sysbench的框架实现介绍

146 阅读4分钟

sysbench是一个非常经典的综合性能测试工具,它支持CPU,IO,内存,尤其是数据库的性能测试。那它是怎么做到通用性的呢,总结一句话是大量运用了重载的方法。

sysbench总体架构

sysbench是一个总体框架,它用来操作各个测性能的计算,那各个部门只需要做的一件事情是声明需要的实现。只要理解了这三个struct就可以了:

/* 某个测试用例的整体结构 */
typedef struct sb_test
{
  const char        *sname;
  const char        *lname;
  /* 下面有具体说明 */
  sb_operations_t   ops;
  sb_builtin_cmds_t builtin_cmds;
  sb_arg_t          *args;

  sb_list_item_t    listitem;
} sb_test_t;
/* 某个测试用例的具体操作实现结构 */
typedef struct
{
  sb_op_init            *init;            /* initialization function */
  sb_op_prepare         *prepare;         /* called after timers start,  but
                                             before thread execution */
  sb_op_thread_init     *thread_init;     /* thread initialization
                                             (called when each thread starts) */
  sb_op_print_mode      *print_mode;      /* print mode function */
  sb_op_next_event      *next_event;      /* event generation function */
  sb_op_execute_event   *execute_event;   /* event execution function */
  sb_op_report          *report_intermediate; /* intermediate reports handler */
  sb_op_report          *report_cumulative;   /* cumulative reports handler */
  sb_op_thread_run      *thread_run;      /* main thread loop */
  sb_op_thread_done     *thread_done;     /* thread finalize function */
  sb_op_cleanup         *cleanup;         /* called after exit from thread,
                                             but before timers stop */ 
  sb_op_done            *done;            /* finalize function */
} sb_operations_t;
/* 某个测试用例的三阶段实现结构 */
typedef struct
{
  sb_builtin_cmd_func_t *help;  /* print help */
  sb_builtin_cmd_func_t *prepare; /* prepare for the test */
  sb_builtin_cmd_func_t *run;   /* run the test */
  sb_builtin_cmd_func_t *cleanup; /* cleanup the test database, files, etc. */
} sb_builtin_cmds_t;

拿最简单的CPU性能计算举例,它需要实现的是:

static sb_test_t cpu_test =
{
  .sname = "cpu", /*case简称*/
  .lname = "CPU performance test",/*case全称*/
  .ops = {
    .init = cpu_init, /* 初始化case */
    .print_mode = cpu_print_mode, /* case启动前,做说明 */
    .next_event = cpu_next_event, /* 拿到下一个event的数据 */
    .execute_event = cpu_execute_event, /* 具体执行这个event */
    .report_cumulative = cpu_report_cumulative,  /* 阶段性报告输出 */
    .done = cpu_done /* case结束后,处理干净 */
  },
  .args = cpu_args /*子case需要的参数说明*/
};

看到这个后,把一个case需要做的事情描述很清楚了,从需要什么参数,到初始化,逐个event执行,函数定义很清晰。sysbench的其他case也都这样需要一个完整的结构说明,如io操作,则需要多一个case的prepare和cleandown声明。
那sysbench的完整流程是怎样呢?黄色部分是测试用例需要实现的。

至此,可以清晰地看到sysbench的框架还是非常好理解。
上面struct里面有个event概念,不同的测试event的定义都不一样:比如CPU的测试case,一个event是完成求得小于某数(默认10000)的所有质数。比如fileio的测试case,一次read或者一次write操作就是一个event。

sysbench的线程介绍

  1. worker_thread具体实现是怎样呢:欣赏下sysbench.c里面某子线程是如何执行的,代码非常清晰易懂:

    static int thread_run(sb_test_t *test, int thread_id) { sb_event_t event; int rc = 0;

    while (sb_more_events(thread_id) && rc == 0) { event = test->ops.next_event(thread_id); if (event.type == SB_REQ_TYPE_NULL) break;

    sb_event_start(thread_id);
    
    rc = test->ops.execute_event(&event, thread_id);
    
    sb_event_stop(thread_id);
    

    }

    return rc; }

  2. intermediate_report线程:周期性输出性能数据,参数项为:--report-interval=N,对CPU的测试用例举例:sysbench cpu --report-interval=1,截取部分输出结果如下:

    Threads started!

    [ 1s ] thds: 1 eps: 922.10 lat (ms,95%): 1.08 [ 2s ] thds: 1 eps: 925.19 lat (ms,95%): 1.08 [ 3s ] thds: 1 eps: 926.00 lat (ms,95%): 1.08 [ 4s ] thds: 1 eps: 926.00 lat (ms,95%): 1.08 [ 5s ] thds: 1 eps: 926.00 lat (ms,95%): 1.08 [ 6s ] thds: 1 eps: 926.00 lat (ms,95%): 1.08 [ 7s ] thds: 1 eps: 925.00 lat (ms,95%): 1.08 [ 8s ] thds: 1 eps: 926.02 lat (ms,95%): 1.08 [ 9s ] thds: 1 eps: 925.99 lat (ms,95%): 1.08 [ 10s ] thds: 1 eps: 924.98 lat (ms,95%): 1.08

每一秒输出一个结果,eps是每一秒的event数,lat单位是毫秒,95分位延迟数据是1.08

  1. checkpoints_report线程:如果嫌周期性输出不够多,那么可以在某几个时间点整体输出,参数项为: --report-checkpoints=[LIST,...]
    还是对CPU测试用例举例:sysbench cpu --report-checkpoints=3,8 run,截取部分输出结果如下:

    Threads started!

    [ 3s ] Checkpoint report: CPU speed: events per second: 923.01

    General statistics: total time: 3.0001s total number of events: 2771

    Latency (ms): min: 1.08 avg: 1.08 max: 1.22 95th percentile: 1.08 sum: 3000.88

    Threads fairness: events (avg/stddev): 2773.0000/0.00 execution time (avg/stddev): 3.0009/0.00

    [ 8s ] Checkpoint report: CPU speed: events per second: 924.47

    General statistics: total time: 8.0001s total number of events: 4622

    Latency (ms): min: 1.08 avg: 1.08 max: 1.16 95th percentile: 1.08 sum: 4998.04

    Threads fairness: events (avg/stddev): 4621.0000/0.00 execution time (avg/stddev): 4.9980/0.00

  2. tx_rate_controll线程,控制每秒输出量的一个线程:参数项为: --rate=N,默认是不做控制的。
    还是拿CPU测试用例举例,控制每秒跑10个event:sysbench cpu run --rate=10,截取部分输出结果如下:

    Running the test with following options: Number of threads: 1 Target transaction rate: 10/sec Initializing random number generator from current time

    Prime numbers limit: 10000

    Initializing worker threads...

    Threads started!

    CPU speed: events per second: 8.87 #没那么精准哈

输出速率控制在哪里呢?眼尖的人马上可以看到是在sb_more_events函数。那sb_more_events函数主要是做什么呢:

  1. 判断是否超时,默认是10秒
  2. 判断是否到达最大event数,如果设置了的话
  3. 就是速率控制。

综上,大概介绍了sysbench框架的总体实现,关于数据库性能测试容下篇再介绍。