- Haskell-Cafe - Haskell.org

Performance of concurrent array access
by Andreas Voellmy 24 Aug '11

24 Aug '11

Hi, I am getting my feet writing concurrent programs in Haskell with GHC for multicore machines. As a first step I decided to write a program that reads and writes concurrently to an IOArray, with no synchronization whatsoever. I'm doing this to establish a baseline to compare with the performance of other data structures that do use appropriate synchronization mechanisms. I ran in to some surprising results, namely that in many cases, I am not getting any speed up at all. Here are the details... I write a couple variations on a single program. The main idea is that I wrote a DirectAddressTable data structure, which is simply a wrapper around an IOArray providing insert and lookup methods: -- file DirectAddressTable.hs module DirectAddressTable ( DAT , newDAT , lookupDAT , insertDAT , getAssocsDAT ) where import Data.Array.IO import Data.Array.MArray import Data.Int (Int32) data DAT = DAT (IOArray Int32 Char) -- create a fixed size array; missing keys have value '-'. newDAT :: Int32 -> IO (DAT) newDAT n = do a <- newArray (0, n - 1) '-' return (DAT a) -- lookup an item. lookupDAT :: DAT -> Int32 -> IO (Maybe Char) lookupDAT (DAT a) i = do c <- readArray a i return (if c=='-' then Nothing else Just c) -- insert an item insertDAT :: DAT -> Int32 -> Char -> IO () insertDAT (DAT a) i v = writeArray a i v -- get all associations (exclude missing items, i.e. those whose value is '-'). getAssocsDAT :: DAT -> IO [(Int32,Char)] getAssocsDAT (DAT a) = do assocs <- getAssocs a return [ (k,c) | (k,c) <- assocs, c /= '-' ] I then have a main program that initializes a new table, forks some threads, with each thread writing and reading some fixed number of values to the just initialized table. The overall number of elements to write is fixed. The number of threads to use is a taken from a command line argument, and the elements to process are evenly divided among the threads. -- file DirectTableTest.hs import DirectAddressTable import Data.Int (Int32) import Control.Concurrent import Control.Parallel import System.Environment main = do args <- getArgs let numThreads = read (args !! 0) vs <- sequence (replicate numThreads newEmptyMVar) a <- newDAT arraySize sequence_ [ forkIO (doLotsOfStuff numThreads i a >>= putMVar v) | (i,v) <- zip [1..] vs] sequence_ [ takeMVar v >>= \a -> getAssocsDAT a >>= \xs -> print (last xs) | v <- vs] doLotsOfStuff :: Int -> Int -> DAT -> IO DAT doLotsOfStuff numThreads i a = do let p j c = insertDAT a j c >> lookupDAT a j >>= \v -> v `pseq` return () sequence_ [ p j c | (j,c) <- bunchOfKeys j ] return a where bunchOfKeys i = take numElems $ zip cyclicIndices $ drop i cyclicChars numElems = numberOfElems `div` numThreads cyclicIndices = cycle [0..highestIndex] cyclicChars = cycle chars chars = ['a'..'z'] -- Parameters arraySize :: Int32 arraySize = 100 highestIndex = arraySize - 1 numberOfElems = 10 * 1000 * 1000 I compiled this with "ghc --make -rtsopts -threaded -fforce-recomp -O2 DirectTableTest.hs". Running "time ./DirectTableTest 1 +RTS -N1" takes about 1.4 seconds and running "time ./DirectTableTest 2 +RTS -N2" take about 2.0 seconds! Using one more core than worker threads is a little better, with "time ./DirectTableTest 1 +RTS -N1" takes about 1.4 seconds and running "time ./DirectTableTest 1 +RTS -N2" and "time ./DirectTableTest 2 +RTS -N3" both taking about 1.4 seconds. Running with the "-N2 -s" option shows that productivity is 95.4% and GC is 4.3%. Looking at a run of the program with ThreadScope I don't see anything too alarming. Each HEC yields once per ms when a GC occurs. Running with 4 cores gives a time of about 1.2 seconds, which is at least a little better than 1 core. More cores doesn't improve over this. I found that changing the array type used in the implementation of DirectAddressTable from IOArray to IOUArray fixes this problem. With this change, the running time of "time ./DirectTableTest 1 +RTS -N1" is takes about 1.4 seconds whereas the running "time ./DirectTableTest 2 +RTS -N2" is about 1.0 seconds. Increasing to 4 cores gives a run time of 0.55 seconds. Running with "-s" shows a GC time of %3.9 percent. Under ThreadScope I can see that both threads yield every 0.4 ms, more frequently than in the previous program. Finally, I tried one more variation. Instead of having the threads work on the same shared array, I had each thread work on its own array. This scales nicely (as you would expect), more or less like the second program, with either IOArray or IOUArray implementing the DirectAddressTable data structure. I understand why IOUArray might perform better than IOArray, but I don't know why it scales better to multiple threads and cores. Does anyone know why this might be happening or what I can do to find out what is going on? Regards, Andreas

