Add virtual-threads-benchmark-webflux: the WebFlux leg of the three-way benchmark
Re-uses the platform/virtual-thread transcripts already committed in virtual-threads-benchmark rather than re-measuring them, and adds a WebFlux module measured the same way (client-side java.net.http.HttpClient load generator against a real Netty server). Covers I/O-bound throughput, a naive-vs-offloaded CPU-bound comparison that turned up a real, counter- intuitive fact about this sandbox's default thread-pool sizes (LoopResources.DEFAULT_IO_WORKER_COUNT=4 vs Schedulers.parallel()'s availableProcessors()=2), and an event-loop-starvation scenario showing what the naive endpoint actually costs unrelated /io traffic on the same server. Companion code for the ankurm.com post "Virtual Threads vs WebFlux vs Platform Threads: Benchmarks and a Decision Framework." Also fixes the root README's module table, which was missing the virtual-threads-benchmark row from an earlier commit.
This commit is contained in:
Claude
+29
@@ -0,0 +1,29 @@
|
||||
package com.ankurm.vthreadswebflux;
|
||||
|
||||
import java.security.MessageDigest;
|
||||
import java.security.NoSuchAlgorithmException;
|
||||
|
||||
/**
|
||||
* The exact same 20,000x SHA-256 loop as {@code DemoController.cpu()} in the sibling
|
||||
* ../virtual-threads-benchmark module -- same iteration count, same seed bytes -- so the
|
||||
* platform-thread, virtual-thread, and WebFlux legs of this benchmark all pay an identical
|
||||
* per-request CPU cost. Only the thread model serving the request differs.
|
||||
*/
|
||||
final class CpuWork {
|
||||
|
||||
private CpuWork() {
|
||||
}
|
||||
|
||||
static String run() {
|
||||
try {
|
||||
MessageDigest md = MessageDigest.getInstance("SHA-256");
|
||||
byte[] data = "spring-boot-4.1-virtual-threads-benchmark".getBytes();
|
||||
for (int i = 0; i < 20_000; i++) {
|
||||
data = md.digest(data);
|
||||
}
|
||||
return "cpu:" + Thread.currentThread() + ":" + data.length;
|
||||
} catch (NoSuchAlgorithmException e) {
|
||||
throw new IllegalStateException(e);
|
||||
}
|
||||
}
|
||||
}
|
||||
+53
@@ -0,0 +1,53 @@
|
||||
package com.ankurm.vthreadswebflux;
|
||||
|
||||
import org.springframework.web.bind.annotation.GetMapping;
|
||||
import org.springframework.web.bind.annotation.RestController;
|
||||
import reactor.core.publisher.Mono;
|
||||
import reactor.core.scheduler.Schedulers;
|
||||
|
||||
import java.time.Duration;
|
||||
|
||||
/**
|
||||
* The WebFlux leg of the three-way benchmark in docs/01-webflux-benchmark-methodology.md.
|
||||
* Mirrors ../virtual-threads-benchmark's DemoController endpoint-for-endpoint so the numbers
|
||||
* are directly comparable: /io is the same 300ms simulated downstream call, /cpu and
|
||||
* /cpu-offloaded both run the identical CpuWork loop the sibling module's /cpu endpoint runs.
|
||||
*
|
||||
* /cpu and /cpu-offloaded exist as a pair on purpose. /cpu wraps blocking CPU work in
|
||||
* Mono.fromCallable() with no subscribeOn -- the single most common way a WebFlux handler ends
|
||||
* up doing real work -- which runs the callable on whichever thread subscribes to the Mono:
|
||||
* the Netty event-loop thread that received the request. /cpu-offloaded moves the same work to
|
||||
* Schedulers.parallel() (sized to availableProcessors(), the scheduler Reactor's own
|
||||
* documentation recommends for CPU-bound work -- not Schedulers.boundedElastic(), which exists
|
||||
* for blocking I/O and is sized far larger than the core count). See
|
||||
* docs/01-webflux-benchmark-methodology.md for what each does to the event loop under load.
