Merge pull request #13475 from jtattermusch/better_completion_registry_benchmark

Use spinlock in CompletionRegistry (and improve the benchmark).

src/csharp/Grpc.Core/Internal/CompletionRegistry.cs

@@ -19,7 +19,9 @@
 using System;
 using System.Collections.Concurrent;
 using System.Collections.Generic;
+using System.Diagnostics;
 using System.Runtime.InteropServices;
+using System.Threading;
 using Grpc.Core.Logging;
 using Grpc.Core.Utils;
 
@@ -35,7 +37,7 @@ namespace Grpc.Core.Internal
 
         readonly GrpcEnvironment environment;
         readonly Dictionary<IntPtr, IOpCompletionCallback> dict = new Dictionary<IntPtr, IOpCompletionCallback>(new IntPtrComparer());
-        readonly object myLock = new object();
+        SpinLock spinLock = new SpinLock(Debugger.IsAttached);
         IntPtr lastRegisteredKey;  // only for testing
 
         public CompletionRegistry(GrpcEnvironment environment)
@@ -46,11 +48,19 @@ namespace Grpc.Core.Internal
         public void Register(IntPtr key, IOpCompletionCallback callback)
         {
             environment.DebugStats.PendingBatchCompletions.Increment();
-            lock (myLock)
+
+            bool lockTaken = false;
+            try
             {
+                spinLock.Enter(ref lockTaken);
+
                 dict.Add(key, callback);
                 this.lastRegisteredKey = key;
             }
+            finally
+            {
+                if (lockTaken) spinLock.Exit();
+            }
         }
 
         public void RegisterBatchCompletion(BatchContextSafeHandle ctx, BatchCompletionDelegate callback, object state)
@@ -68,11 +78,18 @@ namespace Grpc.Core.Internal
         public IOpCompletionCallback Extract(IntPtr key)
         {
             IOpCompletionCallback value = null;
-            lock (myLock)
+            bool lockTaken = false;
+            try
             {
+                spinLock.Enter(ref lockTaken);
+
                 value = dict[key];
                 dict.Remove(key);
             }
+            finally
+            {
+                if (lockTaken) spinLock.Exit();
+            }
             environment.DebugStats.PendingBatchCompletions.Decrement();
             return value;
         }

src/csharp/Grpc.Microbenchmarks/CompletionRegistryBenchmark.cs

@@ -40,24 +40,25 @@ namespace Grpc.Microbenchmarks
             GrpcEnvironment.ReleaseAsync().Wait();
         }
 
-        public void Run(int threadCount, int iterations)
+        public void Run(int threadCount, int iterations, bool useSharedRegistry)
         {
-            Console.WriteLine(string.Format("CompletionRegistryBenchmark: threads={0}, iterations={1}", threadCount, iterations));
+            Console.WriteLine(string.Format("CompletionRegistryBenchmark: threads={0}, iterations={1}, useSharedRegistry={2}", threadCount, iterations, useSharedRegistry));
-            var threadedBenchmark = new ThreadedBenchmark(threadCount, () => ThreadBody(iterations));
+            CompletionRegistry sharedRegistry = useSharedRegistry ? new CompletionRegistry(environment) : null;
+            var threadedBenchmark = new ThreadedBenchmark(threadCount, () => ThreadBody(iterations, sharedRegistry));
             threadedBenchmark.Run();
             // TODO: parametrize by number of pending completions
         }
 
-        private void ThreadBody(int iterations)
+        private void ThreadBody(int iterations, CompletionRegistry optionalSharedRegistry)
         {
-            var completionRegistry = new CompletionRegistry(environment);
+            var completionRegistry = optionalSharedRegistry ?? new CompletionRegistry(environment);
             var ctx = BatchContextSafeHandle.Create();
   
             var stopwatch = Stopwatch.StartNew();
             for (int i = 0; i < iterations; i++)
             {
                 completionRegistry.Register(ctx.Handle, ctx);
-                var callback = completionRegistry.Extract(completionRegistry.LastRegisteredKey);
+                var callback = completionRegistry.Extract(ctx.Handle);
                 // NOTE: we are not calling the callback to avoid disposing ctx.
             }
             stopwatch.Stop();

src/csharp/Grpc.Microbenchmarks/Program.cs

@@ -77,7 +77,10 @@ namespace Grpc.Microbenchmarks
             benchmark.Init();
             foreach (int threadCount in new int[] {1, 1, 2, 4, 8, 12})
             {
-                benchmark.Run(threadCount, 4 * 1000 * 1000);
+                foreach (bool useSharedRegistry in new bool[] {false, true})
+                {
+                    benchmark.Run(threadCount, 4 * 1000 * 1000, useSharedRegistry);
+                }
             }
             benchmark.Cleanup();
         }