http://linux.derkeiler.com/Mailing-Lists/Kernel/2008-01/msg01289.html

Re: Improve hackbench



Ingo Molnar wrote:
* Zhang, Yanmin <[EMAIL PROTECTED]> wrote:


hackbench is to test Linux scheduler. The original program is at
http://devresources.linux-foundation.org/craiger/hackbench/src/hackbench.c
Based on this multi-process version, a nice person created a
multi-thread version. Pls. see
http://www.bullopensource.org/posix/pi-futex/hackbench_pth.c


great. I've uploaded your unified & improved version to:

http://redhat.com/~mingo/cfs-scheduler/tools/hackbench.c

(i made some small changes - two warning fixes on gcc 4.2 and a default
of 10 groups when hackbench is called without parameters, plus a
printout.)


On x86-64 there's a bug [*], that causes hackbench to segfault when
compiled with optimizations:
in reap_worker():
int status;
...
pthread_join(id, (void **)(void *)&status);

That is not correct, sizeof(void*) > sizeof(int) on x86-64.
Something gets overwritten on the stack, I tried with gcc
-fstack-protector, but it doesn't detect it !?
After applying the patch, it no longer segfaults.

This patch fixes it:
--- hackbench.c 2008-01-04 10:08:26.000000000 +0200
+++ ../hackbench.c 2008-01-04 13:45:22.000000000 +0200
@@ -241,8 +241,10 @@
wait(&status);
if (!WIFEXITED(status))
exit(1);
- } else
- pthread_join(id, (void **)(void *)&status);
+ } else {
+ void* status;
+ pthread_join(id, (void **)&status);
+ }
}

/* One group of senders and receivers */

----------------


I also notice that the thread version is slower, than process version:

$ ./hackbench 5 thread
Running with 5*40 (== 200) tasks.
Time: 0.413
$ ./hackbench 5 thread
Running with 5*40 (== 200) tasks.
Time: 0.423
$ ./hackbench 5 thread 20
Running with 5*40 (== 200) tasks.
Time: 0.093
$ ./hackbench 5 thread 200
Running with 5*40 (== 200) tasks.
Time: 0.827
$ ./hackbench 5 thread 2000
Running with 5*40 (== 200) tasks.
Time: 8.409
$ ./hackbench 5 process 2000
Running with 5*40 (== 200) tasks.
Time: 7.669
$ ./hackbench -pipe 5 process 2000
Running with 5*40 (== 200) tasks.
Time: 3.416
$ ./hackbench -pipe 5 thread 2000
Running with 5*40 (== 200) tasks.
Time: 4.320

[*]
$ uname -a
Linux lightspeed2 2.6.24-rc6-ge697789d #3 Wed Jan 2 11:15:05 EET 2008
x86_64 GNU/Linux
$ gcc -v
Using built-in specs.
Target: x86_64-linux-gnu
Configured with: ../src/configure -v
--enable-languages=c,c++,fortran,objc,obj-c++,treelang --prefix=/usr
--enable-shared --with-system-zlib --libexecdir=/usr/lib
--without-included-gettext --enable-threads=posix --enable-nls
--with-gxx-include-dir=/usr/include/c++/4.2 --program-suffix=-4.2
--enable-clocale=gnu --enable-libstdcxx-debug --enable-mpfr
--enable-checking=release --build=x86_64-linux-gnu
--host=x86_64-linux-gnu --target=x86_64-linux-gnu
Thread model: posix
gcc version 4.2.3 20071123 (prerelease) (Debian 4.2.2-4)
$ wget http://redhat.com/~mingo/cfs-scheduler/tools/hackbench.c
--13:40:53-- http://redhat.com/~mingo/cfs-scheduler/tools/hackbench.c
=> `hackbench.c'
Resolving redhat.com... 209.132.177.50
Connecting to redhat.com|209.132.177.50|:80... connected.
HTTP request sent, awaiting response... 302 Found
Location: http://www.redhat.com/~mingo/cfs-scheduler/tools/hackbench.c
[following]
--13:40:54-- http://www.redhat.com/~mingo/cfs-scheduler/tools/hackbench.c
=> `hackbench.c'
Resolving www.redhat.com... 209.132.177.50
Connecting to www.redhat.com|209.132.177.50|:80... connected.
HTTP request sent, awaiting response... 301 Moved Permanently
Location: http://people.redhat.com/mingo/cfs-scheduler/tools/hackbench.c
[following]
--13:40:54-- http://people.redhat.com/mingo/cfs-scheduler/tools/hackbench.c
=> `hackbench.c'
Resolving people.redhat.com... 66.187.233.237
Connecting to people.redhat.com|66.187.233.237|:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 8,455 (8.3K) [text/plain]

100%[====================================================================================================================>]
8,455 --.--K/s

13:40:55 (61.93 KB/s) - `hackbench.c' saved [8455/8455]