|
||||
*/
|
||||
@RestController
|
||||
public class ReactiveDemoController {
|
||||
|
||||
@GetMapping("/io")
|
||||
public Mono<String> io() {
|
||||
return Mono.delay(Duration.ofMillis(300))
|
||||
.map(tick -> "io:" + Thread.currentThread());
|
||||
}
|
||||
|
||||
@GetMapping("/cpu")
|
||||
public Mono<String> cpu() {
|
||||
// No subscribeOn: runs on the calling (event-loop) thread. This is the naive version.
|
||||
return Mono.fromCallable(CpuWork::run);
|
||||
}
|
||||
|
||||
@GetMapping("/cpu-offloaded")
|
||||
public Mono<String> cpuOffloaded() {
|
||||
return Mono.fromCallable(CpuWork::run)
|
||||
.subscribeOn(Schedulers.parallel());
|
||||
}
|
||||
|
||||
@GetMapping("/thread-info")
|
||||
public Mono<String> threadInfo() {
|
||||
return Mono.fromSupplier(() -> {
|
||||
Thread t = Thread.currentThread();
|
||||
return "Thread: " + t + " | Virtual: " + t.isVirtual();
|
||||
});
|
||||
}
|
||||
}
|
||||
+12
@@ -0,0 +1,12 @@
|
||||
package com.ankurm.vthreadswebflux;
|
||||
|
||||
import org.springframework.boot.SpringApplication;
|
||||
import org.springframework.boot.autoconfigure.SpringBootApplication;
|
||||
|
||||
@SpringBootApplication
|
||||
public class VirtualThreadsWebfluxBenchmarkApplication {
|
||||
|
||||
public static void main(String[] args) {
|
||||
SpringApplication.run(VirtualThreadsWebfluxBenchmarkApplication.class, args);
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,16 @@
|
||||
spring:
|
||||
application:
|
||||
name: virtual-threads-webflux-benchmark
|
||||
|
||||
server:
|
||||
port: 8080
|
||||
|
||||
management:
|
||||
endpoints:
|
||||
web:
|
||||
exposure:
|
||||
include: health,metrics
|
||||
|
||||
logging:
|
||||
level:
|
||||
com.ankurm.vthreadswebflux: INFO
|
||||
+50
@@ -0,0 +1,50 @@
|
||||
package com.ankurm.vthreadswebflux;
|
||||
|
||||
import java.io.IOException;
|
||||
import java.io.UncheckedIOException;
|
||||
import java.nio.file.Files;
|
||||
import java.nio.file.Path;
|
||||
import java.nio.file.Paths;
|
||||
import java.nio.file.StandardOpenOption;
|
||||
|
||||
/**
|
||||
* Writes docs/output/NN-*.txt while a test runs, so every number quoted in the blog post
|
||||
* is backed by a file produced by an assertion that would fail the build if it stopped
|
||||
* being true. Never hand-edit files under docs/output/ — regenerate with scripts/run-all.sh.
