Lev Walkin wrote:
> Valentin Micic wrote:
>> Do you think that HiPE may make any difference here?
>
> I do not think so, since it uses extra-modular calls, such
> as gen_tcp and lists. But let's try it anyway...
Kostis Sagonas has enarmed me with a hipe:c(Module) hint with
which you can recompile existing (standard) modules into
native ones.
Therefore I redid the hipe benchmark. Here's what I did:
hipe:c(lists).
hipe:c(proplists).
hipe:c(gen_tcp).
hipe:c(inet).
% Test:
18> code:is_module_native(gen_tcp).
true
19>
and, of course, I compiled Yucan as native, using
c(yucan, [native]).
Here's an updated stats:
Rate,Received reply rate,Normalized error rate (1/100%),"Generated
request rate (also, expected reply rate)",Error rate,Attempt 1, Attempt
2, Attempt 3, Error 1, Error 2, Error 3
1000 rps,1000,0,1000,0,1000,1000,1000,0,0,0
2000 rps,1999,0,2000,0,1999,2000,2000,0,0,0
3000 rps,2999,0,3000,0,2999,2999,2999,0,0,0
4000 rps,3997,0,4000,0,3997,3997,3997,0,0,0
5000 rps,4997,0,5000,0,4998,4998,4997,0,0,0
6000 rps,5999,0,6000,0,5999,5999,5999,0,0,0
7000 rps,6999,0,7000,0,7000,6999,7000,0,0,0
8000 rps,7998,0,8000,0,7996,7996,8003,0,0,0
9000 rps,8422,66,9000,0,8972,8988,7308,0,1,1
10000 rps,8523,900,10000,9,7773,8234,9563,9,7,11
Compared to the earlier ones:
> SMP-128-nohipe
>
> 7000 rps,7000,0,7000,0,7000,7000,7000,0,0,0
> 8000 rps,7982,0,8000,0,7998,7954,7994,0,0,0
> 9000 rps,8755,(533),9000,5,8785,8736,8744,5,6,5
> 10000 rps,8969,(1500),10000,15,9012,8953,8942,14,13,18
>
> SMP-128-HIPE
>
> 7000 rps,6706,0,7000,0,6119,7000,7000,0,0,0
> 8000 rps,8001,0,8000,0,8002,8000,8002,0,0,0
> 9000 rps,8937,(33),9000,0,8923,8892,8998,0,1,0
> 10000 rps,9159,(1166),10000,11,9234,8661,9582,15,10,10
Here's a summary:
8000 RPS err % 9k RPS err% 10kRPS err%
Non-HIPE 0% 5% 15%
HiPE(yucan) 0% 0.3% 11%
HiPE([...]) 0% 0.6% 9%
Interpretation: contrary to my earlier assumption, compiling
the yucan.erl module using HiPE was getting me the most benefit.
I now attribute it to using {packet, http} filter, which is
already implemented in C as part of Erlang VM. No significant
amount of extra-modular calls are employed in Yucan, thus
compiling the rest of the system modules (gen_tcp, gen_server,
inet) into HiPE helped too much.
Overall, the effect from HiPE on the numer of error-free
connections per second seems to be about 10%, which,
in my particular case, does not justify any additional
deployment complexities.
> As you see, HIPE scales marginally better at 0.3% errors @ 9kRPS,
> whereas nohipe has 5% errors @9kRPS.
>
> Overall, this picture tells that it is not worth it. Those intra-modular
> calls between HIPE and non-hipe code can kill the whole idea. Also, HIPE
> assembled code has proven unstable under real workload in my project,
> which is far more important factor.
>
>> Yes, Yucan!
>>
>> V.
>>
>> -----Original Message-----
>> From:
[hidden email] [mailto:
[hidden email]] On
>> Behalf Of Lev Walkin
>> Sent: 19 September 2009 07:30 PM
>> To: Evans, Matthew
>> Cc: Erlang Questions
>> Subject: Re: [erlang-questions] Erlang, Yaws, and the deadly Tornado
>>
>> Evans, Matthew wrote:
>>> Thanks for sharing.
>>> I can't see in your tests what version of Erlang you are using? If R12,
>> would R13 be a better performer?
>>
>> R13B01
>>
>>> ________________________________________
>>> From:
[hidden email] [
[hidden email]] On
>>> Behalf
>> Of Lev Walkin [
[hidden email]]
>>> Sent: Saturday, September 19, 2009 3:54 AM
>>> To: Erlang Questions
>>> Subject: [erlang-questions] Erlang, Yaws, and the deadly Tornado
>>>
>>> ======================
>>> Since Facebook acquisition of FriendFeed, a bunch of technologies were
>>> released to the wild, including, most notably, a Tornado web server
>>> written in Python. The Tornado is touted as a <a scalable, non-blocking
>>> web server and web framework>. See Wikipedia article
>>>
http://en.wikipedia.org/wiki/Tornado_HTTP_Server on some details on the
>>> performance of that server, as well as some comparison with other web
>>> servers.
>>>
>>> The numbers looked interesting, so I decided to benchmark Tornado myself
>>> to check out how it fares against some Erlang tools.
>>> ======================
>>>
>>> Since the data is highly graphical, I can only give the link to the rest
>>> of the benchmark.
>>>
>>>
http://lionet.livejournal.com/42016.html>>>
>>>
>>>
>>> --
>>> vlm
>>>
>>> ________________________________________________________________
>>> erlang-questions mailing list. See
http://www.erlang.org/faq.html>>> erlang-questions (at) erlang.org
>>>
>>>
>>> ________________________________________________________________
>>> erlang-questions mailing list. See
http://www.erlang.org/faq.html>>> erlang-questions (at) erlang.org
>>>
>>>
>>
>>
>> ________________________________________________________________
>> erlang-questions mailing list. See
http://www.erlang.org/faq.html>> erlang-questions (at) erlang.org
>>
>
>
> ________________________________________________________________
> erlang-questions mailing list. See
http://www.erlang.org/faq.html> erlang-questions (at) erlang.org
>
________________________________________________________________
erlang-questions mailing list. See
http://www.erlang.org/faq.htmlerlang-questions (at) erlang.org
Locked
