add Compojure with htp-kit and dbcp.clj

[shenfeng]

Another test added to the suite using Compojure, but using http-kit[1] as the server and dbcp.clj[2] to talk to database

1. http://http-kit.org
2. https://github.com/http-kit/dbcp.clj

It should run quite fast and make it quite an interesting option.

[goshacmd]

👍

[TechEmpower]

Thanks for the contribution @shenfeng

[shenfeng]

Hi, thanks. I am more than happy to answer any question about http-kit.

http-kit is written in JAVA, I guess it should have better performance than Netty or servlet [1], and a nice Clojure API is exported, so, we can have much better API than Netty or servlet can offer. The combined result is interesting: we have runtime performance and developer performance at the same time.

[1] https://github.com/ptaoussanis/clojure-web-server-benchmarks

[bhauer]

Hi, Shen. Thanks again for this test for http-kit.

We included http-kit in the runs that we kicked off this Tuesday using @wg's Wrk in lieu of WeigHTTP. We switched to Wrk in order to get latency data but a side-effect is that Wrk is able to squeeze a bit more performance out of the frameworks, so many results have gone up slightly versus our first round of tests.

On our i7 hardware, the results for http-kit are respectable: approximately 68K rps for the JSON serialization test, but that does not rival Netty or Servlets. This has me curious about two things:

• Are there perhaps additional configuration settings or tuning that we could include in our next round (round 3) to improve the results for http-kit?
• Have you had a chance to run this particular test in your own environment using Wrk or WeigHTTP (and not ApacheBench)? We avoid ApacheBench because it's single-threaded and tends to hide the real performance potential of servers. Doing so might highlight something.

As with every test, we want to be sure that we're not crippling http-kit by way of some silly configuration oversight.

[shenfeng]

Are there perhaps additional configuration settings or tuning that we could include in our next round (round 3) to improve the results for http-kit?

http-kit has little to configure.

#setup.py
 command = "lein run -- --db-host " + args.database_host + " --worker=${CPU_CORES}"

# should have better raw performance

Servlet can do 213,322 and netty 203,970? while http-kit's can only do 68k. That's faster than I think.

I can understand netty and servlet faster than http-kit. I've read the netty one's code, it's pretty low level. And I can not beat Java with Clojure: Clojure is write on top of Java :)

I run wrk on the netty and http-kit on a Macbook Air with 1.7 GHZ cpu:

# netty
./wrk -t3 -c400 -r100k http://localhost:8080/
Making 100000 requests to http://localhost:8080/
  3 threads and 400 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    13.06ms    9.30ms  69.33ms   82.43%
    Req/Sec     7.84k     0.95k    9.00k    74.32%
  100096 requests in 3.69s, 17.18MB read
  Socket errors: connect 0, read 266, write 0, timeout 0
Requests/sec:  27151.18
Transfer/sec:      4.66MB

#http-kit
 ./wrk -t3 -c400 -r100k http://localhost:8080/http-kit/json
Making 100000 requests to http://localhost:8080/http-kit/json
  3 threads and 400 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    16.00ms   12.53ms  63.93ms   63.48%
    Req/Sec     5.14k   793.47     7.00k    51.30%
  100017 requests in 5.80s, 16.02MB read
  Socket errors: connect 0, read 808, write 0, timeout 7
Requests/sec:  17240.70
Transfer/sec:      2.76MB

I am curious about the database access performance ranking. I expect it to be slower than netty or servlet.

[shenfeng]

Hi, I just sent another pull request #103, hopefully, will make http-kit run faster.

[bhauer]

Excellent. We'll get that applied soon!

[shenfeng]

Hi, I just sent another pull request #147.
hopefully, will make http-kit run faster.