|
||||
*/
|
||||
public final class Transcript {
|
||||
|
||||
private final StringBuilder buf = new StringBuilder();
|
||||
private final Path outFile;
|
||||
|
||||
private Transcript(String fileName) {
|
||||
this.outFile = Paths.get("docs/output", fileName);
|
||||
}
|
||||
|
||||
public static Transcript start(String fileName, String header) {
|
||||
Transcript t = new Transcript(fileName);
|
||||
t.line(header);
|
||||
t.line("=".repeat(header.length()));
|
||||
return t;
|
||||
}
|
||||
|
||||
public Transcript line(String s) {
|
||||
buf.append(s).append('\n');
|
||||
return this;
|
||||
}
|
||||
|
||||
public Transcript blank() {
|
||||
buf.append('\n');
|
||||
return this;
|
||||
}
|
||||
|
||||
public void save() {
|
||||
try {
|
||||
Files.createDirectories(outFile.getParent());
|
||||
Files.writeString(outFile, buf.toString(), StandardOpenOption.CREATE,
|
||||
StandardOpenOption.TRUNCATE_EXISTING);
|
||||
} catch (IOException e) {
|
||||
throw new UncheckedIOException(e);
|
||||
}
|
||||
}
|
||||
}
|
||||
+248
@@ -0,0 +1,248 @@
|
||||
package com.ankurm.vthreadswebflux;
|
||||
|
||||
import org.junit.jupiter.api.Test;
|
||||
import org.springframework.boot.builder.SpringApplicationBuilder;
|
||||
import org.springframework.boot.web.server.context.WebServerInitializedEvent;
|
||||
import org.springframework.context.ApplicationListener;
|
||||
import org.springframework.context.ConfigurableApplicationContext;
|
||||
import reactor.core.scheduler.Schedulers;
|
||||
import reactor.netty.resources.LoopResources;
|
||||
|
||||
import java.net.URI;
|
||||
import java.net.http.HttpClient;
|
||||
import java.net.http.HttpRequest;
|
||||
import java.net.http.HttpResponse;
|
||||
import java.util.ArrayList;
|
||||
import java.util.Collections;
|
||||
import java.util.List;
|
||||
import java.util.concurrent.CountDownLatch;
|
||||
import java.util.concurrent.Executors;
|
||||
import java.util.concurrent.TimeUnit;
|
||||
import java.util.concurrent.atomic.AtomicInteger;
|
||||
|
||||
import static org.assertj.core.api.Assertions.assertThat;
|
||||
|
||||
/**
|
||||
* The WebFlux leg of the three-way benchmark. Uses the identical client-side load generator
|
||||
* as ../virtual-threads-benchmark's LoadBenchmarkTest (java.net.http.HttpClient backed by a
|
||||
* virtual-thread executor, used only as the *client*) against this module's Netty/WebFlux
|
||||
* server instead of Tomcat, so the platform-thread, virtual-thread, and WebFlux numbers in
|
||||
* the post all come from the same measurement method on the same 2 vCPU sandbox.
|
||||
*
|
||||
* Output: docs/output/01-io-bound-webflux.txt, docs/output/02-cpu-bound-webflux.txt.
|
||||
* See docs/01-webflux-benchmark-methodology.md.
|
||||
*/
|
||||
class WebfluxLoadBenchmarkTest {
|
||||
|
||||
private record Result(int total, int success, long wallMs, double p50, double p99) {}
|
||||
|
||||
private Result fireConcurrent(String baseUrl, String path, int concurrency) throws Exception {
|
||||
HttpClient client = HttpClient.newBuilder()
|
||||
.executor(Executors.newVirtualThreadPerTaskExecutor())
|
||||
.build();
|
||||
List<Long> latencies = Collections.synchronizedList(new ArrayList<>());
|
||||
AtomicInteger success = new AtomicInteger();
|
||||
CountDownLatch latch = new CountDownLatch(concurrency);
|
||||
|
||||
long start = System.nanoTime();
|
||||
for (int i = 0; i < concurrency; i++) {
|
||||
Thread.ofVirtual().start(() -> {
|
||||
long reqStart = System.nanoTime();
|
||||
try {
|
||||
HttpRequest req = HttpRequest.newBuilder(URI.create(baseUrl + path)).build();
|
||||
