perf bench: Add sched-messaging.c: Benchmark for scheduler and IPC mechanisms based...
Hitoshi Mitake [Thu, 5 Nov 2009 00:31:32 +0000 (09:31 +0900)]
This patch adds bench/sched-messaging.c.

This benchmark measures performance of scheduler and IPC
mechanisms, and is based on hackbench by Rusty Russell.

Example of usage:

  % perf bench sched messaging -g 20 -l 1000 -s
  5.432                             # in sec

  % perf bench sched messaging                 # run with default
  options (20 sender and receiver processes per group)
  (10 groups == 400 processes run)

        Total time:0.308 sec

  % perf bench sched messaging -t -g 20      # # be multi-thread,
  with 20 groups (20 sender and receiver threads per group)
  (20 groups == 800 threads run)

        Total time:0.582 sec

( Rusty is the original author of hackbench.c and he said the code is
  and was under the GPLv2 so fine to be merged. )

Signed-off-by: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Acked-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: fweisbec@gmail.com
Cc: Jiri Kosina <jkosina@suse.cz>
LKML-Reference: <1257381097-4743-3-git-send-email-mitake@dcl.info.waseda.ac.jp>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

tools/perf/bench/sched-messaging.c [new file with mode: 0644]

diff --git a/tools/perf/bench/sched-messaging.c b/tools/perf/bench/sched-messaging.c
new file mode 100644 (file)
index 0000000..36b62c5
--- /dev/null
@@ -0,0 +1,332 @@
+/*
+ *
+ * builtin-bench-messaging.c
+ *
+ * messaging: Benchmark for scheduler and IPC mechanisms
+ *
+ * Based on hackbench by Rusty Russell <rusty@rustcorp.com.au>
+ * Ported to perf by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
+ *
+ */
+
+#include "../perf.h"
+#include "../util/util.h"
+#include "../util/parse-options.h"
+#include "../builtin.h"
+#include "bench.h"
+
+/* Test groups of 20 processes spraying to 20 receivers */
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/socket.h>
+#include <sys/wait.h>
+#include <sys/time.h>
+#include <sys/poll.h>
+#include <limits.h>
+
+#define DATASIZE 100
+
+static int use_pipes = 0;
+static unsigned int loops = 100;
+static unsigned int thread_mode = 0;
+static unsigned int num_groups = 10;
+static int simple = 0;
+
+struct sender_context {
+       unsigned int num_fds;
+       int ready_out;
+       int wakefd;
+       int out_fds[0];
+};
+
+struct receiver_context {
+       unsigned int num_packets;
+       int in_fds[2];
+       int ready_out;
+       int wakefd;
+};
+
+static void barf(const char *msg)
+{
+       fprintf(stderr, "%s (error: %s)\n", msg, strerror(errno));
+       exit(1);
+}
+
+static void fdpair(int fds[2])
+{
+       if (use_pipes) {
+               if (pipe(fds) == 0)
+                       return;
+       } else {
+               if (socketpair(AF_UNIX, SOCK_STREAM, 0, fds) == 0)
+                       return;
+       }
+
+       barf(use_pipes ? "pipe()" : "socketpair()");
+}
+
+/* Block until we're ready to go */
+static void ready(int ready_out, int wakefd)
+{
+       char dummy;
+       struct pollfd pollfd = { .fd = wakefd, .events = POLLIN };
+
+       /* Tell them we're ready. */
+       if (write(ready_out, &dummy, 1) != 1)
+               barf("CLIENT: ready write");
+
+       /* Wait for "GO" signal */
+       if (poll(&pollfd, 1, -1) != 1)
+               barf("poll");
+}
+
+/* Sender sprays loops messages down each file descriptor */
+static void *sender(struct sender_context *ctx)
+{
+       char data[DATASIZE];
+       unsigned int i, j;
+
+       ready(ctx->ready_out, ctx->wakefd);
+
+       /* Now pump to every receiver. */
+       for (i = 0; i < loops; i++) {
+               for (j = 0; j < ctx->num_fds; j++) {
+                       int ret, done = 0;
+
+again:
+                       ret = write(ctx->out_fds[j], data + done,
+                                   sizeof(data)-done);
+                       if (ret < 0)
+                               barf("SENDER: write");
+                       done += ret;
+                       if (done < DATASIZE)
+                               goto again;
+               }
+       }
+
+       return NULL;
+}
+
+
+/* One receiver per fd */
+static void *receiver(struct receiver_context* ctx)
+{
+       unsigned int i;
+
+       if (!thread_mode)
+               close(ctx->in_fds[1]);
+
+       /* Wait for start... */
+       ready(ctx->ready_out, ctx->wakefd);
+
+       /* Receive them all */
+       for (i = 0; i < ctx->num_packets; i++) {
+               char data[DATASIZE];
+               int ret, done = 0;
+
+again:
+               ret = read(ctx->in_fds[0], data + done, DATASIZE - done);
+               if (ret < 0)
+                       barf("SERVER: read");
+               done += ret;
+               if (done < DATASIZE)
+                       goto again;
+       }
+
+       return NULL;
+}
+
+static pthread_t create_worker(void *ctx, void *(*func)(void *))
+{
+       pthread_attr_t attr;
+       pthread_t childid;
+       int err;
+
+       if (!thread_mode) {
+               /* process mode */
+               /* Fork the receiver. */
+               switch (fork()) {
+               case -1:
+                       barf("fork()");
+                       break;
+               case 0:
+                       (*func) (ctx);
+                       exit(0);
+                       break;
