Benchmarking Go and Python Web servers(ziutek.github.com)
ziutek.github.com
Benchmarking Go and Python Web servers
http://ziutek.github.com/web_bench/
14 comments
I'm mostly pumped that tornado is holding up pretty well against a compiled language. :)
Don't forget that Go still has a lot to improve wrt code optimization (no function inlining yet, simple goroutines scheduler, for example), and that its regex implementation is still simple and not optimal (I heard they plan to replace it with RE2 later).
Wondering if Go is going to take 5 more years to be "ready" - or at least at a stage where there are no caveats with benchmarks.
Couldn't Google fast track development? Or is it that it was built for internal purposes and it's not a priority to mainstream it?
http://golang.org/doc/devel/roadmap.html
Couldn't Google fast track development? Or is it that it was built for internal purposes and it's not a priority to mainstream it?
http://golang.org/doc/devel/roadmap.html
Here's profile of http.go https://github.com/traviscline/web_bench/blob/master/http.go:
287.36% runtime.mach_semaphore_wait
80.46% syscall.Syscall6
2.30% runtime.cas
2.30% runtime.memmove
1.15% MCentral_Free
1.15% MHeap_FreeLocked
1.15% bytes.*Buffer·WriteString
1.15% crypto/sha256.Init·
1.15% hash_remove
1.15% runtime.MCache_Alloc
1.15% runtime.MHeap_LookupMaybe
1.15% runtime.breakpoint
1.15% runtime.mach_semaphore_signal
1.15% runtime.mallocgc
1.15% runtime.xadd
1.15% syscall.Syscall
I kinda expected Go to lose. It's HTTP library seems optimized for simplicity, and I get the impression that very little performance tuning has been done on the scheduler.
I'd like someone to do a Shootout-style multi-language continuous benchmark for network-type apps to help judge the general suitability of languages to IO and memory-management heavy tasks.
Stop the press, news are news, a compiled language built for concurrency is better at creating performant web servers than an interpreted language with a GIL.
Next in line, is an Nginx in-module application faster than Rails on WEBRick? We're not sure yet, but we're going to tell you. Stay tuned.
Next in line, is an Nginx in-module application faster than Rails on WEBRick? We're not sure yet, but we're going to tell you. Stay tuned.
I think it's some sort of interesting how small the gap is.
GIL isn't really relevant for blocking IO-bound code, because the lock is released (and not contended) when blocking. For nonblocking single-threaded code such as Tornado's event loop, the GIL also doesn't matter. Not to mention that the real speed concerns had nothing to do with interpreted versus compiled, because Tornado very clearly held its own, and web.py+flup (etc.) have architectural issues with regard to performance, far beyond mere interpreter overhead.
It looks like he used the built-in webpy cherrypy server, but if I recall correctly that server is recommended only for debugging, not for production.
It's interesting how you can predict bias depending on ordering of the title.
Throw in native event loops with twisted and compare.
Added a basic gevent example here that is apparently about 10% better than the http.go wrt transaction rate.
https://github.com/traviscline/web_bench
(notice: I realize this is a highly synthetic benchmark and reading siege results like this is not good benchmarking)
https://github.com/traviscline/web_bench
(notice: I realize this is a highly synthetic benchmark and reading siege results like this is not good benchmarking)
It's too bad the author didn't include gevent in the post, as it seems to be the fastest way to do concurrency with python (see http://nichol.as/benchmark-of-python-web-servers). I'm surprised it was faster than the Go example. I wonder if a Cython-based handler would take it even further...
I don't see the use of regexp in gvent.py.
Good call, added it, initial run didn't show a huge difference.
edit: here are numbers from a few runs: gvent,http.go 361.43,386.68 354.69,397.15 388.96,377.69 424.81,430.34
http.go averaged 26% utilization gvent.py averaged 15% utilization
I'm embarrassed posting this because of how grossly unscientific this is..
Not quite as fast as I thought initially but likely faster than the other python competitors.
edit: here are numbers from a few runs: gvent,http.go 361.43,386.68 354.69,397.15 388.96,377.69 424.81,430.34
http.go averaged 26% utilization gvent.py averaged 15% utilization
I'm embarrassed posting this because of how grossly unscientific this is..
Not quite as fast as I thought initially but likely faster than the other python competitors.
Good. Next, there's a check for "GET" method and if request contains parameters at all (which returns 404 if not) in the Go version. Though, I don't think it will make a big difference.
I don't see how benchmarks like these are really useful to anyone other than the author. I care more about reliability than raw performance on a simple test, so I would rather see a fairly complex application of average quality being run under load.
Stuff like this is the reason why people are saying HN quality is going down.
As submitter, I mostly agree. Microbenchmarks are the celebrity gossip of the tech world. They are mostly meaningless, but they get more attention and controversy than they deserve.
And yet, I submitted it anyway. I saw a link, found it interesting, and as such posted it. I'm a little surprised it made it to the front page, but if there were more interesting content available, I assume it wouldn't have.
And yet, I submitted it anyway. I saw a link, found it interesting, and as such posted it. I'm a little surprised it made it to the front page, but if there were more interesting content available, I assume it wouldn't have.
My comment was not directed to you only. Such things rising to the front page is, in equal part, our(voters/commenters) fault also.
A year back HN front page was all I needed for my daily dose of interesting tech stories(I hate Twitter). Now really good content is buried away in later pages most of the time.
A year back HN front page was all I needed for my daily dose of interesting tech stories(I hate Twitter). Now really good content is buried away in later pages most of the time.
I wonder why he didn't compare against gevent-based servers. In the comparisons I saw between Python-based http servers, those came out highest in perf.
See my other comment, gevent appears to outperform in my initial test.
My experience with siege is that it actually measures siege's terrible performance.
That is pretty much what I would expect from a compiled vs non-compiled language.