HttpResponse<String> resp = client.send(req, HttpResponse.BodyHandlers.ofString());
|
||||
if (resp.statusCode() == 200) {
|
||||
success.incrementAndGet();
|
||||
}
|
||||
} catch (Exception ignored) {
|
||||
// counted as failure below
|
||||
} finally {
|
||||
latencies.add((System.nanoTime() - reqStart) / 1_000_000);
|
||||
latch.countDown();
|
||||
}
|
||||
});
|
||||
}
|
||||
latch.await(60, TimeUnit.SECONDS);
|
||||
long wallMs = (System.nanoTime() - start) / 1_000_000;
|
||||
|
||||
List<Long> sorted = new ArrayList<>(latencies);
|
||||
Collections.sort(sorted);
|
||||
double p50 = sorted.isEmpty() ? 0 : sorted.get(sorted.size() / 2);
|
||||
int p99Idx = sorted.isEmpty() ? 0 : Math.min(sorted.size() - 1, (int) (sorted.size() * 0.99));
|
||||
double p99 = sorted.isEmpty() ? 0 : sorted.get(p99Idx);
|
||||
|
||||
return new Result(concurrency, success.get(), wallMs, p50, p99);
|
||||
}
|
||||
|
||||
@Test
|
||||
void ioBoundScenario() throws Exception {
|
||||
int concurrency = 600; // same concurrency as the platform/virtual-thread /io benchmark
|
||||
|
||||
String baseUrl = startApp();
|
||||
ConfigurableApplicationContext ctx = currentCtx;
|
||||
try {
|
||||
fireConcurrent(baseUrl, "/io", Math.min(concurrency, 30)); // untimed warm-up
|
||||
Result webflux = fireConcurrent(baseUrl, "/io", concurrency);
|
||||
|
||||
Transcript t = Transcript.start("01-io-bound-webflux.txt",
|
||||
"I/O-bound endpoint (/io, Mono.delay(300ms)), concurrency=" + concurrency
|
||||
+ ", 2 vCPU sandbox, Spring Boot 4.1.1 / JDK " + System.getProperty("java.version"));
|
||||
t.line(String.format("webflux (netty) : total=%d success=%d wall=%dms p50=%.0fms p99=%.0fms",
|
||||
webflux.total, webflux.success, webflux.wallMs, webflux.p50, webflux.p99));
|
||||
t.blank();
|
||||
t.line("Mono.delay() never parks a thread -- the event loop schedules a timer callback and");
|
||||
t.line("goes back to the selector loop immediately. Wall time for 600 concurrent requests");
|
||||
t.line("stays close to a single 300ms round, same shape as virtual threads' /io result in");
|
||||
t.line("../virtual-threads-benchmark/docs/output/01-io-bound-benchmark.txt.");
|
||||
t.save();
|
||||
|
||||
assertThat(webflux.success).isEqualTo(concurrency);
|
||||
} finally {
|
||||
ctx.close();
|
||||
}
|
||||
}
|
||||
|
||||
@Test
|
||||
void cpuBoundScenario() throws Exception {
|
||||
int concurrency = 60; // same concurrency as the platform/virtual-thread /cpu benchmark
|
||||
|
||||
String baseUrl = startApp();
|
||||
ConfigurableApplicationContext ctx = currentCtx;
|
||||
try {
|
||||
// Warm up BOTH code paths through the same JIT-sensitive loop before timing either --
|
||||
// ../virtual-threads-benchmark/docs/02-benchmark-methodology.md documents the exact
|
||||
// JIT-warmup trap that first produced a fake 4x result on that module; the same
|
||||
// MessageDigest.digest hot loop is reused here unmodified, so the same trap applies.
|
||||
fireConcurrent(baseUrl, "/cpu", Math.min(concurrency, 20));
|
||||
fireConcurrent(baseUrl, "/cpu-offloaded", Math.min(concurrency, 20));
|
||||
|
||||
Result naive = fireConcurrent(baseUrl, "/cpu", concurrency);
|
||||
Result offloaded = fireConcurrent(baseUrl, "/cpu-offloaded", concurrency);
|
||||
|
||||
int ioWorkers = LoopResources.DEFAULT_IO_WORKER_COUNT;
|
||||
int parallelWorkers = Runtime.getRuntime().availableProcessors();
|
||||
|
||||
Transcript t = Transcript.start("02-cpu-bound-webflux.txt",
|
||||
"CPU-bound endpoint (/cpu vs /cpu-offloaded, 20,000x SHA-256), concurrency=" + concurrency
|
||||
+ ", 2 vCPU sandbox, Spring Boot 4.1.1 / JDK " + System.getProperty("java.version"));
