Proto optimization

bench/src/main/scala/perfio/BufferedInputUncheckedBenchmark.scala

@@ -2,6 +2,7 @@ package perfio
 
 import org.openjdk.jmh.annotations.*
 import org.openjdk.jmh.infra.*
+import perfio.internal.BufferUtil
 
 import java.util.concurrent.TimeUnit
 

bench/src/main/scala/perfio/BufferedOutputNumBenchmark.scala

@@ -2,6 +2,7 @@ package perfio
 
 import org.openjdk.jmh.annotations.*
 import org.openjdk.jmh.infra.*
+import perfio.internal.BufferUtil
 
 import java.io.*
 import java.nio.ByteBuffer

bench/src/main/scala/perfio/BufferedOutputStringBenchmark.scala

@@ -99,21 +99,21 @@ class BufferedOutputStringBenchmark extends BenchUtil:
     bh.consume(bout.getBuffer)
 
   @Benchmark
-  def array_FullyBufferedOutput_growing(bh: Blackhole): Unit =
+  def array_ArrayBufferedOutput_growing(bh: Blackhole): Unit =
     val out = BufferedOutput.growing()
     writeTo(out)
     bh.consume(out.buffer)
     bh.consume(out.length)
 
   @Benchmark
-  def array_FullyBufferedOutput_growing_preallocated(bh: Blackhole): Unit =
+  def array_ArrayBufferedOutput_growing_preallocated(bh: Blackhole): Unit =
     val out = BufferedOutput.growing(totalLength)
     writeTo(out)
     bh.consume(out.buffer)
     bh.consume(out.length)
 
   @Benchmark
-  def array_FullyBufferedOutput_fixed(bh: Blackhole): Unit =
+  def array_ArrayBufferedOutput_fixed(bh: Blackhole): Unit =
     val out = BufferedOutput.ofArray(new Array[Byte](totalLength))
     writeTo(out)
     bh.consume(out.buffer)

bench/src/main/scala/perfio/BufferedOutputUncheckedBenchmark.scala

@@ -2,6 +2,7 @@ package perfio
 
 import org.openjdk.jmh.annotations.*
 import org.openjdk.jmh.infra.*
+import perfio.internal.BufferUtil
 
 import java.util.concurrent.TimeUnit
 

bench/src/main/scala/perfio/LineTokenizerBenchmark.scala

@@ -2,6 +2,7 @@ package perfio
 
 import org.openjdk.jmh.annotations.*
 import org.openjdk.jmh.infra.*
+import perfio.internal.StringInternals
 
 import java.io.*
 import java.lang.foreign.MemorySegment

bench/src/main/scala/perfio/TextOutputBenchmark.scala

@@ -2,6 +2,7 @@ package perfio
 
 import org.openjdk.jmh.annotations.*
 import org.openjdk.jmh.infra.*
+import perfio.internal.StringInternals
 
 import java.io.*
 import java.nio.charset.Charset

proto-bench/src/main/scala/ParseBenchmark.scala → proto-bench/src/main/scala/ReadBenchmark.scala

@@ -18,7 +18,7 @@ import java.nio.file.{Files, Path}
 @Measurement(iterations = 7, time = 1)
 @OutputTimeUnit(TimeUnit.MICROSECONDS)
 @State(Scope.Benchmark)
-class ParseBenchmark:
+class ReadBenchmark:
 
   private var testData: Array[Byte] = null
   private val testDataFile = Path.of("src/main/resources/CodeGeneratorRequest-pack.bin")

