HPCC-12739 Add an option to not return memory back to OS

* Retaining the committed pages avoids them being cleared when
  they are required the next time.

Signed-off-by: Gavin Halliday <gavin.halliday@lexisnexis.com>

common/thorhelper/roxierow.cpp

@@ -697,7 +697,7 @@ protected:
 
     void testSetup()
     {
-        setTotalMemoryLimit(false, true, 40*HEAP_ALIGNMENT_SIZE, 0, NULL, NULL);
+        setTotalMemoryLimit(false, true, false, 40*HEAP_ALIGNMENT_SIZE, 0, NULL, NULL);
     }
 
     void testCleanup()

ecl/eclagent/eclagent.cpp

@@ -2027,6 +2027,9 @@ void EclAgent::runProcess(IEclProcess *process)
     bool allowTransparentHugePages = agentTopology->getPropBool("@heapUseTransparentHugePages", true);
     allowTransparentHugePages = globals->getPropBool("heapUseTransparentHugePages", allowTransparentHugePages);
 
+    bool retainMemory = agentTopology->getPropBool("@heapRetainMemory", false);
+    retainMemory = globals->getPropBool("heapRetainMemory", retainMemory);
+
 #ifndef __64BIT__
     if (memLimitMB > 4096)
     {
@@ -2036,7 +2039,7 @@ void EclAgent::runProcess(IEclProcess *process)
     }
 #endif
     memsize_t memLimitBytes = (memsize_t)memLimitMB * 1024 * 1024;
-    roxiemem::setTotalMemoryLimit(allowHugePages, allowTransparentHugePages, memLimitBytes, 0, NULL, NULL);
+    roxiemem::setTotalMemoryLimit(allowHugePages, allowTransparentHugePages, retainMemory, memLimitBytes, 0, NULL, NULL);
 
     rowManager.setown(roxiemem::createRowManager(0, NULL, queryDummyContextLogger(), allocatorMetaCache, false));
     setHThorRowManager(rowManager.get());

initfiles/componentfiles/configxml/eclagent_config.xsd.in

@@ -198,6 +198,14 @@
       </xs:annotation>
     </xs:attribute>
 
+    <xs:attribute name="heapRetainMemory" type="xs:boolean" default="false">
+      <xs:annotation>
+        <xs:appinfo>
+          <tooltip>Retain and do not return unused memory to the operating system.</tooltip>
+        </xs:appinfo>
+      </xs:annotation>
+    </xs:attribute>
+
     <xs:attribute name="pluginDirectory" type="absolutePath" use="optional" default="${PLUGINS_PATH}/">
       <xs:annotation>
         <xs:appinfo>

initfiles/componentfiles/configxml/roxie.xsd.in

@@ -765,6 +765,13 @@
         </xs:appinfo>
       </xs:annotation>
     </xs:attribute>
+    <xs:attribute name="heapRetainMemory" type="xs:boolean" default="false">
+      <xs:annotation>
+        <xs:appinfo>
+          <tooltip>Retain and do not return unused memory to the operating system.</tooltip>
+        </xs:appinfo>
+      </xs:annotation>
+    </xs:attribute>
     <xs:attribute name="trapTooManyActiveQueries" type="xs:boolean" use="optional" default="true">
       <xs:annotation>
         <xs:appinfo>

initfiles/componentfiles/configxml/thor.xsd.in

@@ -388,6 +388,13 @@
           </xs:appinfo>
         </xs:annotation>
       </xs:attribute>
+      <xs:attribute name="heapRetainMemory" type="xs:boolean" default="false">
+        <xs:annotation>
+          <xs:appinfo>
+            <tooltip>Retain and do not return unused memory to the operating system.</tooltip>
+          </xs:appinfo>
+        </xs:annotation>
+      </xs:attribute>
       <xs:attribute name="pluginsPath" type="relativePath" default="${PLUGINS_PATH}/"/>
       <xs:attribute name="nodeGroup" type="xs:string" use="optional">
         <xs:annotation>

