为什么Greenplum 的CPU有大量是%ni的占用

在使用Greenplum的过程中，发现CPU监控有大量的%ni的占比。
ni是指低优先级的用户模式，通过setpriority可以设置进程的优先级。数字越大，优先级越低。
TOP中CPU统计的分类解释如下：

   2c. SUMMARY Area Fields
       The summary area fields describing CPU statistics are abbreviated.  They provide information about times spent in:
           us = user mode
           sy = system mode
           ni = low priority user mode (nice)
           id = idle task
           wa = I/O waiting
           hi = servicing IRQs
           si = servicing soft IRQs
           st = steal (time given to other DomU instances)

setpriority 用法如下：

NAME
       getpriority, setpriority - get/set program scheduling priority

SYNOPSIS
       #include <sys/time.h>
       #include <sys/resource.h>

       int getpriority(int which, int who);
       int setpriority(int which, int who, int prio);

查看greenplum的源码，在src/backend/tcop/postgres.c中发现了setpriority的踪迹。

/*
 * Change the priority of the current process to the specified level
 * (bigger nice_level values correspond to lower priority).
*/
static bool renice_current_process(int nice_level)
{
#ifdef WIN32
        elog(DEBUG2, "Renicing of processes on Windows currently not supported.");
        return false;
#else
        int prio_out = -1;
        elog(DEBUG2, "Current nice level of the process: %d",
                        getpriority(PRIO_PROCESS, 0));
        prio_out = setpriority(PRIO_PROCESS, 0, nice_level);
        if (prio_out == -1)
        {
                switch (errno)
                {
                case EACCES:
                        elog(DEBUG1, "Could not change priority of the query process, errno: %d (%m).", errno);
                        break;
                case ESRCH:
                        /* ignore this, the backend went away when we weren't looking */
                        break;
                default:
                        elog(DEBUG1, "Could not change priority of the query process, errno: %d (%m).", errno);
                }
                return false;
        }

        elog(DEBUG2, "Reniced process to level %d", getpriority(PRIO_PROCESS, 0));
        return true;
#endif
}

以上函数在exec_mpp_query时被调用：

/*
 * exec_mpp_query
 *
 * Called in a qExec process to read and execute a query plan sent by
 * cdbdisp_dispatchPlan().
 *
 * query_string -- optional query text (C string).
 * serializedQuerytree[len]  -- Query node or (NULL,0) if plan provided.
 * serializedPlantree[len] -- PlannedStmt node, or (NULL,0) if query provided.
 * serializedParms[len] -- optional parameters
 * serializedSliceInfo[len] -- optional SliceTable
 * localSlice -- slice table index
 *
 * Caller may supply either a Query (representing utility command) or
 * a PlannedStmt (representing a planned DML command), but not both.
 */
static void
exec_mpp_query(const char *query_string, 
                           const char * serializedQuerytree, int serializedQuerytreelen,
                           const char * serializedPlantree, int serializedPlantreelen,
                           const char * serializedParams, int serializedParamslen,
                           const char * serializedSliceInfo, int serializedSliceInfolen,
                           const char * seqServerHost, int seqServerPort,
                           int localSlice)
{
...
        /* Downgrade segworker process priority */
                if (gp_segworker_relative_priority != 0)
                {
                        renice_current_process(PostmasterPriority + gp_segworker_relative_priority);
                }

gp_segworker_relative_priority 是一个启动参数，默认是20：

        {
                {"gp_segworker_relative_priority", PGC_POSTMASTER, RESOURCES_MGM,
                        gettext_noop("Priority for the segworkers relative to the postmaster's priority."),
                        NULL,
                        GUC_NO_SHOW_ALL | GUC_NOT_IN_SAMPLE
                },
                &gp_segworker_relative_priority,
                PRIO_MAX,
                0, PRIO_MAX, NULL, NULL
        },
...

src/include/cdb/cdbvars.h:#define PRIO_MAX 20

因此就能解释，为什么使用greenplum会发现大量的nice占比的CPU统计。