4 6

Function composition in run-time?
by dokondr 24 Aug '11

24 Aug '11

Hi, What is the Haskell way to compose functions in run-time? Depending on configuration parameters I need to be able to compose function in several ways without recompilation. When program starts it reads configuration parameters from a text file. For example, I have three functions, f1, f2, f3, each doing some string processing. I need to support two configurations of string processors : if param1 then sp = f1 . f2 . f3 else sp = f1 . f3 I'd like to avoid 'if' somehow and instead use some declarative way to specify code to run in external configuration file. In other words I need some easy tools to create mini DSLs without all the efforts usually involved with implementing full-blown DSL. Thanks, Dmitri

5 7

different results after compilation
by Комар Максим 24 Aug '11

24 Aug '11

I have some script: $ runhaskell readfile.hs fromList [(Merchant {nick = "01010", location = "prontera", x = 184, y = 94},Shop {buy = ShopBuy {titleB = "AB> Green Salad=5k", itemsB = fromList [(Item {itemId = 12065, price = 5000, refine = "", card1 = 0, card2 = 0, card3 = 0, card4 = 0},(100,97))]}, sell = ShopSell {titleS = "", itemsS = fromList []}})] $ ghc --make readfile.hs Linking readfile ... $ ./readfile $ why results are different?

3 2

- Try to install ssh package by cabal.
by Loïc Maury 24 Aug '11

24 Aug '11

Hello, I try to install the ssh package with cabal, but unfortunately that doesn't work, I have this error : src/SSH/Crypto.hs:70:16: Couldn't match expected type `Data.ASN1.Types.ASN1t' with actual type `ASN1ConstructionType' I'am on Debian with ghc-7.0.1 I don't know how I can resolve this issue ? Thank you for help Loïc

2 2

Conflict with Network.CGI and System.Random
by william murphy 23 Aug '11

23 Aug '11

Hello all, I'm trying to write a program to be converted into a cgi file using Network.CGI and Text.Xhtml, and I want to have random numbers in it. Sounds simple, but I need to call the global random generator (I think) and I can't do that in the CGI block, i.e. cgiTMain = do gen <- getStdGen ... simply doesn't work, nor an variation of calling a function in the cgiTMainblock with a type of IO(), which is what I need to get the generator. Is it impossible both to write a program using Network.CGI and have random values? That seems very limiting, the people that wrote Haskell were smarter than that... Thanks, Will

2 1

Please take the State of Haskell, 2011 survey
by Johan Tibell 23 Aug '11

23 Aug '11

Hi all, I've put together a quick, 12-question State of Haskell, 2011 survey: http://blog.johantibell.com/2011/07/its-time-for-this-years-state-of.html The survey will hopefully give us some insight into how people use Haskell and perhaps also some ideas on how Haskell tools and libraries could be improved. This is the second year I run this survey. New from last year are specific questions on library support and reasoning about run-time performance. P.S. Please direct replies to this email to haskell-cafe(a)haskell.org. Cheers, Johan

