anjy7 commented on code in PR #22669: URL: https://github.com/apache/kafka/pull/22669#discussion_r3496284459
########## jmh-benchmarks/src/main/java/org/apache/kafka/jmh/raft/KRaftBenchmarkingCounters.java: ########## @@ -0,0 +1,151 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.kafka.jmh.raft; + +import org.apache.kafka.common.protocol.ApiKeys; +import org.apache.kafka.raft.RaftClientBenchmarkContext; + +import org.openjdk.jmh.annotations.AuxCounters; +import org.openjdk.jmh.annotations.Level; +import org.openjdk.jmh.annotations.Scope; +import org.openjdk.jmh.annotations.Setup; +import org.openjdk.jmh.annotations.State; +import org.openjdk.jmh.infra.BenchmarkParams; + +import java.util.Optional; + +/** + * Secondary, machine-independent work counters reported by the raft benchmarks alongside the timing + * score, as {@code benchmark:counter} rows. + * + * <p>Throughout this class, an <em>operation</em> is JMH's unit of work: a single invocation of a + * {@code @Benchmark}-annotated method. (One operation equals one invocation here because we don't use + * {@code @OperationsPerInvocation}.) JMH reports the timing score in {@code ns/op}, and these work + * counters are reported {@code PerOp} to match. + * + * <p>Each benchmark calls {@link #drainFrom} every invocation to accumulate the work deltas drained + * from {@link RaftClientBenchmarkContext}. The raw totals are private accumulators; what we report + * are the per-operation values from the {@code *PerOp()} methods (the quantity of interest), plus + * {@link #operations}. + * + * <p>JMH aggregates {@code Type.EVENTS} secondary results with {@code SUM} across all measurement + * data points i.e {@code forks x measurement iterations}. To make the <em>summary</em> row + * report the true per-operation value rather than that value multiplied by the data-point count, each + * method pre-divides by the data-point count obtained from {@link BenchmarkParams} in + * {@link #captureRunShape}. The SUM then reconstitutes the exact per-operation value (e.g. + * {@code logReadsPerOp = 1.0}) in the summary, for any {@code -f}/{@code -i} configuration. (The + * per-iteration console values are correspondingly a small fraction of the per-op value; read the + * summary row.) + * + * <p>The per-operation values are integer-exact and should be stable across a correct refactor of + * {@code KafkaRaftClient}: a flush count moving from 1 to 2 per operation is a behavioral diff, not + * measurement noise. The counters that are zero on a path (e.g. log flushes on a caught-up fetch) + * are the most useful tripwires, since zero is speed-independent. + */ +@State(Scope.Thread) +@AuxCounters(AuxCounters.Type.EVENTS) +public class KRaftBenchmarkingCounters { + private long logFlushesTotal; + private long logReadsTotal; + private long logTruncationsTotal; + private long rpcRequestsSentTotal; + private long rpcResponsesSentTotal; + private long quorumStateWritesTotal; + private long quorumStateReadsTotal; + + // Reported: the number of operations (i.e. @Benchmark method invocations) measured in the + // iteration, and the divisor for the per-operation values below. + public long operations; + + // The number of measurement data points JMH will SUM the per-op methods over, i.e. + // (forks x measurement iterations) for this run. Captured from BenchmarkParams so it tracks the + // actual run shape (including -f/-i overrides) rather than being hardcoded. + private double measurementDataPoints = 1.0; + + @Setup(Level.Trial) + public void captureRunShape(BenchmarkParams params) { + // forks() is 0 when forking is disabled (in-process), which is still one set of iterations. + int forks = Math.max(1, params.getForks()); + measurementDataPoints = (double) forks * params.getMeasurement().getCount(); + } Review Comment: Different forks run their own iterations and the counter values are serialized and given back to the parent jmh process, which then sums all of them up. And yes, state is per thread but our benchmarks run on a single thread (@Threads(1) default) and there shouldn't be multithreaded benchmarks as discussed in the design proposal. To explain this snippet, I'll take an example to show how its used in the `perOperation()` function: Lets say, the leaderBenchmark (AverageTime fork=5 iterations=30), runs 100,000 times in 1 iteration and we get 100,000 log reads. To calculate per operation, we do counter / operations, which gives us 1 log read per op. Now, this across 5 forks with 30 iterations will give use 150 log reads total in the output (jmh sums them all). Hence to get the actual per op number in the output, we divide it by measurementDataPoints ( can be seen in `perOperation` function). Which means, 1.0 / 30 = 0.0333… This single iteration now reports 0.0333, not 1.0. JMH collects one such reading per iteration per fork = 30 readings, and adds them up: 0.0333 × 30 = 1.0 (the true per op value) And params.getMeasurement() returns the measurement-phase config (an IterationParams), and .getCount() is the number of measurement iterations ( excluding warmup ) And `Math.max(1, params.getForks())` is to avoid garbage values in the output incase anyone uses fork=0 for debugging purposes (because of division by zero in `perOperation` function). Another thing to note is that the default fork value is 5 in JMH and JMH even warns fork=0 is for debugging only, and results from it are unreliable as it forces the benchmark to run inside the same JVM as JMH's own machinery. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
