protobuf: cache temp buffers

Before:
TransportBenchmark.unaryCall1024                                       true        NETTY  sample  4564   2188854.745 ±  71456.423   ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.00                   true        NETTY  sample         1875968.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.50                   true        NETTY  sample         2105344.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.90                   true        NETTY  sample         2396160.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.95                   true        NETTY  sample         2535424.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.99                   true        NETTY  sample         3011993.600                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.999                  true        NETTY  sample         7471595.520                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.9999                 true        NETTY  sample        99090432.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p1.00                   true        NETTY  sample        99090432.000                ns/op
TransportBenchmark.unaryCall1024:·gc.alloc.rate                        true        NETTY  sample    10      2787.784 ±    169.945  MB/sec
TransportBenchmark.unaryCall1024:·gc.alloc.rate.norm                   true        NETTY  sample    10   6415272.837 ±    262.046    B/op
TransportBenchmark.unaryCall1024:·gc.churn.PS_Eden_Space               true        NETTY  sample    10      2815.863 ±    429.465  MB/sec
TransportBenchmark.unaryCall1024:·gc.churn.PS_Eden_Space.norm          true        NETTY  sample    10   6483440.294 ± 947355.959    B/op
TransportBenchmark.unaryCall1024:·gc.churn.PS_Survivor_Space           true        NETTY  sample    10         2.143 ±      1.623  MB/sec
TransportBenchmark.unaryCall1024:·gc.churn.PS_Survivor_Space.norm      true        NETTY  sample    10      4873.798 ±   3679.598    B/op
TransportBenchmark.unaryCall1024:·gc.count                             true        NETTY  sample    10        42.000               counts
TransportBenchmark.unaryCall1024:·gc.time                              true        NETTY  sample    10       155.000                   ms

After:
Benchmark                                                          (direct)  (transport)    Mode   Cnt         Score        Error   Units
TransportBenchmark.unaryCall1024                                       true        NETTY  sample  5037   1982881.569 ±  16738.841   ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.00                   true        NETTY  sample         1683456.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.50                   true        NETTY  sample         1918976.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.90                   true        NETTY  sample         2232320.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.95                   true        NETTY  sample         2330624.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.99                   true        NETTY  sample         2729574.400                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.999                  true        NETTY  sample         6127304.704                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p0.9999                 true        NETTY  sample        15515648.000                ns/op
TransportBenchmark.unaryCall1024:unaryCall1024·p1.00                   true        NETTY  sample        15515648.000                ns/op
TransportBenchmark.unaryCall1024:·gc.alloc.rate                        true        NETTY  sample    10      2071.435 ±    141.669  MB/sec
TransportBenchmark.unaryCall1024:·gc.alloc.rate.norm                   true        NETTY  sample    10   4318096.849 ±    269.655    B/op
TransportBenchmark.unaryCall1024:·gc.churn.PS_Eden_Space               true        NETTY  sample    10      2076.282 ±    323.504  MB/sec
TransportBenchmark.unaryCall1024:·gc.churn.PS_Eden_Space.norm          true        NETTY  sample    10   4335884.918 ± 729189.378    B/op
TransportBenchmark.unaryCall1024:·gc.churn.PS_Survivor_Space           true        NETTY  sample    10         1.567 ±      1.238  MB/sec
TransportBenchmark.unaryCall1024:·gc.churn.PS_Survivor_Space.norm      true        NETTY  sample    10      3274.883 ±   2640.345    B/op
TransportBenchmark.unaryCall1024:·gc.count                             true        NETTY  sample    10        31.000               counts
TransportBenchmark.unaryCall1024:·gc.time                              true        NETTY  sample    10        51.000                   ms

1 files changed, 22 insertions, 5 deletions

diff --git a/protobuf-lite/src/main/java/io/grpc/protobuf/lite/ProtoLiteUtils.java b/protobuf-lite/src/main/java/io/grpc/protobuf/lite/ProtoLiteUtils.java
index 98e781885..8cf95115f 100644
--- a/protobuf-lite/src/main/java/io/grpc/protobuf/lite/ProtoLiteUtils.java
+++ b/protobuf-lite/src/main/java/io/grpc/protobuf/lite/ProtoLiteUtils.java
@@ -78,6 +78,18 @@ public class ProtoLiteUtils {
     globalRegistry = checkNotNull(newRegistry, "newRegistry");
   }
 
+  /**
+   * Local cache of buffers to use for parsing.  ThreadLocal used a WeakReference internally, so
+   * these will not be retained.
+   */
+  private static final ThreadLocal<byte[]> bufs = new ThreadLocal<byte[]>() {
+
+    @Override
+    protected byte[] initialValue() {
+      return new byte[4096]; // Picked at random.
+    }
+  };
+
   /** Create a {@code Marshaller} for protos of the same type as {@code defaultInstance}. */
   public static <T extends MessageLite> Marshaller<T> marshaller(final T defaultInstance) {
     @SuppressWarnings("unchecked")
@@ -129,16 +141,21 @@ public class ProtoLiteUtils {
           if (stream instanceof KnownLength) {
             int size = stream.available();
             if (size > 0 && size <= GrpcUtil.DEFAULT_MAX_MESSAGE_SIZE) {
-              byte[] buf = new byte[size];
+              // Coded Input stream does not escape, so buf does not escape.
+              byte[] buf = bufs.get();
+              if (buf.length < size) {
+                buf = new byte[size];
+                bufs.set(buf);
+              }
               int chunkSize;
               int position = 0;
-              while ((chunkSize = stream.read(buf, position, buf.length - position)) != -1) {
+              while ((chunkSize = stream.read(buf, position, size - position)) != -1) {
                 position += chunkSize;
               }
-              if (buf.length != position) {
-                throw new RuntimeException("size inaccurate: " + buf.length + " != " + position);
+              if (size != position) {
+                throw new RuntimeException("size inaccurate: " + size + " != " + position);
               }
-              cis = CodedInputStream.newInstance(buf);
+              cis = CodedInputStream.newInstance(buf, 0, size);
             } else if (size == 0) {
               return defaultInstance;
             }