roxie/ccd/ccdmain.cpp

@@ -759,9 +759,10 @@ int STARTQUERY_API start_query(int argc, const char *argv[])
         memsize_t totalMemoryLimit = (memsize_t) topology->getPropInt64("@totalMemoryLimit", 0);
         bool allowHugePages = topology->getPropBool("@heapUseHugePages", false);
         bool allowTransparentHugePages = topology->getPropBool("@heapUseTransparentHugePages", true);
+        bool retainMemory = topology->getPropBool("@heapRetainMemory", false);
         if (!totalMemoryLimit)
             totalMemoryLimit = 1024 * 0x100000;  // 1 Gb;
-        roxiemem::setTotalMemoryLimit(allowHugePages, allowTransparentHugePages, totalMemoryLimit, 0, NULL, NULL);
+        roxiemem::setTotalMemoryLimit(allowHugePages, allowTransparentHugePages, retainMemory, totalMemoryLimit, 0, NULL, NULL);
 
         traceStartStop = topology->getPropBool("@traceStartStop", false);
         traceServerSideCache = topology->getPropBool("@traceServerSideCache", false);

roxie/ccd/ccdserver.cpp

@@ -26607,7 +26607,7 @@ protected:
 
     void testSetup()
     {
-        roxiemem::setTotalMemoryLimit(false, true, 100 * 1024 * 1024, 0, NULL, NULL);
+        roxiemem::setTotalMemoryLimit(false, true, false, 100 * 1024 * 1024, 0, NULL, NULL);
     }
 
     void testCleanup()

roxie/roxiemem/roxiemem.cpp

@@ -138,7 +138,7 @@ typedef MapBetween<unsigned, unsigned, memsize_t, memsize_t> MapActivityToMemsiz
 
 static CriticalSection heapBitCrit;
 
-static void initializeHeap(bool allowHugePages, bool allowTransparentHugePages, unsigned pages, unsigned largeBlockGranularity, ILargeMemCallback * largeBlockCallback)
+static void initializeHeap(bool allowHugePages, bool allowTransparentHugePages, bool retainMemory, unsigned pages, unsigned largeBlockGranularity, ILargeMemCallback * largeBlockCallback)
 {
     if (heapBase) return;
 
@@ -148,6 +148,9 @@ static void initializeHeap(bool allowHugePages, bool allowTransparentHugePages,
     heapLargeBlockGranularity = largeBlockGranularity;
     heapLargeBlockCallback = largeBlockCallback;
     memsize_t memsize = memsize_t(heapTotalPages) * HEAP_ALIGNMENT_SIZE;
+
+    heapNotifyUnusedEachFree = !retainMemory;
+    heapNotifyUnusedEachBlock = false;
 #ifdef _WIN32
     // Not the world's best code but will do 
     char *next = (char *) HEAP_ALIGNMENT_SIZE;
@@ -174,8 +177,7 @@ static void initializeHeap(bool allowHugePages, bool allowTransparentHugePages,
         {
             heapUseHugePages = true;
             heapNotifyUnusedEachFree = false;
-            //MORE: At the moment I'm not sure calling madvise() has any benefit, it may be better for the following to stay false;
-            heapNotifyUnusedEachBlock = true;
+            heapNotifyUnusedEachBlock = !retainMemory;
             DBGLOG("Using Huge Pages for roxiemem");
         }
         else
@@ -234,13 +236,12 @@ static void initializeHeap(bool allowHugePages, bool allowTransparentHugePages,
                 if ((heapBlockSize % hugePageSize) == 0)
                 {
                     //If we notify heapBlockSize items at a time it will always be a multiple of hugePageSize so shouldn't trigger defragmentation
-                    heapNotifyUnusedEachBlock = true;
-                    DBGLOG("Transparent huge pages used for roxiemem heap - memory released in blocks");
+                    heapNotifyUnusedEachBlock = !retainMemory;
+                    DBGLOG("Transparent huge pages used for roxiemem heap");
                 }
                 else
                 {
-                    DBGLOG("Transparent huge pages used for roxiemem heap- MEMORY WILL NOT BE RELEASED.");
-                    DBGLOG("Increase HEAP_ALIGNMENT_SIZE so HEAP_ALIGNMENT_SIZE*32 is a multiple of huge page size"");");
+                    DBGLOG("Transparent huge pages used for roxiemem heap");
                 }
             }
         }
@@ -260,6 +261,17 @@ static void initializeHeap(bool allowHugePages, bool allowTransparentHugePages,
     }
 #endif
 
