Rework and publish metric benchmarks

.github/workflows/benchmark.yml

@@ -42,10 +42,15 @@ jobs:
         env:
           DEVELOCITY_ACCESS_KEY: ${{ secrets.DEVELOCITY_ACCESS_KEY }}
 
+      # TODO (jack-berg): Select or build appropriate benchmarks for other key areas:
+      # - Log SDK record & export
+      # - Trace SDK record & export
+      # - Metric SDK export
+      # - Noop implementation
       - name: Run Benchmark
         run: |
-          cd sdk/trace/build
-          java -jar libs/opentelemetry-sdk-trace-*-jmh.jar -rf json SpanBenchmark SpanPipelineBenchmark ExporterBenchmark
+          cd sdk/all/build
+          java -jar libs/opentelemetry-sdk-*-jmh.jar -rf json MetricRecordBenchmark
 
       - name: Use CLA approved github bot
         run: .github/scripts/use-cla-approved-bot.sh
@@ -54,7 +59,7 @@ jobs:
         uses: benchmark-action/github-action-benchmark@4bdcce38c94cec68da58d012ac24b7b1155efe8b # v1.20.7
         with:
           tool: 'jmh'
-          output-file-path: sdk/trace/build/jmh-result.json
+          output-file-path: sdk/all/build/jmh-result.json
           gh-pages-branch: benchmarks
           github-token: ${{ secrets.GITHUB_TOKEN }}
           benchmark-data-dir-path: "benchmarks"

sdk/all/build.gradle.kts

@@ -22,4 +22,6 @@ dependencies {
   testAnnotationProcessor("com.google.auto.value:auto-value")
 