+               default:
+                       break;
+               }
+
+               return (pthread_t)0;
+       }
+
+       if (pthread_attr_init(&attr) != 0)
+               barf("pthread_attr_init:");
+
+#ifndef __ia64__
+       if (pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN) != 0)
+               barf("pthread_attr_setstacksize");
+#endif
+
+       err = pthread_create(&childid, &attr, func, ctx);
+       if (err != 0) {
+               fprintf(stderr, "pthread_create failed: %s (%d)\n",
+                       strerror(err), err);
+               exit(-1);
+       }
+       return childid;
+}
+
+static void reap_worker(pthread_t id)
+{
+       int proc_status;
+       void *thread_status;
+
+       if (!thread_mode) {
+               /* process mode */
+               wait(&proc_status);
+               if (!WIFEXITED(proc_status))
+                       exit(1);
+       } else {
+               pthread_join(id, &thread_status);
+       }
+}
+
+/* One group of senders and receivers */
+static unsigned int group(pthread_t *pth,
+               unsigned int num_fds,
+               int ready_out,
+               int wakefd)
+{
+       unsigned int i;
+       struct sender_context *snd_ctx = malloc(sizeof(struct sender_context)
+                       + num_fds * sizeof(int));
+
+       if (!snd_ctx)
+               barf("malloc()");
+
+       for (i = 0; i < num_fds; i++) {
+               int fds[2];
+               struct receiver_context *ctx = malloc(sizeof(*ctx));
+
+               if (!ctx)
+                       barf("malloc()");
+
+
+               /* Create the pipe between client and server */
+               fdpair(fds);
+
+               ctx->num_packets = num_fds * loops;
+               ctx->in_fds[0] = fds[0];
+               ctx->in_fds[1] = fds[1];
+               ctx->ready_out = ready_out;
+               ctx->wakefd = wakefd;
+
+               pth[i] = create_worker(ctx, (void *)receiver);
+
+               snd_ctx->out_fds[i] = fds[1];
+               if (!thread_mode)
+                       close(fds[0]);
+       }
+
+       /* Now we have all the fds, fork the senders */
+       for (i = 0; i < num_fds; i++) {
+               snd_ctx->ready_out = ready_out;
+               snd_ctx->wakefd = wakefd;
+               snd_ctx->num_fds = num_fds;
+
+               pth[num_fds+i] = create_worker(snd_ctx, (void *)sender);
+       }
+
+       /* Close the fds we have left */
+       if (!thread_mode)
+               for (i = 0; i < num_fds; i++)
+                       close(snd_ctx->out_fds[i]);
+
+       /* Return number of children to reap */
+       return num_fds * 2;
+}
+
+static const struct option options[] = {
+       OPT_BOOLEAN('p', "pipe", &use_pipes,
+                   "Use pipe() instead of socketpair()"),
+       OPT_BOOLEAN('t', "thread", &thread_mode,
+                   "Be multi thread instead of multi process"),
+       OPT_INTEGER('g', "group", &num_groups,
+                   "Specify number of groups"),
+       OPT_INTEGER('l', "loop", &loops,
+                   "Specify number of loops"),
+       OPT_BOOLEAN('s', "simple-output", &simple,
+                   "Do simple output (this maybe useful for"
+                   "processing by scripts or graph tools like gnuplot)"),
+       OPT_END()
+};
+
+static const char * const bench_sched_message_usage[] = {
+       "perf bench sched messaging <options>",
+       NULL
+};
+
+int bench_sched_messaging(int argc, const char **argv,
+                   const char *prefix __used)
+{
+       unsigned int i, total_children;
+       struct timeval start, stop, diff;
+       unsigned int num_fds = 20;
+       int readyfds[2], wakefds[2];
+       char dummy;
+       pthread_t *pth_tab;
+
+       argc = parse_options(argc, argv, options,
+                            bench_sched_message_usage, 0);
+
+       pth_tab = malloc(num_fds * 2 * num_groups * sizeof(pthread_t));
+       if (!pth_tab)
+               barf("main:malloc()");
+
+       fdpair(readyfds);
+       fdpair(wakefds);
+
+       total_children = 0;
+       for (i = 0; i < num_groups; i++)
+               total_children += group(pth_tab+total_children, num_fds,
+                                       readyfds[1], wakefds[0]);
+
+       /* Wait for everyone to be ready */
+       for (i = 0; i < total_children; i++)
+               if (read(readyfds[0], &dummy, 1) != 1)
+                       barf("Reading for readyfds");
+
+       gettimeofday(&start, NULL);
+
+       /* Kick them off */
+       if (write(wakefds[1], &dummy, 1) != 1)
+               barf("Writing to start them");
+
+       /* Reap them all */
+       for (i = 0; i < total_children; i++)
+               reap_worker(pth_tab[i]);
+
+       gettimeofday(&stop, NULL);
+
+       timersub(&stop, &start, &diff);
+
+       if (simple)
+               printf("%lu.%03lu\n", diff.tv_sec, diff.tv_usec/1000);
+       else {
+               printf("(%d sender and receiver %s per group)\n",
+                      num_fds, thread_mode ? "threads" : "processes");
+               printf("(%d groups == %d %s run)\n\n",
+                      num_groups, num_groups * 2 * num_fds,
+                      thread_mode ? "threads" : "processes");
+               printf("\tTotal time:%lu.%03lu sec\n",
+                      diff.tv_sec, diff.tv_usec/1000);
+       }
+
+       return 0;
+}