+    if (heapNotifyUnusedEachFree)
+        DBGLOG("Memory released to OS on each %uk 'page'", (unsigned)(HEAP_ALIGNMENT_SIZE/1024));
+    else if (heapNotifyUnusedEachBlock)
+        DBGLOG("Memory released to OS in %uk blocks", (unsigned)(HEAP_ALIGNMENT_SIZE*UNSIGNED_BITS/1024));
+    else
+    {
+        DBGLOG("MEMORY WILL NOT BE RELEASED TO OS");
+        if (!retainMemory)
+            DBGLOG("Increase HEAP_ALIGNMENT_SIZE so HEAP_ALIGNMENT_SIZE*32 is a multiple of system huge page size");
+    }
+
     assertex(((memsize_t)heapBase & (HEAP_ALIGNMENT_SIZE-1)) == 0);
 
     heapEnd = heapBase + memsize;
@@ -4250,7 +4262,7 @@ extern void setMemoryStatsInterval(unsigned secs)
     lastStatsCycles = get_cycles_now();
 }
 
-extern void setTotalMemoryLimit(bool allowHugePages, bool allowTransparentHugePages, memsize_t max, memsize_t largeBlockSize, const unsigned * allocSizes, ILargeMemCallback * largeBlockCallback)
+extern void setTotalMemoryLimit(bool allowHugePages, bool allowTransparentHugePages, bool retainMemory, memsize_t max, memsize_t largeBlockSize, const unsigned * allocSizes, ILargeMemCallback * largeBlockCallback)
 {
     assertex(largeBlockSize == align_pow2(largeBlockSize, HEAP_ALIGNMENT_SIZE));
     unsigned totalMemoryLimit = (unsigned) (max / HEAP_ALIGNMENT_SIZE);
@@ -4259,7 +4271,7 @@ extern void setTotalMemoryLimit(bool allowHugePages, bool allowTransparentHugePa
         totalMemoryLimit = 1;
     if (memTraceLevel)
         DBGLOG("RoxieMemMgr: Setting memory limit to %"I64F"d bytes (%u pages)", (unsigned __int64) max, totalMemoryLimit);
-    initializeHeap(allowHugePages, allowTransparentHugePages, totalMemoryLimit, largeBlockGranularity, largeBlockCallback);
+    initializeHeap(allowHugePages, allowTransparentHugePages, retainMemory, totalMemoryLimit, largeBlockGranularity, largeBlockCallback);
     initAllocSizeMappings(allocSizes ? allocSizes : defaultAllocSizes);
 }
 
@@ -4512,7 +4524,7 @@ protected:
         ASSERT(HugeHeaplet::dataOffset() >= sizeof(HugeHeaplet));
 
         memsize_t memory = (useLargeMemory ? largeMemory : smallMemory) * (unsigned __int64)0x100000U;
-        initializeHeap(false, true, (unsigned)(memory / HEAP_ALIGNMENT_SIZE), 0, NULL);
+        initializeHeap(false, true, false, (unsigned)(memory / HEAP_ALIGNMENT_SIZE), 0, NULL);
         initAllocSizeMappings(defaultAllocSizes);
     }
 
