Disapointed by Vertx versus raw servlet - performance

Question

Vertx is way slower than raw asyncservlet with exactly the same code to execute
I have implemented a very basic GET http method that does nothing than Thread.sleep(30) before returning http response text "ok". I did this with jetty-servlet and vertx-verticles
Vertx vertx = Vertx.vertx();
DeploymentOptions d = new DeploymentOptions();
d.setInstances(400);
d.setWorker(true);
d.setWorkerPoolSize(400);
vertx.deployVerticle("com.vertx.Ping",d);
public class Ping extends AbstractVerticle {
private HttpServer server;
public void start(Future<Void> startFuture) {
server = vertx.createHttpServer().requestHandler(req -> {
try {
Thread.currentThread().sleep(30);
} catch (InterruptedException e) {
e.printStackTrace();
}
req.response().end("ok");
});
server.listen(9999);
}
}
ab -k -c 1 -n 1 => same result for servlet and vertx : ~31ms
Servlet: ab -k -c 1000 -n 100000
Percentage of the requests served within a certain time (ms)
50% 219
66% 224
75% 227
80% 229
90% 235
95% 240
98% 245
99% 247
100% 264 (longest request)
Vertx: ab -k -c 1000 -n 100000 => after more than 2 minutes i stopped the test
Let's try something easier :
Vertx: ab -k -c 1000 -n 10000 (10 times less)
Percentage of the requests served within a certain time (ms)
50% 3930
66% 3943
75% 3964
80% 3977
90% 3997
95% 4009
98% 4019
99% 4028
100% 4038 (longest request)
Vertx is so damn slow, what do i do wrong folks ? Thank you

Answer 1

You are not using Vert.x as it should be. Remember that Vert.x is based on asynchronous operations.
First off, you don't need 400 worker verticle instances. Just 1 regular (non-worker) verticle instance is more than enough. You are doing this because you call Thread.sleep, which causes a thread to block, but there is a much simpler way.
Next: don't use blocking operations! You are running code on an event-loop, so you must not block. Your code should look like:
vertx
.createHttpServer()
.requestHandler(req -> {
vertx.setTimer(30, tid -> {
req.response().end("ok");
}
})
.listen(9999);
You should see a drastic improvement in your measures, and all of that with just 1 thread compared to the 400 you had.
Note: when you need to call a blocking operation, you should have a look at executeBlocking in the Vertx class, which offloads some code to a worker, and then dispatches the result as a new event.