4 5

Post-doc position
by Jacques Carette 23 Aug '11

23 Aug '11

People interested in (typed) code generation might be interested (or know people who are) in the following post-doc position. While the 'topic' is design for multiple visual scales, the tools used in practice will be DSLs and (typed) code generators. We use Haskell for implementing all of our DSLs (see [1] for details of part of the ongoing work). Feel free to contact me for further details. Jacques [1] http://www.cas.mcmaster.ca/~carette/SAGA/ [Sorry for the bad formatting, this 'ad' was unfortunately born as a Word document. Not my choice.] ============= Council on Library and Information Resources (CLIR) Postdoctoral Fellowship in Design for Multiple Visual Scales at McMaster University The Computing and Software Department and the University Library at McMaster University in Hamilton, Ontario, Canada invite applications for a two-year CLIR Postdoctoral Fellowship in software design, with an emphasis on games, for multiple visual scales, from the iPod Touch to 60inch high-resolution screen panels. The Fellowship is part of the Gaming Scalability Environment (G-ScalE) project involving Dr. Jacques Carette (Computing and Software, Faculty of Engineering), Dr. Andrew Mactavish (Multimedia, Faculty of Humanities), and Jeff Trzeciak (University Librarian). The fellow will start immediately and report to Dr. Carette but also work closely with the library on certain projects. Duties • Work with Dr. Carette and his students to identify all of the issues relating to usability at different scales, then help design (and implement) a set of domain-specific languages which capture the constraints of user-interface features with respect to scale. o From the psychology and user-interface literature, identify cognitive issues relating to scale perception, and capture these in a formal model. Examples: Fitts’ Law, Hick’s Law and the Steering Law; o Identify examples of such concerns, and pre-existing solutions, from existing software which works at multiple scales; put an emphasis on solutions concerning gaming software; o Relate these models and solutions to design constraints for features in software; o Embody these design constraint into a set of domain specific languages; and o Provide interpreters for these languages. • The above work will be done in conjunction with a number of other researchers and students, and is expected to lead to several peer-reviewed publications, as well as working software. • The candidate will be expected to teach one course (in the CAS department). • The candidate will be expected to work with the library to: o Identify the current and future library needs of Engineering faculty and students; o Assess current state of library resources, services and facilities in support of Engineering; o Make recommendations to the library administration for improvement; o Promote greater awareness of library resources, services and facilities within Engineering; o Promote greater awareness within the library of the faculty’s teaching and research priorities ; and o Provide support as needed to students and faculty within Engineering. Required Skills: • Ph.D. in Computer Science or Software Engineering is preferred, but a Ph.D. in psychology, human factors, or mathematics accompanied by a demonstrable expertise in software development and mathematical modelling. • Interest in multi-disciplinary work. • Capable of understanding the requirements of faculty and students in Engineering with regards to library resources, services and facilities. How to Apply If you have received your Ph.D. within the last five years and are interested in this fellowship, please email your curriculum vitae to abishop(a)clir.org. The Council on Library and Information Resources (CLIR) is an independent, nonprofit organization that forges strategies to enhance research, teaching, and learning environments in collaboration with libraries, cultural institutions, and communities of higher learning. CLIR promotes forward-looking collaborative solutions that transcend disciplinary, institutional, professional, and geographic boundaries in support of the public good.

1 0

Is there anyone interested in translating Haskell Report 2010 to Chinese?
by Qiuchi Jian 23 Aug '11

23 Aug '11

Hi, I am trying to read and translate the 2010 Haskell Report to Chinese for fun. I am wondering if anyone would like to join me. Cheers, Qiuchi

2 1

Why no Monoid, Num, Integral, or Floating RVarT instances?
by Barend Venter 23 Aug '11

23 Aug '11

It seems like you should be able to get instances for: (Monoid m) => Monoid (RVarT n m) (Num m) => Num (RVarT n m) and so forth for integral and floating Just wondering if I am missing something here or if there is any particular thing I'm missing. I am aware you would probably not be able to do an Eq or Show instance but me understanding was that in the case of the (Num b) => Num (a -> b) instance the solution was simply to leave those things undefined.

3 2

how long could it wait for an account on hackageDB?
by Qi Qi 23 Aug '11

23 Aug '11

Hello all, I wanted to have an account on hackageDB to upload a package. I followed http://hackage.haskell.org/packages/accounts.html, and sent an email to ross(a)soi.city.ac.uk for asking an account. It's been a couple days that I haven't heard back. Is this normal, or I have to do something else? Thanks, -- Qi Qi

2 1