[TD-2375]<enhance>: configure the number of CPU cores available for query processing.

packaging/cfg/taos.cfg

@@ -29,8 +29,11 @@
 # number of threads per CPU core
 # numOfThreadsPerCore       1.0
 
-# the proportion of total threads responsible for query
-# ratioOfQueryThreads       0.5
+# the proportion of total CPU cores available for query processing
+# 1.0: all CPU cores are available for query processing
+# 0.5: only half of the CPU cores are available for query
+# 0.0: only one core available
+# ratioOfQueryThreads       1.0
 
 # number of management nodes in the system
 # numOfMnodes               3
@@ -265,5 +268,5 @@
 # enable/disable stream (continuous query)
 # stream                1
 
-# only 50% CPU resources will be used in query processing
-# halfCoresForQuery     0
+# in retrieve blocking model, only in 50% query threads will be used in query processing in dnode
+# retrieveBlockModel    0

src/common/inc/tglobal.h

@@ -46,7 +46,7 @@ extern int32_t  tsShellActivityTimer;
 extern uint32_t tsMaxTmrCtrl;
 extern float    tsNumOfThreadsPerCore;
 extern int32_t  tsNumOfCommitThreads;
-extern float    tsRatioOfQueryThreads;  // todo remove it
+extern float    tsRatioOfQueryThreads;
 extern int8_t   tsDaylight;
 extern char     tsTimezone[];
 extern char     tsLocale[];
@@ -57,7 +57,7 @@ extern char     tsTempDir[];
 
 //query buffer management
 extern int32_t  tsQueryBufferSize;      // maximum allowed usage buffer for each data node during query processing
-extern int32_t  tsHalfCoresForQuery;         // only 50% will be used in query processing
+extern int32_t  tsRetrieveBlockModel;         // only 50% will be used in query processing
 
 // client
 extern int32_t tsTableMetaKeepTimer;

src/common/src/tglobal.c

@@ -52,7 +52,7 @@ int32_t tsMaxConnections = 5000;
 int32_t tsShellActivityTimer  = 3;  // second
 float   tsNumOfThreadsPerCore = 1.0f;
 int32_t tsNumOfCommitThreads = 1;
-float   tsRatioOfQueryThreads = 0.5f;
+float   tsRatioOfQueryThreads = 1.0f;
 int8_t  tsDaylight       = 0;
 char    tsTimezone[TSDB_TIMEZONE_LEN] = {0};
 char    tsLocale[TSDB_LOCALE_LEN] = {0};
@@ -107,8 +107,8 @@ int64_t tsMaxRetentWindow = 24 * 3600L;  // maximum time window tolerance
 // positive value (in MB)
 int32_t tsQueryBufferSize = -1;
 
-// only 50% cpu will be used in query processing in dnode
-int32_t tsHalfCoresForQuery = 0;
+// in retrieve blocking model, only in 50% query threads will be used in query processing in dnode
+int32_t tsRetrieveBlockModel = 0;
 
 // db parameters
 int32_t tsCacheBlockSize = TSDB_DEFAULT_CACHE_BLOCK_SIZE;
@@ -887,8 +887,8 @@ static void doInitGlobalConfig(void) {
   cfg.unitType = TAOS_CFG_UTYPE_BYTE;
   taosInitConfigOption(cfg);
 
-  cfg.option = "halfCoresForQuery";
-  cfg.ptr = &tsHalfCoresForQuery;
+  cfg.option = "retrieveBlockModel";
+  cfg.ptr = &tsRetrieveBlockModel;
   cfg.valType = TAOS_CFG_VTYPE_INT32;
   cfg.cfgType = TSDB_CFG_CTYPE_B_CONFIG | TSDB_CFG_CTYPE_B_SHOW;
   cfg.minValue = 0;

src/dnode/src/dnodeShell.c

@@ -70,8 +70,7 @@ int32_t dnodeInitShell() {
 
   dnodeProcessShellMsgFp[TSDB_MSG_TYPE_NETWORK_TEST]   = dnodeSendStartupStep;
 
-  int32_t numOfThreads = tsNumOfCores * tsNumOfThreadsPerCore;
-  numOfThreads = (int32_t) ((1.0 - tsRatioOfQueryThreads) * numOfThreads / 2.0);
+  int32_t numOfThreads = (tsNumOfCores * tsNumOfThreadsPerCore) / 2.0;
   if (numOfThreads < 1) {
     numOfThreads = 1;
   }

src/dnode/src/dnodeVRead.c

@@ -26,16 +26,17 @@ static SWorkerPool tsVQueryWP;
 static SWorkerPool tsVFetchWP;
 
 int32_t dnodeInitVRead() {
+  const int32_t maxFetchThreads = 4;
+
   tsVQueryWP.name = "vquery";
   tsVQueryWP.workerFp = dnodeProcessReadQueue;
-  tsVQueryWP.min = tsNumOfCores;
-  tsVQueryWP.max = tsNumOfCores/* * tsNumOfThreadsPerCore*/;
-//  if (tsVQueryWP.max <= tsVQueryWP.min * 2) tsVQueryWP.max = 2 * tsVQueryWP.min;
+  tsVQueryWP.min = tsNumOfCores * tsRatioOfQueryThreads;
+  tsVQueryWP.max = tsVQueryWP.min;
   if (tWorkerInit(&tsVQueryWP) != 0) return -1;
 
   tsVFetchWP.name = "vfetch";
   tsVFetchWP.workerFp = dnodeProcessReadQueue;
-  tsVFetchWP.min = MIN(4, tsNumOfCores);
+  tsVFetchWP.min = MIN(maxFetchThreads, tsNumOfCores);
   tsVFetchWP.max = tsVFetchWP.min;
   if (tWorkerInit(&tsVFetchWP) != 0) return -1;
 

src/query/src/qExecutor.c

@@ -7635,7 +7635,7 @@ int32_t qRetrieveQueryResultInfo(qinfo_t qinfo, bool* buildRes, void* pRspContex
 
   int32_t code = TSDB_CODE_SUCCESS;
 
-  if (tsHalfCoresForQuery) {
+  if (tsRetrieveBlockModel) {
     pQInfo->rspContext = pRspContext;
     tsem_wait(&pQInfo->ready);
     *buildRes = true;

src/vnode/src/vnodeRead.c

@@ -281,7 +281,7 @@ static int32_t vnodeProcessQueryMsg(SVnodeObj *pVnode, SVReadMsg *pRead) {
     vDebug("vgId:%d, QInfo:%p, dnode continues to exec query", pVnode->vgId, *qhandle);
 
     // In the retrieve blocking model, only 50% CPU will be used in query processing
-    if (tsHalfCoresForQuery) {
+    if (tsRetrieveBlockModel) {
       qTableQuery(*qhandle);  // do execute query
       qReleaseQInfo(pVnode->qMgmt, (void **)&qhandle, false);
     } else {
@@ -380,7 +380,7 @@ static int32_t vnodeProcessFetchMsg(SVnodeObj *pVnode, SVReadMsg *pRead) {
     freeHandle = true;
   } else {  // result is not ready, return immediately
     // Only effects in the non-blocking model
-    if (!tsHalfCoresForQuery) {
+    if (!tsRetrieveBlockModel) {
       if (!buildRes) {
         assert(pRead->rpcHandle != NULL);