$ gcc -O2 -g -Wall -o hackbench hackbench.c -lpthread
hackbench.c:32:66: warning: missing terminating ' character
$ ./hackbench 1 thread
Running with 1*40 (== 40) tasks.
Segmentation fault

$ valgrind --trace-children=yes ./hackbench 1 thread
==27332== Memcheck, a memory error detector.
==27332== Copyright (C) 2002-2007, and GNU GPL'd, by Julian Seward et al.
==27332== Using LibVEX rev 1804, a library for dynamic binary translation.
==27332== Copyright (C) 2004-2007, and GNU GPL'd, by OpenWorks LLP.
==27332== Using valgrind-3.3.0-Debian, a dynamic binary instrumentation
framework.
==27332== Copyright (C) 2000-2007, and GNU GPL'd, by Julian Seward et al.
==27332== For more details, rerun with: -v
==27332==
Running with 1*40 (== 40) tasks.
==27332== Thread 2:
==27332== Syscall param write(buf) points to uninitialised byte(s)
==27332== at 0x4C1854B: (within /usr/lib/debug/libpthread-2.7.so)
==27332== by 0x400C34: ready (hackbench.c:138)
==27332== by 0x400C97: receiver (hackbench.c:182)
==27332== by 0x4C113F6: start_thread (pthread_create.c:297)
==27332== by 0x4EFD91C: clone (in /usr/lib/debug/libc-2.7.so)
==27332== Address 0x558a09f is on thread 2's stack
==27332==
==27332== Thread 22:
==27332== Syscall param write(buf) points to uninitialised byte(s)
==27332== at 0x4C1854B: (within /usr/lib/debug/libpthread-2.7.so)
==27332== by 0x400C34: ready (hackbench.c:138)
==27332== by 0x400D33: sender (hackbench.c:152)
==27332== by 0x4C113F6: start_thread (pthread_create.c:297)
==27332== by 0x4EFD91C: clone (in /usr/lib/debug/libc-2.7.so)
==27332== Address 0x55da07f is on thread 22's stack
==27332==
==27332== Thread 40:
==27332== Syscall param write(buf) points to uninitialised byte(s)
==27332== at 0x4C1854B: (within /usr/lib/debug/libpthread-2.7.so)
==27332== by 0x400D6E: sender (hackbench.c:160)
==27332== by 0x4C113F6: start_thread (pthread_create.c:297)
==27332== by 0x4EFD91C: clone (in /usr/lib/debug/libc-2.7.so)
==27332== Address 0x56220a0 is on thread 40's stack
==27332==
==27332== Thread 1:
==27332== Jump to the invalid address stated on the next line
==27332== at 0x0: ???
==27332== by 0x518702F: ???
==27332== by 0xFFFFFFFF: ???
==27332== by 0x518702F: ???
==27332== by 0x2800000000: ???
==27332== Address 0x0 is not stack'd, malloc'd or (recently) free'd
==27332==
==27332== Process terminating with default action of signal 11 (SIGSEGV)
==27332== Bad permissions for mapped region at address 0x0
==27332== at 0x0: ???
==27332== by 0x518702F: ???
==27332== by 0xFFFFFFFF: ???
==27332== by 0x518702F: ???
==27332== by 0x2800000000: ???
==27332==
==27332== ERROR SUMMARY: 40041 errors from 4 contexts (suppressed: 8 from 1)
==27332== malloc/free: in use at exit: 11,420 bytes in 61 blocks.
==27332== malloc/free: 62 allocs, 1 frees, 11,692 bytes allocated.
==27332== For counts of detected errors, rerun with: -v
==27332== searching for pointers to 61 not-freed blocks.
==27332== checked 560,688 bytes.
==27332==
==27332== LEAK SUMMARY:
==27332== definitely lost: 20 bytes in 1 blocks.
==27332== possibly lost: 10,608 bytes in 39 blocks.
==27332== still reachable: 792 bytes in 21 blocks.
==27332== suppressed: 0 bytes in 0 blocks.
==27332== Rerun with --leak-check=full to see details of leaked memory.
Segmentation fault



Best regards,

 

/*
 * This is the latest version of hackbench.c, that tests scheduler and
 * unix-socket (or pipe) performance.
 *
 * Usage: hackbench [-pipe] <num groups> [process|thread] [loops]
 *
 * Build it with:
 *   gcc -g -Wall -O2 -o hackbench hackbench.c -lpthread
 */
#if 0

Date: Fri, 04 Jan 2008 14:06:26 +0800
From: "Zhang, Yanmin" <[EMAIL PROTECTED]>
To: LKML <[EMAIL PROTECTED]>
Subject: Improve hackbench
Cc: Ingo Molnar <[EMAIL PROTECTED]>, Arjan van de Ven <[EMAIL PROTECTED]>

hackbench tests the Linux scheduler. The original program is at 
http://devresources.linux-foundation.org/craiger/hackbench/src/hackbench.c
Based on this multi-process version, a nice person created a multi-thread
version. Pls. see
http://www.bullopensource.org/posix/pi-futex/hackbench_pth.c

When I integrated them into my automation testing system, I found
a couple of issues and did some improvements.

