- Glasgow-haskell-users

"+RTS -A" parameter and CPU cache size
by Bulat Ziganshin 16 Jun '06

16 Jun '06

Hello glasgow-haskell-users, "+RTS -A" parameter sets up size of memory allocation area. by default it set to 256kb and this means that after each 256 kb allocated minor GC occurs and data from this block that is still alive are moved to main heap area (what is scanned only on major GCs) while rest of space is freed. also on each minor GC all mutable variables (IORef/STRef) and mutavle arrays (IOArray/STArray) are scanned and therefore programs that use them heavily (including GHC itself) works faster when value of "-A" parameter is increased. For example, by setting this parameter to 10mb you can decrease count of minor GCs in 40 times. The speed change depends on concrete program, for GHC itself it may be 20-50% faster. that is the example of using this option to speed up compilation: ghc --make main.hs +RTS -A10m well, but what if we don't have many global variables and therefore don't need to increase -A? in this case it's better to have such value of this parameter that all "local data" hold in L1+L2 CPU caches. that is especially important when program generates large number of temporary data cells with very short lifetime. i've found a good example of such program in my own vGetLine-testing code - it reads millions of 20-char lines in one loop. my CPU has 128k total data cache and for this task the most efficient -A value was about 100kb, what allows this benchmark to run 2x faster that with standard -A256k value in general, data cache should be enough to hold: 1) "nursery area", whose size is set by -A 2) part of data from this nursery area what remains alive after minor GC (it may be anything between 0 kb and whole nursery area) 3) all data outside of nursery area what program access between two minor GCs (unpredictable value, in general) knowing this all, we can try to calculate best size of nursery area (i.e., "-A" value), depended on CPU cache size. although we can't find exact optimum because some ratios in list above are unknown, we still can say that optimization should be done for cases where amount of "external" (outside nursery) data accessed between GCs is no more than size of nursery itself and that in typical cases more than 50% of data in nursery will be freed by minor GC. based on these assumptions, we can make conclusion that nursery size should be about 1/2 of cache size. if nursery size is greater or equal to cpu cache size, then it will work inefficient, overfilling L2 cache and "trashing" memory - just like the program that uses little more memory than physically available trashes the swap disk now let's analyze effective data cache size of today CPUs: celeron - 128k-256k sempron, athlon, athlon xp - 192k-320k (64k of L1 + 128k-256k of L2 cache) celeron d, athlon 64, pentium4, pentium d - 512k-2048k as you can see, most of cpus, used in today systems, don't have cache large enough to work efficiently with 256k nursery size (according to my calculations above, it requires 512k or larger cache). this results in sub-optimal performance of ghc-compiled programs. according to my tests, every algorithm tested can be made faster either by increasing -A to 10mb, or decreasing to 1/2 of my cpu cache! the only problem is to automatically decide what to do for each particular algorithm - increase or decrease? :) moreover, when decreasing -A to 64k. speed loss for some algorithms can be also significant - up to 1.5 times. of course, this is the same algorithms that take advantage when -A is significantly increased (see first paragraph for explanations). this problem don't allow me to recommend automatic decreasing of -A for low-level cpus with GHC 6.4, but fortunately it should almost gone in ghc 6.6 taking this all into account, i propose the following: 1) GHC faqs already suggests using of RTS -A/-H option to speed up compilation. i propose to move this suggestion right to compiler itself and add the line char *ghc_rts_opts = "-A10m"; to GHC 6.4 sources 2) i propose to write "L2 cache size detection" code and use it in GHC 6.6 RTS to setup initial value of "-A" option. in order to allow program tune itself to any cpu architecture, with cache sizes ranging from 128kb to 4mb. this will allow low-level cpus to run significantly faster on some algorithms (up to 2x, as i said above) and can give 5-10% speedup for high-level cpus, that is also not so bad :) -- Best regards, Bulat mailto:Bulat.Ziganshin@gmail.com

3 5

Re: installing a Cabal package non-exposed
by Christian Maeder 16 Jun '06

16 Jun '06

Simon Marlow schrieb: > Before we add support for this to Cabal, we should ask why it is needed. > The only reason (that I know of) to hide a package by default is > because its namespace conflicts with another exposed package. Is that > the case here? Yes, I had identical hierarchical module names in two packages, one was a HaXml-13-CVS Cabal package, the other an old uni-haxml package that's usually included via ghc's -package-conf option into our sources. Unfortunately both versions have been modified in different directions (and almost nobody wants to touch our old uni code). Nevertheless, I thought, it should be possible to install different hidden versions of libraries such that users (or user groups) are free to explicitly pick the version that suits them most (although they'll have to stick to that single version then). Cheers Christian