|
||||
t.line("LoopResources.DEFAULT_IO_WORKER_COUNT (Netty event-loop threads) = " + ioWorkers);
|
||||
t.line("Runtime.availableProcessors() (Schedulers.parallel() thread count) = " + parallelWorkers);
|
||||
t.blank();
|
||||
t.line(String.format("webflux naive (Mono.fromCallable, no subscribeOn) : total=%d success=%d wall=%dms p50=%.0fms p99=%.0fms",
|
||||
naive.total, naive.success, naive.wallMs, naive.p50, naive.p99));
|
||||
t.line(String.format("webflux offloaded (subscribeOn(Schedulers.parallel())) : total=%d success=%d wall=%dms p50=%.0fms p99=%.0fms",
|
||||
offloaded.total, offloaded.success, offloaded.wallMs, offloaded.p50, offloaded.p99));
|
||||
t.blank();
|
||||
t.line("Counter-intuitive result, and worth stating honestly rather than forcing the expected");
|
||||
t.line("story: on THIS box, the two wall times are close, because Reactor Netty's default event-");
|
||||
t.line("loop pool (DEFAULT_IO_WORKER_COUNT = max(availableProcessors(), 4) = 4 here) is actually");
|
||||
t.line("LARGER than Schedulers.parallel()'s pool (sized to availableProcessors() = 2). The naive");
|
||||
t.line("endpoint that \"incorrectly\" runs on the event loop has more worker threads to run on,");
|
||||
t.line("at this modest concurrency, than the \"correctly offloaded\" one. This does not mean the");
|
||||
t.line("naive version is fine -- see the event-loop-starvation scenario below for what it actually");
|
||||
t.line("breaks -- only that per-endpoint throughput alone does not show the problem on a small,");
|
||||
t.line("under-loaded box like this one.");
|
||||
t.save();
|
||||
|
||||
assertThat(naive.success).isEqualTo(concurrency);
|
||||
assertThat(offloaded.success).isEqualTo(concurrency);
|
||||
} finally {
|
||||
ctx.close();
|
||||
}
|
||||
}
|
||||
|
||||
@Test
|
||||
void eventLoopStarvationScenario() throws Exception {
|
||||
String baseUrl = startApp();
|
||||
ConfigurableApplicationContext ctx = currentCtx;
|
||||
try {
|
||||
// Warm-up through both code paths first, same JIT reason as cpuBoundScenario.
|
||||
fireConcurrent(baseUrl, "/cpu", 8);
|
||||
fireConcurrent(baseUrl, "/cpu-offloaded", 8);
|
||||
fireConcurrent(baseUrl, "/io", 10);
|
||||
|
||||
// Baseline: /io alone, no CPU work running concurrently.
|
||||
Result ioBaseline = fireConcurrent(baseUrl, "/io", 20);
|
||||
|
||||
// Fire 8 concurrent /cpu (naive) requests and, while they are still in flight,
|
||||
// fire 20 /io requests at the SAME server -- this is the actual failure mode of
|
||||
// blocking the event loop: it is not that the CPU endpoint itself is slow, it is
|
||||
// that the CPU endpoint holds event-loop threads other requests need.
|
||||
Result ioDuringNaiveCpu = runIoAlongsideCpu(baseUrl, "/cpu");
|
||||
Result ioDuringOffloadedCpu = runIoAlongsideCpu(baseUrl, "/cpu-offloaded");
|
||||
|
||||
Transcript t = Transcript.start("03-event-loop-starvation.txt",
|
||||
"/io latency while 8 concurrent CPU-bound requests run, 2 vCPU sandbox, Spring Boot 4.1.1 / JDK "
|
||||
+ System.getProperty("java.version"));
|
||||
t.line(String.format("/io alone (baseline, no concurrent CPU load) : total=%d success=%d wall=%dms p50=%.0fms p99=%.0fms",
|
||||
ioBaseline.total, ioBaseline.success, ioBaseline.wallMs, ioBaseline.p50, ioBaseline.p99));
|
||||
t.line(String.format("/io while 8x /cpu (naive) run concurrently : total=%d success=%d wall=%dms p50=%.0fms p99=%.0fms",
|
||||
ioDuringNaiveCpu.total, ioDuringNaiveCpu.success, ioDuringNaiveCpu.wallMs, ioDuringNaiveCpu.p50, ioDuringNaiveCpu.p99));
|
||||
t.line(String.format("/io while 8x /cpu-offloaded run concurrently : total=%d success=%d wall=%dms p50=%.0fms p99=%.0fms",
