[Haskell-cafe] Slow mvar when compiled with threaded

[Bob Ippolito]

So I dug a bit deeper (trying N different ways to write the code,
ThreadScope, etc.) and I have some good news some bad news.

The good news is that I figured out how to fix it, and it's an "easy" fix.

The bad news is that the fix is to compile it with GHC HEAD (~7.8, I built
ec4af3f), which has the new Mio high-performance multicore IO manager [1].
Apparently the old IO manager wasn't well suited to this use case.

With the latest GHC, the -threaded version outperforms the single threaded
version, and the numbers for both are better than with GHC 7.6.3.

I've updated the gist with a README that has some more details:
https://gist.github.com/etrepum/8312165

[1]
http://haskell.cs.yale.edu/wp-content/uploads/2013/08/hask035-voellmy.pdf

-bob



On Wed, Jan 8, 2014 at 7:50 AM, Bob Ippolito <bob at redivi.com> wrote:

> I haven't yet been able to sort out the performance difference, and
> probably won't have time to dig in deeper with profiling tools today. I've
> put together a gist with all of the files and a cabal file so it's easy to
> build with the right options.
>
> https://gist.github.com/etrepum/8312165
>
>
> On Tue, Jan 7, 2014 at 7:47 PM, Bob Ippolito <bob at redivi.com> wrote:
>
>>
>>
>> On Tuesday, January 7, 2014, Branimir Maksimovic wrote:
>>
>>>  On 01/08/2014 04:15 AM, Bob Ippolito wrote:
>>>
>>> On Tuesday, January 7, 2014, Branimir Maksimovic wrote:
>>>
>>>> On 01/07/2014 09:39 PM, Bob Ippolito wrote:
>>>>
>>>>> Here's a much simpler implementation for that sort of pattern, using
>>>>> channels to fan out work to threads. I added a dependency on Criterion
>>>>> because getCPUTime is basically useless for this kind of measurement on Mac
>>>>> OS X since it doesn't include the time that the process spent waiting on IO:
>>>>>
>>>> Great, thank you very much. You gave me material for learning ;)
>>>> However, my version is significantly faster when compiling without
>>>> -threaded.
>>>> With -threaded option, your version is much faster than mine, but both
>>>> are significantly slower
>>>> then compile without -threaded.
>>>>
>>>
>>>  Happy to look into it, I didn't have time today to do benchmarks
>>> (and Mac OS X is the worst platform to do this kind of testing on
>>> regardless of language, its network stack is inconsistent at best). I need
>>> to know more:
>>>
>>>  How exactly are you compiling?
>>>
>>>
>>> ghc-7.6.3 --make -O2 client.hs
>>>
>>>  Which OS?
>>>
>>>
>>> Ubuntu 13.10 with 3.13-rc7 kernel.
>>>
>>>  What version of GHC and Haskell Platform?
>>>
>>>
>>> bmaxa at maxa:~$ apt-cache policy haskell-platform
>>> haskell-platform:
>>>   Installed: 2013.2.0.0
>>>   Candidate: 2013.2.0.0
>>>   Version table:
>>>  *** 2013.2.0.0 0
>>>         500 http://archive.ubuntu.com/ubuntu/ saucy/universe amd64
>>> Packages
>>>         100 /var/lib/dpkg/status
>>>
>>>
>>>  What is the exact command line you execute it with?
>>>
>>>
>>> time ./client maxa 5055 1000 100000
>>>
>>
>> What happens if you add +RTS -N to the end of that command line?
>>
>>
>>>
>>>
>>>  What timings do you get?
>>>
>>> with your version:
>>> real    0m4.235s
>>> user    0m1.589s
>>> sys    0m2.642s
>>>
>>> with my version
>>> real    0m3.010s
>>> user    0m0.590s
>>> sys    0m2.417s
>>> that is, of course, without -threaded
>>>
>>>  What's the code for the server are you connecting to?
>>>
>>> import Network (listenOn,PortID(..))
>>> import Network.Socket (accept,close)
>>> import Network.Socket.ByteString
>>> import System.Environment
>>> import Control.Concurrent (forkIO)
>>>
>>> main = do
>>>     n <- getArgs
>>>     let nn = (read.head) n :: Int
>>>     putStrLn $ "Listening on " ++ show nn
>>>     sock <- listenOn $ PortNumber $ fromIntegral nn
>>>     serve sock
>>>
>>> serve sock = do
>>>     (s,_) <- accept sock
>>>     forkIO $ process s
>>>     serve sock
>>>
>>> process sock = do
>>>     buf <- recv sock 1024
>>>     sendAll sock buf
>>>     close sock
>>>
>>>  Loopback, local network, or internet?
>>>
>>>
>>> loopback. Testing is on same computer.
>>>
>>> Thanks!
>>>
>>
>> Great, I'll look closer later tonight or tomorrow morning.
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/haskell-cafe/attachments/20140109/5872b6cd/attachment.html>