1) Merge hackbench: I integrated hackbench_pth.c into hackbench and added a
new parameter which can be used to choose process mode or thread mode. The
default mode is process.

2) It runs too fast and ends in a couple of seconds. Sometimes it's too hard to debug
the issues. On my ia64 Montecito machines, the result looks weird when comparing
process mode and thread mode.
I want a stable result and hope the testing could run for a stable longer time, so I
might use performance tools to debug issues.
I added another new parameter,`loops`, which can be used to change variable loops,
so more messages will be passed from writers to receivers. Parameter 'loops' is equal to
100 by default.

For example on my 8-core x86_64:
[EMAIL PROTECTED] hackbench]$ uname -a
Linux lkp-st01-x8664 2.6.24-rc6 #1 SMP Fri Dec 21 08:32:31 CST 2007 x86_64 x86_64 x86_64 GNU/Linux
[EMAIL PROTECTED] hackbench]$ ./hackbench 
Usage: hackbench [-pipe] <num groups> [process|thread] [loops]
[EMAIL PROTECTED] hackbench]$ ./hackbench 150 process 1000
Time: 151.533
[EMAIL PROTECTED] hackbench]$ ./hackbench 150 thread 1000
Time: 153.666


With the same new parameters, I did captured the SLUB issue discussed on LKML recently.

3) hackbench_pth.c will fail on ia64 machine because pthread_attr_setstacksize always
fails if the stack size is less than 196*1024. I moved this statement within a __ia64__ check.


This new program could be compiled with command line:
#gcc -g -Wall  -o hackbench hackbench.c -lpthread


Thank Ingo for his great comments!

-yanmin

---

* Nathan Lynch <[EMAIL PROTECTED]> wrote:

> Here's a fixlet for the hackbench program found at
>
> http://people.redhat.com/mingo/cfs-scheduler/tools/hackbench.c
>
> When redirecting hackbench output I am seeing multiple copies of the
> "Running with %d*40 (== %d) tasks" line.  Need to flush the buffered
> output before forking.

#endif

/* 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 unsigned int loops = 100;
/*
 * 0 means thread mode and others mean process (default)
 */
static unsigned int process_mode = 1;

static int use_pipes = 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 print_usage_exit()
{
	printf("Usage: hackbench [-pipe] <num groups> [process|thread] [loops]\n");
	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("Creating fdpair");
}

/* 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 < sizeof(data))
				goto again;
		}
	}

	return NULL;
}


/* One receiver per fd */
static void *receiver(struct receiver_context* ctx)
{
	unsigned int i;

	if (process_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;
}

pthread_t create_worker(void *ctx, void *(*func)(void *))
{
	pthread_attr_t attr;
	pthread_t childid;
	int err;

	if (process_mode) {
		/* process mode */
		/* Fork the receiver. */
		switch (fork()) {
			case -1: barf("fork()");
			case 0:
				(*func) (ctx);
				exit(0);
		}

		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

	if ((err=pthread_create(&childid, &attr, func, ctx)) != 0) {
		fprintf(stderr, "pthread_create failed: %s (%d)\n", strerror(err), err);
		exit(-1);
	}
	return (childid);
}

void reap_worker(pthread_t id)
{
	int status;

	if (process_mode) {
		/* process mode */
		wait(&status);
		if (!WIFEXITED(status))
			exit(1);
	} else {
		void *status;

		pthread_join(id, &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));

	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 *)(void *)receiver);

		snd_ctx->out_fds[i] = fds[1];
		if (process_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 *)(void *)sender);
	}

	/* Close the fds we have left */
	if (process_mode)
		for (i = 0; i < num_fds; i++)
			close(snd_ctx->out_fds[i]);

	/* Return number of children to reap */
	return num_fds * 2;
}

int main(int argc, char *argv[])
{
	unsigned int i, num_groups = 10, total_children;
	struct timeval start, stop, diff;
	unsigned int num_fds = 20;
	int readyfds[2], wakefds[2];
	char dummy;
	pthread_t *pth_tab;

	if (argv[1] && strcmp(argv[1], "-pipe") == 0) {
		use_pipes = 1;
		argc--;
		argv++;
	}

	if (argc >= 2 && (num_groups = atoi(argv[1])) == 0)
		print_usage_exit();

	printf("Running with %d*40 (== %d) tasks.\n",
		num_groups, num_groups*40);

	fflush(NULL);

	if (argc > 2) {
		if ( !strcmp(argv[2], "process") )
			process_mode = 1;
		else if ( !strcmp(argv[2], "thread") )
			process_mode = 0;
		else
			print_usage_exit();
	}

	if (argc > 3)
		loops = atoi(argv[3]);

	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);

	/* Print time... */
	timersub(&stop, &start, &diff);
	printf("Time: %lu.%03lu\n", diff.tv_sec, diff.tv_usec/1000);
	exit(0);
}



Reply via email to