|
||||
ioDuringOffloadedCpu.total, ioDuringOffloadedCpu.success, ioDuringOffloadedCpu.wallMs, ioDuringOffloadedCpu.p50, ioDuringOffloadedCpu.p99));
|
||||
t.blank();
|
||||
t.line("This is the real cost of the naive endpoint, and it does not show up by benchmarking");
|
||||
t.line("/cpu in isolation: /io shares the same small event-loop pool with /cpu. On this run,");
|
||||
t.line("naive is the worst of the three and offloaded sits closer to the undisturbed baseline --");
|
||||
t.line("the ordering this test asserts on. The gap is modest here, not a cliff, because");
|
||||
t.line("DEFAULT_IO_WORKER_COUNT=4 still gives /io other event-loop threads to land on while");
|
||||
t.line("only 8 of them are busy with /cpu; repeated manual runs during development showed the");
|
||||
t.line("same ordering with the gap sometimes larger. At higher concurrency, or with a smaller");
|
||||
t.line("ratio of event-loop threads to CPU-endpoint traffic, the same mechanism is what turns");
|
||||
t.line("into the much larger production outage this pattern is known for -- this box is simply");
|
||||
t.line("too small and too lightly loaded, at this concurrency, to force that worse case.");
|
||||
t.save();
|
||||
|
||||
assertThat(ioBaseline.success).isEqualTo(20);
|
||||
assertThat(ioDuringNaiveCpu.success).isEqualTo(20);
|
||||
assertThat(ioDuringOffloadedCpu.success).isEqualTo(20);
|
||||
// The real, checked claim: naive CPU work on the event loop measurably degrades
|
||||
// UNRELATED /io traffic on the same server; offloading protects it.
|
||||
assertThat(ioDuringNaiveCpu.p99).isGreaterThan(ioDuringOffloadedCpu.p99);
|
||||
} finally {
|
||||
ctx.close();
|
||||
}
|
||||
}
|
||||
|
||||
/** Fires 8 concurrent requests to cpuPath and, without waiting for them, fires 20
|
||||
* concurrent /io requests against the same server, returning the /io Result only. */
|
||||
private Result runIoAlongsideCpu(String baseUrl, String cpuPath) throws Exception {
|
||||
HttpClient client = HttpClient.newBuilder()
|
||||
.executor(Executors.newVirtualThreadPerTaskExecutor())
|
||||
.build();
|
||||
|
||||
// Fire the CPU load in the background, not waited on.
|
||||
for (int i = 0; i < 8; i++) {
|
||||
Thread.ofVirtual().start(() -> {
|
||||
try {
|
||||
HttpRequest req = HttpRequest.newBuilder(URI.create(baseUrl + cpuPath)).build();
|
||||
client.send(req, HttpResponse.BodyHandlers.ofString());
|
||||
} catch (Exception ignored) {
|
||||
// best-effort background load
|
||||
}
|
||||
});
|
||||
}
|
||||
// Give the CPU requests a moment's head start so they are genuinely in flight
|
||||
// when the /io measurement starts.
|
||||
Thread.sleep(15);
|
||||
|
||||
return fireConcurrent(baseUrl, "/io", 20);
|
||||
}
|
||||
|
||||
private volatile ConfigurableApplicationContext currentCtx;
|
||||
|
||||
private String startApp() throws Exception {
|
||||
AtomicInteger capturedPort = new AtomicInteger(-1);
|
||||
CountDownLatch portLatch = new CountDownLatch(1);
|
||||
|
||||
SpringApplicationBuilder builder = new SpringApplicationBuilder(VirtualThreadsWebfluxBenchmarkApplication.class)
|
||||
.initializers(ctx -> ctx.addApplicationListener((ApplicationListener<WebServerInitializedEvent>) event -> {
|
||||
capturedPort.set(event.getWebServer().getPort());
|
||||
portLatch.countDown();
|
||||
}));
|
||||
|
||||
currentCtx = builder.run("--server.port=0", "--spring.jmx.enabled=false");
|
||||
portLatch.await(10, TimeUnit.SECONDS);
|
||||
return "http://localhost:" + capturedPort.get();
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user