   testImplementation(project(":sdk:testing"))
+
+  jmh(project(":sdk:testing"))
 }

sdk/all/src/jmh/java/io/opentelemetry/sdk/MetricRecordBenchmark.java

@@ -0,0 +1,246 @@
+/*
+ * Copyright The OpenTelemetry Authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+package io.opentelemetry.sdk;
+
+import static io.opentelemetry.sdk.metrics.InstrumentType.COUNTER;
+import static io.opentelemetry.sdk.metrics.InstrumentType.GAUGE;
+import static io.opentelemetry.sdk.metrics.InstrumentType.HISTOGRAM;
+import static io.opentelemetry.sdk.metrics.InstrumentType.UP_DOWN_COUNTER;
+
+import io.opentelemetry.api.common.AttributeKey;
+import io.opentelemetry.api.common.Attributes;
+import io.opentelemetry.api.metrics.Meter;
+import io.opentelemetry.api.trace.Span;
+import io.opentelemetry.api.trace.Tracer;
+import io.opentelemetry.sdk.common.export.MemoryMode;
+import io.opentelemetry.sdk.metrics.Aggregation;
+import io.opentelemetry.sdk.metrics.ExemplarFilter;
+import io.opentelemetry.sdk.metrics.InstrumentType;
+import io.opentelemetry.sdk.metrics.InstrumentValueType;
+import io.opentelemetry.sdk.metrics.SdkMeterProvider;
+import io.opentelemetry.sdk.metrics.data.AggregationTemporality;
+import io.opentelemetry.sdk.metrics.export.DefaultAggregationSelector;
+import io.opentelemetry.sdk.testing.exporter.InMemoryMetricReader;
+import io.opentelemetry.sdk.trace.SdkTracerProvider;
+import io.opentelemetry.sdk.trace.samplers.Sampler;
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.List;
+import java.util.Random;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Group;
+import org.openjdk.jmh.annotations.GroupThreads;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Param;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.TearDown;
+import org.openjdk.jmh.annotations.Warmup;
+
+/**
+ * Notes on interpreting the data:
+ *
+ * <p>The benchmark has two dimensions which partially overlap: cardinality and thread count.
+ * Cardinality dictates how many unique attribute sets (i.e. series) are recorded to, and thread
+ * count dictates how many threads are simultaneously recording to those series. In all cases, the
+ * record path needs to look up an aggregation handle for the series corresponding to the
+ * measurement's {@link Attributes} in a {@link java.util.concurrent.ConcurrentHashMap}. That will
+ * be the case until otel adds support for <a
+ * href="https://github.com/open-telemetry/opentelemetry-specification/issues/4126">bound
+ * instruments</a>. The cardinality dictates the size of this map, which has some impact on
+ * performance. However, by far the dominant bottleneck is contention. That is, the number of
+ * threads simultaneously trying to record to the same series. Increasing the threads increases
+ * contention. Increasing cardinality decreases contention, as the threads are now spreading their
+ * record activities over more distinct series. The highest contention scenario is cardinality=1,
+ * threads=4. Any scenario with threads=1 has zero contention.
+ *
+ * <p>It's useful to characterize the performance of the metrics system under contention, as some
+ * high-performance applications may have many threads trying to record to the same series. It's
+ * also useful to characterize the performance of the metrics system under low contention, as some
+ * high-performance applications may not frequently be trying to concurrently record to the same
+ * series yet still care about the overhead of each record operation.
+ *
+ * <p>{@link AggregationTemporality} can impact performance because additional concurrency controls
+ * are needed to ensure there are no duplicate, partial, or lost writes while resetting the set of
+ * timeseries each collection.
+ */
+public class MetricRecordBenchmark {
+
+  private static final int recordCount = 10 * 1024;
+
+  @State(Scope.Benchmark)
+  public static class ThreadState {
+
+    @Param InstrumentTypeAndAggregation instrumentTypeAndAggregation;
+
+    @Param AggregationTemporality aggregationTemporality;
+
+    @Param({"1", "100"})
+    int cardinality;
+
+    // The following parameters are excluded from the benchmark to reduce combinatorial explosion
+    // but can optionally be enabled for adhoc evaluation.
+
+    // InstrumentValueType doesn't materially impact performance. Uncomment to evaluate.
+    // @Param
+    // InstrumentValueType instrumentValueType;
+    InstrumentValueType instrumentValueType = InstrumentValueType.LONG;
+
+    // MemoryMode almost exclusively impacts collect from a performance standpoint. Uncomment to
+    // evaluate.
+    // @Param
+    // MemoryMode memoryMode;
+    MemoryMode memoryMode = MemoryMode.REUSABLE_DATA;
+
+    // Exemplars can impact performance, but we skip evaluation to limit test cases. Uncomment to
+    // evaluate.
+    // @Param({"true", "false"})
+    // boolean exemplars;
+    boolean exemplars = false;
+
+    OpenTelemetrySdk openTelemetry;
+    Instrument instrument;
+    List<Long> measurements;
+    List<Attributes> attributesList;
+    Span span;
+    io.opentelemetry.context.Scope contextScope;
+
+    @Setup
+    @SuppressWarnings("MustBeClosedChecker")
+    public void setup() {
