lib pthread：单线程性能与多线程

Question

（注意：我看到这篇文章，告诉我这是否是同样的问题：C: performance of pthread, low than single thrad）

我正在学习 pthread 库。我写了两个版本的同一个C程序。该程序采用大BAM files列表，并使用samtools library计算记录数。

以下是单线程程序：

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include "bam.h"

/** maximum number of threads */
static const int MAX_COUNT_THREADS=4;

struct Param
    {
    char* filename;
    };


static void printCount(const char* filename,unsigned long count)
    {
    fprintf(stdout,"%s\t%ld\n",filename,count);
    }

static void* scan_bam(void* ptr)
    {
    unsigned long count=0;
    struct Param* params=(struct Param*)ptr;

    bamFile in=bam_open(params->filename, "r") ;
    bam_header_t *header= NULL;
    bam1_t *b=bam_init1();

    time_t rawtime;
    time ( &rawtime );

    fprintf(stderr,"STARTING : %s %s",params->filename,ctime(&rawtime));
    if(in==0)
        {
        fprintf(stderr,"Cannot read %s.\n",params->filename);
        exit(EXIT_FAILURE);
        }
    header= bam_header_read(in);
    while((bam_read1(in, b)) > 0)
        {
        ++count;
        }
    bam_destroy1(b);
    bam_header_destroy(header);
           bam_close(in);


    printCount(params->filename,count);

    time ( &rawtime );
    fprintf(stderr,"end for %s %s",params->filename,ctime(&rawtime));
    free(params);
    return NULL;
    }


int main(int argc,char** argv)
    {
    int optind=1;

    while(optind<argc)
    {
    struct Param* params=(struct Param*)malloc(sizeof(struct Param));

    if(params==0)
        {
        fprintf(stderr,"Out of memory.\n");
        exit(EXIT_FAILURE);
        }
    params->filename=argv[optind++];
    scan_bam(params);
    }
    return EXIT_SUCCESS;
    }

和多线程程序。该程序最多可以使用5个线程，并使用条件锁来计算线程数，并告诉主程序在需要时启动新线程。

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <pthread.h>
#include <time.h>
#include "bam.h"

#define VERIFY_ZERO(a) do {if(a!=0) {\
    fprintf(stderr,"Test failed at %s line %d (ret=%d).\n",__FILE__,__LINE__,a);\
    exit(EXIT_FAILURE);\
    }} while(0)



/** maximum number of threads */
static const int MAX_COUNT_THREADS=5;

struct Param
    {
    pthread_t thread;
    char* filename;
    };

struct GLOBALS {
    /** lock to print */
    pthread_mutex_t mutex_print;
    /** condition: wait for free thread */
    pthread_cond_t  accept_new_thread;
    /** condition lock */
    pthread_mutex_t  accept_new_thread_lock;
    /** number of running threads */
    int number_of_threads;
    };

static struct GLOBALS globals={
    PTHREAD_MUTEX_INITIALIZER,
    PTHREAD_COND_INITIALIZER,
    PTHREAD_MUTEX_INITIALIZER,
    0
    };


static void printCount(const char* filename,unsigned long count)
    {
    int ret=pthread_mutex_lock(&globals.mutex_print);
    VERIFY_ZERO(ret);
    fprintf(stdout,"%s\t%ld\n",filename,count);
    ret=pthread_mutex_unlock(&globals.mutex_print);
    VERIFY_ZERO(ret);
    }

static void* scan_bam(void* ptr)
    {
    unsigned long count=0;
    struct Param* params=(struct Param*)ptr;
    bamFile in=bam_open(params->filename, "r") ;
    bam_header_t *header= NULL;
    bam1_t *b=bam_init1();
    time_t rawtime;
    time ( &rawtime );

    fprintf(stderr,"STARTING : %s %s",params->filename,ctime(&rawtime));
    if(in==0)
        {
        fprintf(stderr,"Cannot read %s.\n",params->filename);
        exit(EXIT_FAILURE);
        }
    header= bam_header_read(in);
    while((bam_read1(in, b)) > 0)
        {
        ++count;

