Introduction

This project, based on Java 8 and Spring boot, is to learn and demonstrate the difference between sequential and parallel service call, and how the system performance is influenced by them.

The structure of the apis

A slow API, which sleeps for 2 seconds, is located in /legacy/slow/{id}. {id} can be any Integer number to represent different APIs.
Other three APIs /seq/combined, /parallel/combined, /webflux/combined are represented for a combined API, which combines the backend APIs in a different way.
- For simplify the example, they all call the slow API twice with different paramerters 1 and 2;
- /seq/combined calls the slow APIs by sequence way;
- /parallel/combined calls the slow APIs by parallel way(Asychronous by using okhttp);
- /webflux/combined calls the slow APIs by reactive way(Stream by using spring-webflux);

Obviously, the total time cost is the longest time of one of the backend APIs in parallel and reactive architecture, rather than the sum time for every APIs in sequential system.

ELE-MacBook-Pro:parallel ele$ time curl -Is localhost:8080/seq/combined > /dev/null

real	0m4.033s
user	0m0.006s
sys	0m0.006s
ELE-MacBook-Pro:parallel ele$ time curl -Is localhost:8080/parallel/combined > /dev/null

real	0m2.060s
user	0m0.006s
sys	0m0.006s
ELE-MacBook-Pro:parallel ele$ time curl -Is localhost:8080/webflux/combined > /dev/null

real	0m2.051s
user	0m0.006s
sys	0m0.006s

The performance

I did a benchmark on my laptop, and the data was collected below.

The computer, CPU: 2.7 GHz Intel Core i7, Memory: 16 GB 1600 MHz DDR3
Jvm parameters, -Xms1G -Xmx2G
Use Apache Benchmark to request the apis with two specs. -c 10 -n 200 is currency 10 and total number 200 requests, and -c 100 -n 2000 is currency 100 and total number 2000 requests

Here are some overview indicators, and details are listed below the table.

		Sequcence	Parallel	Reactive
-c 10 -n 200	Time taken	84.617	43.387	43.726
	Maximum Thread number	41	651	58
	Peak Heap Memeory	200Mb	349mb	160Mb
-c 100 -n 2000	Time taken	84.718	-	71.388
	Maximum Thread number	219	3100	226
	Peak Heap Memeory	300Mb	550Mb	510Mb

The performance of Sequence api

With -c 10 -n 200, time taken for tests: 84.617 seconds

With -c 100 -n 2000, time taken for tests: 84.718 seconds

The performance of parallel api

With -c 10 -n 200, time taken for tests: 43.387 seconds

With -c 100 -n 2000, the ab crashed until finished 898 requests completed

The performance of reactive api

With -c 10 -n 200, time taken for tests: 43.726 seconds With -c 100 -n 2000, time taken for tests: 71.388 seconds

Further work

Why does the total of the thread increased so much in Parallel call? Is it the problem of the implementation(okhttp) or the architecture?
Is it possible to decrease the programming complexity in parallel and reactive class?
new version prepared

lijiele / parallel-test