proto-bench/src/main/scala/ReadWriteBenchmark.scala

@@ -0,0 +1,71 @@
+package perfio;
+
+import com.google.protobuf.compiler.PluginProtos as GPluginProtos
+import org.openjdk.jmh.annotations.*
+import org.openjdk.jmh.infra.*
+import perfio.protoapi.PluginProtos as PPluginProtos
+
+import java.io.{BufferedInputStream, BufferedOutputStream, ByteArrayOutputStream, FileInputStream, FileOutputStream}
+import java.nio.ByteOrder
+import java.nio.file.{Files, Path}
+import java.util.concurrent.TimeUnit
+
+@BenchmarkMode(Array(Mode.AverageTime))
+@Fork(value = 1, jvmArgsAppend = Array("-Xmx12g", "-Xss32M", "-XX:+UnlockExperimentalVMOptions", "-XX:+UseZGC", "--enable-native-access=ALL-UNNAMED", "--add-modules", "jdk.incubator.vector"))
+@Threads(1)
+@Warmup(iterations = 7, time = 1)
+@Measurement(iterations = 7, time = 1)
+@OutputTimeUnit(TimeUnit.MICROSECONDS)
+@State(Scope.Benchmark)
+class ReadWriteBenchmark:
+
+  private var testData: Array[Byte] = null
+  private val testDataFile = Path.of("src/main/resources/CodeGeneratorRequest-pack.bin")
+
+  @Setup(Level.Trial)
+  def buildTestData(): Unit =
+    testData = Files.readAllBytes(testDataFile)
+    assert(testData.length == 159366, s"testData.length ${testData.length}")
+
+  @Benchmark
+  def array_google(bh: Blackhole): Unit =
+    val p = GPluginProtos.CodeGeneratorRequest.parseFrom(testData)
+    val out = new ByteArrayOutputStream()
+    p.writeTo(out)
+    out.close()
+    bh.consume(out.size())
+
+  @Benchmark
+  def array_perfio(bh: Blackhole): Unit =
+    val p = PPluginProtos.CodeGeneratorRequest.parseFrom(BufferedInput.ofArray(testData).order(ByteOrder.LITTLE_ENDIAN))
+    val bo = BufferedOutput.growing(4096)
+    p.writeTo(bo)
+    bo.close()
+    bh.consume(bo.length())
+
+  @Benchmark
+  def file_google(bh: Blackhole): Unit =
+    val in = new BufferedInputStream(new FileInputStream(testDataFile.toFile))
+    val p = GPluginProtos.CodeGeneratorRequest.parseFrom(in)
+    in.close()
+    val out = new BufferedOutputStream(new FileOutputStream("/dev/null"))
+    p.writeTo(out)
+    out.close()
+
+  @Benchmark
+  def file_perfio(bh: Blackhole): Unit =
+    val in = BufferedInput.of(new FileInputStream(testDataFile.toFile)).order(ByteOrder.LITTLE_ENDIAN)
+    val p = PPluginProtos.CodeGeneratorRequest.parseFrom(in)
+    in.close()
+    val bo = BufferedOutput.ofFile(Path.of("/dev/null"), 8192)
+    p.writeTo(bo)
+    bo.close()
+
+  @Benchmark
+  def file_perfio_mapped(bh: Blackhole): Unit =
+    val in = BufferedInput.ofMappedFile(testDataFile).order(ByteOrder.LITTLE_ENDIAN)
+    val p = PPluginProtos.CodeGeneratorRequest.parseFrom(in)
+    in.close()
+    val bo = BufferedOutput.ofFile(Path.of("/dev/null"), 8192)
+    p.writeTo(bo)
+    bo.close()