        }
    bam_destroy1(b);
    bam_header_destroy(header);
           bam_close(in);


    printCount(params->filename,count);

    time ( &rawtime );
    fprintf(stderr,"end1 for %s %s",params->filename,ctime(&rawtime));
    pthread_mutex_lock( &globals.accept_new_thread_lock);
    globals.number_of_threads--;
    pthread_cond_signal(&globals.accept_new_thread);
    pthread_mutex_unlock(&globals.accept_new_thread_lock);
    time ( &rawtime );
    fprintf(stderr,"end2 for %s %s",params->filename,ctime(&rawtime));
    free(params);
    return NULL;
    }


int main(int argc,char** argv)
    {
    int optind=1;

    while(optind<argc)
    {
    struct Param* params=(struct Param*)malloc(sizeof(struct Param));
    if(params==0)
        {
        fprintf(stderr,"Out of memory.\n");
        exit(EXIT_FAILURE);
        }
    pthread_mutex_lock(&globals.accept_new_thread_lock);
    while (globals.number_of_threads > MAX_COUNT_THREADS)
        {
            pthread_cond_wait(&globals.accept_new_thread, &globals.accept_new_thread_lock);
        }
    globals.number_of_threads++;
    pthread_mutex_unlock(&globals.accept_new_thread_lock);

    params->filename=argv[optind++];
    fprintf(stderr,"creating %s\n",params->filename);
    pthread_create (&(params->thread), NULL, scan_bam,params);
    pthread_detach(params->thread);

    }
    pthread_mutex_lock(&globals.accept_new_thread_lock);
    while (globals.number_of_threads > 0)
    {
        pthread_cond_wait(&globals.accept_new_thread, &globals.accept_new_thread_lock);
    }
    pthread_mutex_unlock(&globals.accept_new_thread_lock);
    pthread_cond_destroy(&globals.accept_new_thread);
    return EXIT_SUCCESS;
    }

编译并运行多线程程序

gcc -O3 -Wall jeter.c -pthread -I/usr/local/package/samtools-0.1.18 -L/usr/local/package/samtools-0.1.18/  -lbam -lz
$ time (find .// -name "*recal.bam" | grep Item1[0-9] | xargs ./a.out )
creating ./Item10/recal.bam
creating ./Item11/recal.bam
creating ./Item12/recal.bam
creating ./Item13/recal.bam
creating ./Item14/recal.bam
creating ./Item15/recal.bam
STARTING : ./Item10/recal.bam Tue Dec 18 15:12:48 2012
STARTING : ./Item11/recal.bam Tue Dec 18 15:12:48 2012
STARTING : ./Item12/recal.bam Tue Dec 18 15:12:48 2012
STARTING : ./Item14/recal.bam Tue Dec 18 15:12:48 2012
STARTING : ./Item13/recal.bam Tue Dec 18 15:12:48 2012
STARTING : ./Item15/recal.bam Tue Dec 18 15:12:48 2012
./Item10/recal.bam    185784310
end1 for ./Item10/recal.bam Tue Dec 18 15:38:16 2012
end2 for ./Item10/recal.bam Tue Dec 18 15:38:16 2012
creating ./Item16/recal.bam
STARTING : ./Item16/recal.bam Tue Dec 18 15:38:16 2012
./Item11/recal.bam    204408906
end1 for ./Item11/recal.bam Tue Dec 18 15:41:52 2012
end2 for ./Item11/recal.bam Tue Dec 18 15:41:52 2012
creating ./Item17/recal.bam
STARTING : ./Item17/recal.bam Tue Dec 18 15:41:52 2012
./Item12/recal.bam    207766317
end1 for ./Item12/recal.bam Tue Dec 18 15:42:17 2012
end2 for ./Item12/recal.bam Tue Dec 18 15:42:17 2012
creating ./Item18/recal.bam
STARTING : ./Item18/recal.bam Tue Dec 18 15:42:17 2012
./Item15/recal.bam    224957522
end1 for ./Item15/recal.bam Tue Dec 18 15:44:54 2012
end2 for ./Item15/recal.bam Tue Dec 18 15:44:54 2012
creating ./Item19/recal.bam
STARTING : ./Item19/recal.bam Tue Dec 18 15:44:54 2012
./Item13/recal.bam    224548326
end1 for ./Item13/recal.bam Tue Dec 18 15:45:32 2012
end2 for ./Item13/recal.bam Tue Dec 18 15:45:32 2012
./Item14/recal.bam    241267346
end1 for ./Item14/recal.bam Tue Dec 18 15:48:28 2012
end2 for ./Item14/recal.bam Tue Dec 18 15:48:28 2012
./Item16/recal.bam    227446579
end1 for ./Item16/recal.bam Tue Dec 18 16:12:15 2012
end2 for ./Item16/recal.bam Tue Dec 18 16:12:15 2012
./Item17/recal.bam    215307379
end1 for ./Item17/recal.bam Tue Dec 18 16:13:05 2012
end2 for ./Item17/recal.bam Tue Dec 18 16:13:05 2012
./Item18/recal.bam    225914723
end1 for ./Item18/recal.bam Tue Dec 18 16:13:48 2012
end2 for ./Item18/recal.bam Tue Dec 18 16:13:48 2012
./Item19/recal.bam    225509630
end1 for ./Item19/recal.bam Tue Dec 18 16:14:06 2012
end2 for ./Item19/recal.bam Tue Dec 18 16:14:06 2012