+      InstrumentType instrumentType = instrumentTypeAndAggregation.instrumentType;
+      Aggregation aggregation = instrumentTypeAndAggregation.aggregation;
+
+      openTelemetry =
+          OpenTelemetrySdk.builder()
+              .setTracerProvider(SdkTracerProvider.builder().setSampler(Sampler.alwaysOn()).build())
+              .setMeterProvider(
+                  SdkMeterProvider.builder()
+                      .registerMetricReader(
+                          InMemoryMetricReader.builder()
+                              .setAggregationTemporalitySelector(unused -> aggregationTemporality)
+                              .setDefaultAggregationSelector(
+                                  DefaultAggregationSelector.getDefault()
+                                      .with(instrumentType, aggregation))
+                              .setMemoryMode(memoryMode)
+                              .build())
+                      .setExemplarFilter(
+                          exemplars ? ExemplarFilter.traceBased() : ExemplarFilter.alwaysOff())
+                      .build())
+              .build();
+
+      Meter meter = openTelemetry.getMeter("benchmark");
+      instrument = getInstrument(meter, instrumentType, instrumentValueType);
+      Tracer tracer = openTelemetry.getTracer("benchmark");
+      span = tracer.spanBuilder("benchmark").startSpan();
+      // We suppress warnings on closing here, as we rely on tests to make sure context is closed.
+      contextScope = span.makeCurrent();
+
+      Random random = new Random();
+      attributesList = new ArrayList<>(cardinality);
+      AttributeKey<String> key = AttributeKey.stringKey("key");
+      String last = "aaaaaaaaaaaaaaaaaaaaaaaaaa";
+      for (int i = 0; i < cardinality; i++) {
+        char[] chars = last.toCharArray();
+        chars[random.nextInt(last.length())] = (char) (random.nextInt(26) + 'a');
+        last = new String(chars);
+        attributesList.add(Attributes.of(key, last));
+      }
+      Collections.shuffle(attributesList);
+
+      measurements = new ArrayList<>(recordCount);
+      for (int i = 0; i < recordCount; i++) {
+        measurements.add((long) random.nextInt(2000));
+      }
+      Collections.shuffle(measurements);
+    }
+
+    @TearDown
+    public void tearDown() {
+      contextScope.close();
+      span.end();
+      openTelemetry.shutdown();
+    }
+  }
+
+  @Benchmark
+  @Group("threads1")
+  @GroupThreads(1)
+  @Fork(1)
+  @Warmup(iterations = 5, time = 1)
+  @Measurement(iterations = 5, time = 1)
+  public void record_1Thread(ThreadState threadState) {
+    record(threadState);
+  }
+
+  @Benchmark
+  @Group("threads4")
+  @GroupThreads(4)
+  @Fork(1)
+  @Warmup(iterations = 5, time = 1)
+  @Measurement(iterations = 5, time = 1)
+  public void record_4Threads(ThreadState threadState) {
+    record(threadState);
+  }
+
+  private static void record(ThreadState threadState) {
+    for (int i = 0; i < recordCount; i++) {
+      Attributes attributes = threadState.attributesList.get(i % threadState.attributesList.size());
+      long value = threadState.measurements.get(i % threadState.measurements.size());
+      threadState.instrument.record(value, attributes);
+    }
+  }
+
+  @SuppressWarnings("ImmutableEnumChecker")
+  public enum InstrumentTypeAndAggregation {
+    COUNTER_SUM(COUNTER, Aggregation.sum()),
+    UP_DOWN_COUNTER_SUM(UP_DOWN_COUNTER, Aggregation.sum()),
+    GAUGE_LAST_VALUE(GAUGE, Aggregation.lastValue()),
+    HISTOGRAM_EXPLICIT(HISTOGRAM, Aggregation.explicitBucketHistogram()),
+    HISTOGRAM_BASE2_EXPONENTIAL(HISTOGRAM, Aggregation.base2ExponentialBucketHistogram());
+
+    InstrumentTypeAndAggregation(InstrumentType instrumentType, Aggregation aggregation) {
+      this.instrumentType = instrumentType;
+      this.aggregation = aggregation;
+    }
+
+    private final InstrumentType instrumentType;
+    private final Aggregation aggregation;
+  }
+
+  private interface Instrument {
+    void record(long value, Attributes attributes);
+  }
+
+  private static Instrument getInstrument(
+      Meter meter, InstrumentType instrumentType, InstrumentValueType instrumentValueType) {
+    String name = "instrument";
+    switch (instrumentType) {
+      case COUNTER:
+        return instrumentValueType == InstrumentValueType.DOUBLE
+            ? meter.counterBuilder(name).ofDoubles().build()::add
+            : meter.counterBuilder(name).build()::add;
+      case UP_DOWN_COUNTER:
+        return instrumentValueType == InstrumentValueType.DOUBLE
+            ? meter.upDownCounterBuilder(name).ofDoubles().build()::add
+            : meter.upDownCounterBuilder(name).build()::add;
+      case HISTOGRAM:
+        return instrumentValueType == InstrumentValueType.DOUBLE
+            ? meter.histogramBuilder(name).build()::record
+            : meter.histogramBuilder(name).ofLongs().build()::record;
+      case GAUGE:
+        return instrumentValueType == InstrumentValueType.DOUBLE
+            ? meter.gaugeBuilder(name).build()::set
+            : meter.gaugeBuilder(name).ofLongs().build()::set;
+      case OBSERVABLE_COUNTER:
+      case OBSERVABLE_UP_DOWN_COUNTER:
+      case OBSERVABLE_GAUGE:
+    }
+    throw new IllegalArgumentException();
+  }
+}