I would like to improve throughput in my web-server. But I don't understand what is really happening when I am testing it with high load (jmeter).
I run spring boot webflux
app (spring boot 2.0.2, netty
) on a computer with 8 core
cpu.
I created a simple controller with this code:
@GetMapping("/test-delay")
public Mono<String> testGetWithDelay() throws InterruptedException {
Thread.sleep(3000);
return Mono.just("current time:" + LocalDateTime.now());
}
"Thread.sleep(3000)"
- it is an imitation of synchronous works.
Then I run jmeter tests
with 100 threads. And I see a throughput only 2.5 message/sec
. I thought it should be about 100 messages/3 sec
(about 30 messages/sec)
So, I have 2 questions:
Thanks
Your results are correct. You get 2.5 message/sec with delay 3 secs (in every thread), that gives us 2.5 * 3 = 7.5 = 8
cores. Webflux by default uses availableProcessors()
as a default number of threads for processing network/work IO.
So what you need to do is either increase the number of processing threads or move Thread.sleep(3000)
block to separate ThreadPool/Executor
(so working threads are not blocked) or your Thread.sleep(3000)
code should be executed in some kind of non-blocking API (for example in webflux you may use Mono.fromCallable
).
I would recommend you to go with the second/third approach, as non-blocking API should never be blocked.