。

real    61m17.560s
user    66m0.476s
sys    4m5.980s

编译并运行单线程程序

$ gcc -O3 -Wall jeter2.c -I/usr/local/package/samtools-0.1.18 -L/usr/local/package/samtools-0.1.18/  -lbam -lz

time (find .// -name "*recal.bam" | grep Item1[0-9] | xargs ./a.out )
STARTING : ./Item10/recal.bam Tue Dec 18 16:15:25 2012
./Item10/recal.bam    185784310
end for ./Item10/recal.bam Tue Dec 18 16:20:43 2012
STARTING : ./Item11/recal.bam Tue Dec 18 16:20:43 2012
./Item11/recal.bam    204408906
end for ./Item11/recal.bam Tue Dec 18 16:26:20 2012
STARTING : ./Item12/recal.bam Tue Dec 18 16:26:20 2012
./Item12/recal.bam    207766317
end for ./Item12/recal.bam Tue Dec 18 16:31:56 2012
STARTING : ./Item13/recal.bam Tue Dec 18 16:31:56 2012
./Item13/recal.bam    224548326
end for ./Item13/recal.bam Tue Dec 18 16:38:05 2012
STARTING : ./Item14/recal.bam Tue Dec 18 16:38:05 2012
./Item14/recal.bam    241267346
end for ./Item14/recal.bam Tue Dec 18 16:44:59 2012
STARTING : ./Item15/recal.bam Tue Dec 18 16:44:59 2012
./Item15/recal.bam    224957522
end for ./Item15/recal.bam Tue Dec 18 16:50:56 2012
STARTING : ./Item16/recal.bam Tue Dec 18 16:50:56 2012
./Item16/recal.bam    227446579
end for ./Item16/recal.bam Tue Dec 18 16:58:07 2012
STARTING : ./Item17/recal.bam Tue Dec 18 16:58:07 2012
./Item17/recal.bam    215307379
end for ./Item17/recal.bam Tue Dec 18 17:04:58 2012
STARTING : ./Item18/recal.bam Tue Dec 18 17:04:58 2012
./Item18/recal.bam    225914723
end for ./Item18/recal.bam Tue Dec 18 17:11:31 2012
STARTING : ./Item19/recal.bam Tue Dec 18 17:11:31 2012
./Item19/recal.bam    225509630
end for ./Item19/recal.bam Tue Dec 18 17:18:19 2012

real    62m54.503s
user    53m39.529s
sys    3m44.580s

这两个程序已运行〜1H00。所以MT程序的运行速度比另一个慢。为什么？是否可以加速该代码？

Answer 1

看起来I / O操作（从文件读取）在你的程序中占主导地位，所以你很可能不会从线程中获得太多好处，无论它做得多好。

另请注意，多线程变体实际上更快一点;你需要比较实时。由于多线程，用户时间更大，因为它总结了所有线程在用户模式下花费的时间。 内核时间相同。

Answer 2

首先，我想说清楚一点：

多线程并不一定意味着您的程序运行得更快！

这可能意味着，但取决于你的程序真正执行的并行程度。

我在你的MT代码中看到你正在等待条件变量，这意味着那里可能存在争用点，而且这里没有真正的并行性。

实际上，当线程可以执行一些工作时，实际上可以获得真正的加速，而不必在其他线程上等待。如果他们可以完全独立完成所有事情，那么可能可以获得一些加速。

然后加速也依赖于假共享这样的东西（假设是多核CPU），如果你在内存中有一些数组，而线程A需要该数组中的元素2，而另一个运行的线程B需要元素4在那个数组中，元素是sizeof（int），那么你就知道你会遇到缓存问题。每次线程执行对数组的写入时，都需要更新缓存。假设缓存行为64k。这是错误的分享。这可能导致非常显着的减速。

还有其他原因导致您的MT程序无法加速单线程程序。你应该问的最终问题是：

我可以让线程完成一项与任何其他线程没有依赖关系的工作吗？

从所有等待你的线程做起，似乎没有。