1 0

Any way to catch runtime errors in a DLL?
by Michael Marte 16 Jun '06

16 Jun '06

Hello *, if a runtime error occurs inside a DLL compiled by ghc (like "irrefutable pattern match failed" or exceptions caused by error), the application that called the DLL function dies. This is ok for development but unacceptable when it happens with a user sitting in front of the display. (It has not yet happened but it's only a question of time.) So my question is: Is there any way to catch and process runtime errors? I am looking for some way to map those errors to exceptions on the C++ side that can be caught if required. It would be ok to kill the Haskell runtime system and unload the DLL if necessary. Michael

5 6

installing a Cabal package non-exposed
by Christian Maeder 15 Jun '06

15 Jun '06

Hi, how can install a Cabal package hidden directly? Currently I simply install it and have to hide it afterwards, manually. (Meanwhile one could get a conflict with new exposed modules.) I've tried "./setup register --gen-script" and edit the file "register.sh", but that does not install the .a file. Thanks Christian

2 1

STM applications and examples?
by Simon Marlow 15 Jun '06

15 Jun '06

Hi Folks, I'm interested in gathering information about existing STM applications and example code that people might be working on or have lying around. It would be great to maintain a list of such applications and example code on the wiki somewhere, and use it as a resource for learning STM and research into STM implementations. If you have an application using STM, please let us know. If you can supply the code, even better. Similarly if you have some STM code that might be suitable as samples or benchmarks, we'd love to collect it. Thanks! Simon

3 2

Solaris x86 package
by Georg Sauthoff 14 Jun '06

14 Jun '06

Hi, finally I built a working ghc 6.4.2 [1] for Solaris x86. Since I couldn't find any binary package at the GHC homepage, this was a little time consuming procedurce (starting with unregistered build etc). Thus I want to save other Solaris x86 users from that by contributing a Solaris x86 binary package for the ghc web page. How should I proceed? Should I open a ticket at the homepage for that? What should I consider? Is 'make binary-dist' the right target? And what about some ghc libraries? Is it ok to exclude some, like opengl, readline and X11 related packages, because with them the binary package would be too specific to my Solaris environment (or dependent at a specific readline version)? I guess, because 6.4.3 includes some fixes, it make sense to do the packaging for that version instead of 6.4.2. If no one packages for Solaris sparc, I could do that, too. Regards Georg Sauthoff [1] with non-threaded rts

2 5

ghc sparc NCG
by Georg Sauthoff 13 Jun '06

13 Jun '06

Hi, at http://hackage.haskell.org/trac/ghc/wiki/Platforms I read, that the Sparc NCG 'is bitrotted' - does that mean, that it is unmaintained, but used by default at sparc? I am asking because the solaris sparc system has a ghc 6.2 installed, which generates .s files. Or does NCG means something else? Regards Georg Sauthoff

4 6

Monomorphism restriction - is this still needed for efficiency?
by Brian Hulley 12 Jun '06

12 Jun '06

Hi - I'm wondering if the monomorphism restriction is now just an anachronism from Haskell 98 or if it is still needed for efficiency ie should I just now use -fno-monomorphism-restriction when compiling all my code to consign it to the dustbin of history? :-) A related question is, if it can still improve efficiency when applied to local bindings, whether it would be worthwhile to have an additional option to switch monomorphism off for top level bindings but still keep it for local bindings, since I can't think of any situation where I don't want a top level binding to be fully polymorphic. Does anyone know of any situation where a top level binding should not be fully polymorphic? Thanks, Brian. -- Logic empowers us and Love gives us purpose. Yet still phantoms restless for eras long past, congealed in the present in unthought forms, strive mightily unseen to destroy us. http://www.metamilk.com

2 2

Re: [nightly] 07-Jun-2006 build of HEAD on i386-unknown-mingw32 (bling)
by Bulat Ziganshin 07 Jun '06

07 Jun '06

Hello sof, Wednesday, June 7, 2006, 6:27:10 PM, you wrote: > Subject: [nightly] 07-Jun-2006 build of HEAD on i386-unknown-mingw32 (bling) > mingw .tar.gz ready to be uploaded to: > http://www.haskell.org/ghc/dist/current/dist/ i don't see this upload, last mingw version is 28 march -- Best regards, Bulat mailto:Bulat.Ziganshin@gmail.com

1 0

how to get more detail for: <<loop>>
by Johannes Waldmann FbIMN 07 Jun '06

07 Jun '06

How can I get more information if my ghc-compiled program halts with "<<loop>>" What kinds of cyclic dependencies does the runtime system actually recognize? Thanks -- Johannes Waldmann

2 1