid == 1) {
for (i = 0; i < loop + skip; i++) {
MPI_Recv(r_buf, size, MPI_CHAR, 0, 1, MPI_COMM_WORLD,
&reqstat);
MPI_Send(s_buf, size, MPI_CHAR, 0, 1, MPI_COMM_WORLD);
}
}
if (myid == 0) {
double latency;
latency = (t_end - t_start) * 1.0e6 / (2.0 * loop);
fprintf(stdout, "%d\t\t%0.2f\n", size, latency);
}
}
MPI_Finalize();
return 0;
}
CONCEPTUAL (24 lines)
C (73 lines)
Porting a C program to CONCEPTUAL
Advantages of the CONCEPTUAL version
Shorter
More readable
Portable to other messaging layers without source-code
changes
Logs not only performance data but also a wealth of
information about the experimental setuptimer
type & quality, compilers used, environment variables,
complete source code, and more
Replaces hardwired constants with command-line
options (and even supports
--help
)
Fails gracefully when run with too few processors