@@ -5774,7 +5786,7 @@ public:
 protected:
     void testSetup()
     {
-        setTotalMemoryLimit(true, true, memorySize, 0, NULL, NULL);
+        setTotalMemoryLimit(true, true, false, memorySize, 0, NULL, NULL);
     }
 
     void testCleanup()
@@ -5970,7 +5982,7 @@ public:
 protected:
     void testSetup()
     {
-        setTotalMemoryLimit(true, true, hugeMemorySize, 0, NULL, NULL);
+        setTotalMemoryLimit(true, true, false, hugeMemorySize, 0, NULL, NULL);
     }
 
     void testCleanup()

roxie/roxiemem/roxiemem.hpp

@@ -498,7 +498,7 @@ interface IDataBufferManager : extends IInterface
 
 extern roxiemem_decl IDataBufferManager *createDataBufferManager(size32_t size);
 extern roxiemem_decl void setMemoryStatsInterval(unsigned secs);
-extern roxiemem_decl void setTotalMemoryLimit(bool allowHugePages, bool allowTransparentHugePages, memsize_t max, memsize_t largeBlockSize, const unsigned * allocSizes, ILargeMemCallback * largeBlockCallback);
+extern roxiemem_decl void setTotalMemoryLimit(bool allowHugePages, bool allowTransparentHugePages, bool retainMemory, memsize_t max, memsize_t largeBlockSize, const unsigned * allocSizes, ILargeMemCallback * largeBlockCallback);
 extern roxiemem_decl memsize_t getTotalMemoryLimit();
 extern roxiemem_decl void releaseRoxieHeap();
 extern roxiemem_decl bool memPoolExhausted();

roxie/udplib/uttest.cpp

@@ -754,7 +754,7 @@ int main(int argc, char * argv[] )
             printf("ERROR: my ip does not appear to be in range\n");
             usage();
         }
-        roxiemem::setTotalMemoryLimit(false, true, 1048576000, 0, NULL, NULL);
+        roxiemem::setTotalMemoryLimit(false, true, false, 1048576000, 0, NULL, NULL);
         testNxN();
         roxiemem::releaseRoxieHeap();
     }

thorlcr/master/thmastermain.cpp

@@ -684,12 +684,13 @@ int main( int argc, char *argv[]  )
         }
         bool gmemAllowHugePages = globals->getPropBool("@heapUseHugePages", false);
         bool gmemAllowTransparentHugePages = globals->getPropBool("@heapUseTransparentHugePages", true);
+        bool gmemRetainMemory = globals->getPropBool("@heapRetainMemory", false);
 
         // if @masterMemorySize and @globalMemorySize unspecified gmemSize will be default based on h/w
         globals->setPropInt("@masterMemorySize", mmemSize);
 
         PROGLOG("Global memory size = %d MB", mmemSize);
-        roxiemem::setTotalMemoryLimit(gmemAllowHugePages, gmemAllowTransparentHugePages, ((memsize_t)mmemSize) * 0x100000, 0, thorAllocSizes, NULL);
+        roxiemem::setTotalMemoryLimit(gmemAllowHugePages, gmemAllowTransparentHugePages, gmemRetainMemory, ((memsize_t)mmemSize) * 0x100000, 0, thorAllocSizes, NULL);
 
         const char * overrideBaseDirectory = globals->queryProp("@thorDataDirectory");
         const char * overrideReplicateDirectory = globals->queryProp("@thorReplicateDirectory");

thorlcr/slave/slavmain.cpp

@@ -664,11 +664,13 @@ void slaveMain()
     unsigned gmemSize = globals->getPropInt("@globalMemorySize");
     bool gmemAllowHugePages = globals->getPropBool("@heapUseHugePages", false);
     bool gmemAllowTransparentHugePages = globals->getPropBool("@heapUseTransparentHugePages", true);
+    bool gmemRetainMemory = globals->getPropBool("@heapRetainMemory", false);
+
     if (gmemSize >= hdwInfo.totalMemory)
     {
         // should prob. error here
     }
-    roxiemem::setTotalMemoryLimit(gmemAllowHugePages, gmemAllowTransparentHugePages, ((memsize_t)gmemSize) * 0x100000, 0, thorAllocSizes, NULL);
+    roxiemem::setTotalMemoryLimit(gmemAllowHugePages, gmemAllowTransparentHugePages, gmemRetainMemory, ((memsize_t)gmemSize) * 0x100000, 0, thorAllocSizes, NULL);
 
     CJobListener jobListener;
     CThorResourceSlave slaveResource;