proto-bench/src/main/scala/SmallWriteBenchmark.scala

@@ -0,0 +1,114 @@
+package perfio
+
+import com.google.protobuf.compiler.PluginProtos as GPluginProtos
+import com.google.protobuf.{CodedOutputStream, GeneratedMessage as GGeneratedMessage, Parser as GParser}
+import org.openjdk.jmh.annotations.*
+import org.openjdk.jmh.infra.*
+import perfio.proto.runtime.GeneratedMessage as PGeneratedMessage
+import perfio.protoapi.PluginProtos as PPluginProtos
+
+import java.io.{BufferedOutputStream, ByteArrayOutputStream, FileOutputStream, OutputStream}
+import java.lang.invoke.{MethodHandles, MethodType}
+import java.nio.file.Path
+import java.util.concurrent.TimeUnit
+
+@BenchmarkMode(Array(Mode.AverageTime))
+@Fork(value = 1, jvmArgsAppend = Array("-Xmx12g", "-Xss32M", "-XX:+UnlockExperimentalVMOptions", "-XX:+UseZGC", "--enable-native-access=ALL-UNNAMED", "--add-modules", "jdk.incubator.vector"))
+@Threads(1)
+@Warmup(iterations = 15, time = 1)
+@Measurement(iterations = 10, time = 1)
+@OutputTimeUnit(TimeUnit.MICROSECONDS)
+@State(Scope.Benchmark)
+class SmallWriteBenchmark:
+
+  @Param(Array("version1000"))
+  var testCase: String = null
+
+  private var gMessage: GGeneratedMessage = null
+  private var pMessage: PGeneratedMessage = null
+  private var count = 1;
+
+  private def p2g(p: PGeneratedMessage, gp: GParser[? <: GGeneratedMessage]): GGeneratedMessage = {
+    val out = BufferedOutput.growing()
+    p.writeTo(out)
+    out.close()
+    gp.parseFrom(out.toInputStream)
+  }
+
+  val clearMemoMethod = {
+    val lookup = MethodHandles.privateLookupIn(classOf[GGeneratedMessage], MethodHandles.lookup())
+    lookup.findVirtual(classOf[GGeneratedMessage], "setMemoizedSerializedSize", MethodType.methodType(Void.TYPE, Integer.TYPE))
+  }
+
+  private def clearMemo(m: GGeneratedMessage): Unit = {
+    clearMemoMethod.invokeExact(m, -1)
+  }
+
+  @Setup(Level.Trial)
+  def buildInvocationData(): Unit =
+    testCase match
+      case "version1000" =>
+        pMessage = (new PPluginProtos.Version).setMajor(1).setMinor(2).setPatch(3) //.setSuffix("foo")
+        gMessage = p2g(pMessage, GPluginProtos.Version.parser())
+        count = 1000
+      case "versionString1000" =>
+        pMessage = (new PPluginProtos.Version).setMajor(1).setMinor(2).setPatch(3).setSuffix("foo")
+        gMessage = p2g(pMessage, GPluginProtos.Version.parser())
+        count = 1000
+      case "versionString1" =>
+        pMessage = (new PPluginProtos.Version).setMajor(1).setMinor(2).setPatch(3).setSuffix("foo")
+        gMessage = p2g(pMessage, GPluginProtos.Version.parser())
+        count = 1
+      case "strings" =>
+        val r = new PPluginProtos.CodeGeneratorRequest
+        for(i <- 1 to 1000) r.addFileToGenerate(s"file $i")
+        pMessage = r
+        gMessage = p2g(pMessage, GPluginProtos.CodeGeneratorRequest.parser())
+        count = 10
+      case "nested" =>
+        val r = new PPluginProtos.CodeGeneratorRequest
+        r.setCompilerVersion((new PPluginProtos.Version).setMajor(1).setMinor(2).setPatch(3).setSuffix("foo"))
+        pMessage = r
+        gMessage = p2g(pMessage, GPluginProtos.CodeGeneratorRequest.parser())
+        count = 100
+
+  private def writeGoogle(out: OutputStream): Unit =
+    val c = CodedOutputStream.newInstance(out, 4096)
+    var i = 0
+    while i < count do
+      //clearMemo(gMessage.asInstanceOf[GPluginProtos.CodeGeneratorRequest].getCompilerVersion)
+      gMessage.writeTo(c)
+      i += 1
+    c.flush()
+    out.close()
+
+  private def writePerfio(out: BufferedOutput): Unit =
+    var i = 0
+    while i < count do
+      pMessage.writeTo(out)
+      i += 1
+    out.close()
+
+  @Benchmark
+  def array_google(bh: Blackhole): Unit =
+    val baos = new ByteArrayOutputStream()
+    writeGoogle(baos)
+    bh.consume(baos.size())
+
+  @Benchmark
+  def array_perfio(bh: Blackhole): Unit =
+    val bo = BufferedOutput.growing(4096)
+    writePerfio(bo)
+    bh.consume(bo.length())
+
+  @Benchmark
+  def file_google(bh: Blackhole): Unit =
+    val out = new BufferedOutputStream(new FileOutputStream("/dev/null"))
+    writeGoogle(out)
+    out.close()
+
+  @Benchmark
+  def file_perfio(bh: Blackhole): Unit =
+    val bo = BufferedOutput.ofFile(Path.of("/dev/null"), 8192)
+    writePerfio(bo)
+    bo.close()

proto-bench/src/main/scala/WriteBenchmark.scala

@@ -1,59 +1,91 @@
 package perfio
 
-import org.openjdk.jmh.annotations._
-import org.openjdk.jmh.infra._
+import com.google.protobuf.{CodedOutputStream, GeneratedMessage as GGeneratedMessage, Parser as GParser}
+import perfio.proto.runtime.GeneratedMessage as PGeneratedMessage
+import org.openjdk.jmh.annotations.*
+import org.openjdk.jmh.infra.*
 
 import java.util.concurrent.TimeUnit
-import com.google.protobuf.compiler.{PluginProtos => GPluginProtos}
-import perfio.protoapi.{PluginProtos => PPluginProtos}
+import com.google.protobuf.compiler.PluginProtos as GPluginProtos
+import perfio.protoapi.PluginProtos as PPluginProtos
 
-import java.io.{BufferedOutputStream, ByteArrayOutputStream, FileOutputStream}
+import java.io.{BufferedOutputStream, ByteArrayOutputStream, FileOutputStream, OutputStream}
+import java.lang.invoke.{MethodHandles, MethodType}
 import java.nio.ByteOrder
 import java.nio.file.{Files, Path}
 
