Limit number of concurrent workers started by hardware concurrency. (#2753)

* Limit number of concurrent workers started by hardware concurrency.

* Check if std::thread::hardware_concurrency() returns 0.

* Pass in max concurrency from Python.

* Fix Java call to startRaylet.

* Fix typo

* Remove unnecessary cast.

* Fix linting.

* Cleanups on Java side.

* Comment back in actor test.

* Require maximum_startup_concurrency to be at least 1.

* Fix linting and test.

* Improve documentation.

* Fix typo.

java/runtime-native/src/main/java/org/ray/runner/RunManager.java

@@ -623,10 +623,14 @@ public class RunManager {
 
     String resourceArgument = ResourceUtil.getResourcesStringFromMap(staticResources);
 
+    int hardwareConcurrency = Runtime.getRuntime().availableProcessors();
+    int maximumStartupConcurrency = Math.max(1, Math.min(staticResources.get("CPU").intValue(),
+                                                         hardwareConcurrency));
+
     // The second-last arugment is the worker command for Python, not needed for Java.
     String[] cmds = new String[]{filePath, rayletSocketName, storeName, ip, gcsIp,
-        gcsPort, "" + numWorkers, resourceArgument,
-        "", workerCommand};
+        gcsPort, String.valueOf(numWorkers), String.valueOf(maximumStartupConcurrency),
+        resourceArgument, "", workerCommand};
 
     Process p = startProcess(cmds, null, RunInfo.ProcessType.PT_RAYLET,
         "raylet", redisAddress, ip, redirect, cleanup);