Package org.opends.server.util

Class Platform

java.lang.Object
org.opends.server.util.Platform
public final class Platform extends Object
Utility class for getting the default number of worker threads.

  • Method Details

    • computeDefaultNumberOfReaderThreads

      public static int computeDefaultNumberOfReaderThreads()
      Computes the optimal number of reader threads based on the number of CPUs in the system. The formula is the result of searchrate and modrate benchmarks performed on modern quad-core laptops and high performance 16-core NUMA lab machines.

      The returned value is adapted for hyper-threaded machines which report double the processors they really have, hence we divide by two. Matching the number of threads to the number of cores optimizes for CPU intensive workloads (in memory read). However, it is possible to add a few additional threads in order to take up the slack when cores are blocked waiting for memory accesses, hence the additional N/8 fudge factor.

      Returns:
      The optimal number of reader threads based on the number of CPUs in the system.

    • computeDefaultNumberOfWriterThreads

      public static int computeDefaultNumberOfWriterThreads()
      Computes the optimal number of writer threads based on the number of CPUs in the system. This setting should should be used for sizing threads pools in components which are responsible for performing updates to the backend database (e.g. cleaners, TTL, replay threads).

      Testing found that the optimal value is similar to the number of reader threads. Adding more threads does not yield greater performance due to contention on the database transaction log. Beyond this the threads simply queue up behind other blocked threads and essentially become extensions to the work queue.

      The replication replay thread pool is sized to the nearest power of two so, unlike the reader thread count, this method does not add a hyper-threading "fudge factor" in order to avoid further adjustment.

      Returns:
      The optimal number of writer threads based on the number of CPUs in the system.

    • isVirtualThreadsEnabled

      public static @com.forgerock.opendj.util.FeatureFlag boolean isVirtualThreadsEnabled()
      Returns true if virtual threads are available and should be used.
      Returns:
      true if virtual threads are available and should be used

    • runInVirtualThread

      public static Thread runInVirtualThread(String name, Runnable task)
      Creates a new virtual thread with the provided name and task and then starts it.
      Parameters:
      name - The thread name.
      task - The task to run.
      Returns:
      The started thread.

    • newFairReentrantReadWriteLock

      public static ReadWriteLock newFairReentrantReadWriteLock()
      Returns a new fair reentrant read-write lock.
      Returns:
      a new fair reentrant read-write lock

    • newUnfairNonReentrantReadWriteLock

      public static ReadWriteLock newUnfairNonReentrantReadWriteLock()
      Returns a new unfair non-reentrant read-write lock.. Used in cases where the write lock is highly contended such that unfair scheduling brings significant benefits. Testing indicates that stamped locks are slightly more performant than unfair reentrant read-write locks and can be much more performant (~50%) than fair reentrant read-write locks.
      Returns:
      a new unfair non-reentrant read-write lock