+import scala.jdk.CollectionConverters._
+
 @BenchmarkMode(Array(Mode.AverageTime))
 @Fork(value = 1, jvmArgsAppend = Array("-Xmx12g", "-Xss32M", "-XX:+UnlockExperimentalVMOptions", "-XX:+UseZGC", "--enable-native-access=ALL-UNNAMED", "--add-modules", "jdk.incubator.vector"))
 @Threads(1)
-@Warmup(iterations = 7, time = 1)
-@Measurement(iterations = 7, time = 1)
+@Warmup(iterations = 15, time = 1)
+@Measurement(iterations = 10, time = 1)
 @OutputTimeUnit(TimeUnit.MICROSECONDS)
 @State(Scope.Benchmark)
 class WriteBenchmark:
 
-  private var testData: Array[Byte] = null
   private val testDataFile = Path.of("src/main/resources/CodeGeneratorRequest-pack.bin")
-  private var gReq: GPluginProtos.CodeGeneratorRequest = null
-  private var pReq: PPluginProtos.CodeGeneratorRequest = null
+  private var testData: Array[Byte] = null
+  private var gMessage: GGeneratedMessage = null
+  private var pMessage: PGeneratedMessage = null
+
+  private def p2g(p: PGeneratedMessage, gp: GParser[? <: GGeneratedMessage]): GGeneratedMessage = {
+    val out = BufferedOutput.growing()
+    p.writeTo(out)
+    out.close()
+    gp.parseFrom(out.toInputStream)
+  }
 
   @Setup(Level.Trial)
-  def buildTestData(): Unit =
+  def buildTrialData(): Unit =
     testData = Files.readAllBytes(testDataFile)
     assert(testData.length == 159366, s"testData.length ${testData.length}")
-    gReq = GPluginProtos.CodeGeneratorRequest.parseFrom(testData)
-    pReq = PPluginProtos.CodeGeneratorRequest.parseFrom(BufferedInput.ofArray(testData).order(ByteOrder.LITTLE_ENDIAN))
+    val m = PPluginProtos.CodeGeneratorRequest.parseFrom(BufferedInput.ofArray(testData).order(ByteOrder.LITTLE_ENDIAN))
+    val pfs = m.getProtoFileList.asScala.toArray
+    for i <- 1 to 10 do
+      pfs.foreach(m.addProtoFile)
+    val out = BufferedOutput.growing()
+    m.writeTo(out)
+    out.close()
+    testData = out.copyToByteArray()
+
+  @Setup(Level.Invocation)
+  def buildInvocationData(): Unit =
+    gMessage = GPluginProtos.CodeGeneratorRequest.parseFrom(testData)
+    pMessage = PPluginProtos.CodeGeneratorRequest.parseFrom(BufferedInput.ofArray(testData).order(ByteOrder.LITTLE_ENDIAN))
+
+  private def writeGoogle(out: OutputStream): Unit =
+    val c = CodedOutputStream.newInstance(out, 4096)
+    gMessage.writeTo(c)
+    c.flush()
+    out.close()
+
+  private def writePerfio(out: BufferedOutput): Unit =
+    var i = 0
+    pMessage.writeTo(out)
+    out.close()
 
   @Benchmark
   def array_google(bh: Blackhole): Unit =
     val baos = new ByteArrayOutputStream()
-    gReq.writeTo(baos)
-    baos.close()
+    writeGoogle(baos)
     bh.consume(baos.size())
 
   @Benchmark
   def array_perfio(bh: Blackhole): Unit =
     val bo = BufferedOutput.growing(4096)
-    pReq.writeTo(bo)
-    bo.close()
+    writePerfio(bo)
     bh.consume(bo.length())
 
   @Benchmark
   def file_google(bh: Blackhole): Unit =
     val out = new BufferedOutputStream(new FileOutputStream("/dev/null"))
-    gReq.writeTo(out)
+    writeGoogle(out)
     out.close()
 
   @Benchmark
   def file_perfio(bh: Blackhole): Unit =
     val bo = BufferedOutput.ofFile(Path.of("/dev/null"), 8192)
-    pReq.writeTo(bo)
+    writePerfio(bo)
     bo.close()

proto-runtime/src/main/java/perfio/proto/runtime/GeneratedMessage.java

@@ -0,0 +1,9 @@
+package perfio.proto.runtime;
+
+import perfio.BufferedOutput;
+
+import java.io.IOException;
+
+public abstract class GeneratedMessage {
+  public abstract void writeTo(BufferedOutput out) throws IOException;
+}