Add park sleep related options and CPU util caculation code

Enable sleep in park on X86 and Power. Default sleep configuration is set based on experiments done using restCrud on Quarkus with wrk as the load generator. With the current default configuration, we observed: - ~20% throughput improvement on RHEL9 (Intel Skylake) - ~8% improvement on RHEL10 (Intel Raptor Lake) - ~11% improvement on pLinux Depends on eclipse-omr/omr#8039 Signed-off-by: Gengchen Tuo <gengchen.tuo@ibm.com>

Author: thallium

runtime/oti/j9nonbuilder.h

@@ -6562,6 +6562,13 @@ typedef struct J9JavaVM {
 #endif /* JAVA_SPEC_VERSION >= 24 */
 UDATA disclaimableRAMSegmentCount;
 UDATA disclaimableROMSegmentCount;
+#if defined(OMR_THR_YIELD_ALG)
+omrthread_t cpuUtilCalcThread;
+UDATA cpuUtilCacheInterval;
+int64_t prevProcCPUTime;
+int64_t prevProcTimestamp;
+omrthread_monitor_t cpuUtilCacheMutex;
+#endif /* defined(OMR_THR_YIELD_ALG) */
 } J9JavaVM;
 
 #define J9JFR_SAMPLER_STATE_UNINITIALIZED 0

runtime/vm/jvminit.c

@@ -718,6 +718,9 @@ freeJavaVM(J9JavaVM * vm)
 omrthread_monitor_destroy(vm->globalHotFieldPoolMutex);
 }
 }
+#if defined(OMR_THR_YIELD_ALG)
+omrthread_monitor_destroy(vm->cpuUtilCacheMutex);
+#endif /* defined(OMR_THR_YIELD_ALG) */
 
 if (NULL != vm->classMemorySegments) {
 J9ClassWalkState classWalkState;
@@ -7822,6 +7825,9 @@ protectedInitializeJavaVM(J9PortLibrary* portLibrary, void * userData)
 }
 }
 #endif /* defined(J9VM_OPT_JFR) */
+#if defined(OMR_THR_YIELD_ALG)
+omrthread_monitor_init_with_name(&vm->cpuUtilCacheMutex, 0, "CPU Utilization Cache Mutex");
+#endif /* defined(OMR_THR_YIELD_ALG) */
 
 if (JNI_OK != (stageRC = runInitializationStage(vm, ALL_VM_ARGS_CONSUMED))) {
 goto error;

runtime/vm/threadhelp.cpp

@@ -467,6 +467,42 @@ timeCompensationHelper(J9VMThread *vmThread, U_8 threadHelperType, omrthread_mon
 }
 continueTimeCompensation:
 #endif /* defined(J9VM_OPT_CRIU_SUPPORT) */
+#if defined(OMR_THR_YIELD_ALG)
+static UDATA parkPolicy = **(UDATA **)omrthread_global((char *)"parkPolicy");
+if (2 == parkPolicy) {
+J9JavaVM *vm = vmThread->javaVM;
+PORT_ACCESS_FROM_JAVAVM(vm);
+int64_t currentTime = j9time_nano_time();
+omrthread_process_time_t currentProcCPUTimes = {0};
+static const uintptr_t numberOfCpus = j9sysinfo_get_number_CPUs_by_type(J9PORT_CPU_BOUND);
+
+if (-1 == vm->prevProcTimestamp) {
+intptr_t processTimeRC = omrthread_get_process_times(&currentProcCPUTimes);
+if (0 == processTimeRC) {
+if (omrthread_monitor_try_enter(vm->cpuUtilCacheMutex)) {
+vm->prevProcCPUTime = currentProcCPUTimes._userTime + currentProcCPUTimes._systemTime;
+vm->prevProcTimestamp = currentTime;
+}
+}
+} else {
+UDATA timeDelta = currentTime - vm->prevProcTimestamp;
+if (timeDelta >= vm->cpuUtilCacheInterval * 1000000000){
+intptr_t processTimeRC = omrthread_get_process_times(&currentProcCPUTimes);
+if (0 == processTimeRC) {
+if (omrthread_monitor_try_enter(vm->cpuUtilCacheMutex)) {
+double timeElapsedAllCores = (double)numberOfCpus * timeDelta;
+double util = (currentProcCPUTimes._systemTime + currentProcCPUTimes._userTime - vm->prevProcCPUTime) / timeElapsedAllCores * 100;
+
+**(UDATA **)omrthread_global((char *)"cpuUtilCache") = (UDATA)util;
+
+vm->prevProcCPUTime = currentProcCPUTimes._userTime + currentProcCPUTimes._systemTime;
+vm->prevProcTimestamp = currentTime;
+}
+}
+}
+}
+}
+#endif /* defined(OMR_THR_YIELD_ALG) */
 
 switch(threadHelperType) {
 /* Following three APIs return J9THREAD_TIMED_OUT when the timeout expired.

runtime/vm/vmthread.cpp

@@ -739,11 +739,23 @@ threadParseArguments(J9JavaVM *vm, char *optArg)
 
 
 #if defined(OMR_THR_YIELD_ALG)
+#if defined(J9VM_ARCH_X86) || defined(J9VM_ARCH_POWER)
+**(UDATA **)omrthread_global((char *)"parkPolicy") = 2;
+#else
 **(UDATA **)omrthread_global((char *)"parkPolicy") = 0;
+#endif
 **(UDATA **)omrthread_global((char *)"parkSleepMultiplier") = 0;
-**(UDATA **)omrthread_global((char *)"parkSleepTime") = 0;
+**(UDATA **)omrthread_global((char *)"parkSleepTime") = 400;
 **(UDATA **)omrthread_global((char *)"parkSpinCount") = 0;
-**(UDATA **)omrthread_global((char *)"parkSleepCount") = 0;
+**(UDATA **)omrthread_global((char *)"parkSleepCount") = 2;
+#if defined(J9VM_ARCH_POWER)
+**(UDATA **)omrthread_global((char *)"parkSleepCpuUtilThreshold") = 99;
+#elif defined(J9VM_ARCH_X86) /* defined(OMR_ARCH_POWER) */
+**(UDATA **)omrthread_global((char *)"parkSleepCpuUtilThreshold") = 80;
+#else /* defined(J9VM_ARCH_X86) */
+**(UDATA **)omrthread_global((char *)"parkSleepCpuUtilThreshold") = 100;
+#endif /* defined(J9VM_ARCH_X86) */
+vm->cpuUtilCacheInterval = 5;
 #endif /* defined(OMR_THR_YIELD_ALG) */
 
 /* parse arguments */
@@ -1352,6 +1364,24 @@ threadParseArguments(J9JavaVM *vm, char *optArg)
 **(UDATA **)omrthread_global((char *)"parkSleepCount") = parkSleepCount;
 continue;
 }
+
+if (try_scan(&scan_start, "cpuUtilCacheInterval=")) {
+UDATA cacheInterval = 0;
+if (scan_udata(&scan_start, &cacheInterval)) {
+goto _error;
+}
+vm->cpuUtilCacheInterval = cacheInterval;
+continue;
+}
+
+if (try_scan(&scan_start, "parkSleepCpuUtilThreshold=")) {
+UDATA threshold = 0;
+if (scan_udata(&scan_start, &threshold)) {
+goto _error;
+}
+**(UDATA **)omrthread_global((char *)"parkSleepCpuUtilThreshold") = threshold;
+continue;
+}
 #endif /* defined(OMR_THR_YIELD_ALG) */
 
 /* Couldn't find a match for arguments */