You can subscribe to this list here.
2004 |
Jan
|
Feb
(2) |
Mar
(2) |
Apr
|
May
|
Jun
(1) |
Jul
(6) |
Aug
(3) |
Sep
|
Oct
(1) |
Nov
|
Dec
(2) |
---|---|---|---|---|---|---|---|---|---|---|---|---|
2005 |
Jan
(2) |
Feb
(2) |
Mar
|
Apr
(6) |
May
|
Jun
(4) |
Jul
(3) |
Aug
|
Sep
|
Oct
(2) |
Nov
(12) |
Dec
(10) |
2006 |
Jan
(27) |
Feb
(4) |
Mar
(3) |
Apr
(5) |
May
(5) |
Jun
(1) |
Jul
(2) |
Aug
|
Sep
(7) |
Oct
(5) |
Nov
(11) |
Dec
(5) |
2007 |
Jan
(15) |
Feb
|
Mar
|
Apr
|
May
|
Jun
|
Jul
|
Aug
(3) |
Sep
(1) |
Oct
|
Nov
(1) |
Dec
|
2008 |
Jan
(7) |
Feb
(9) |
Mar
(2) |
Apr
(1) |
May
|
Jun
(6) |
Jul
(2) |
Aug
|
Sep
|
Oct
(1) |
Nov
(3) |
Dec
(1) |
2009 |
Jan
(11) |
Feb
|
Mar
(2) |
Apr
(1) |
May
(8) |
Jun
(11) |
Jul
(9) |
Aug
(12) |
Sep
(1) |
Oct
(3) |
Nov
(10) |
Dec
|
2010 |
Jan
(3) |
Feb
(1) |
Mar
(5) |
Apr
|
May
|
Jun
|
Jul
|
Aug
(2) |
Sep
(1) |
Oct
(1) |
Nov
|
Dec
|
2011 |
Jan
(2) |
Feb
(2) |
Mar
(1) |
Apr
|
May
(1) |
Jun
|
Jul
|
Aug
(1) |
Sep
|
Oct
(2) |
Nov
|
Dec
|
2012 |
Jan
|
Feb
|
Mar
|
Apr
|
May
(1) |
Jun
|
Jul
|
Aug
|
Sep
|
Oct
|
Nov
|
Dec
|
2013 |
Jan
(1) |
Feb
|
Mar
|
Apr
(3) |
May
|
Jun
|
Jul
|
Aug
|
Sep
|
Oct
(1) |
Nov
(1) |
Dec
(1) |
2014 |
Jan
|
Feb
|
Mar
|
Apr
|
May
|
Jun
|
Jul
|
Aug
|
Sep
(1) |
Oct
|
Nov
|
Dec
|
2015 |
Jan
|
Feb
|
Mar
|
Apr
|
May
|
Jun
|
Jul
(1) |
Aug
|
Sep
|
Oct
(2) |
Nov
|
Dec
|
From: Rick K. <rk...@nc...> - 2005-07-20 15:07:24
|
Alex, This is a good point and I appreciate the feedback. What's currently out there is very much a "vanilla" build system. Part of that is due to not-extensive use here of PerfSuite/psrun on the AMD64 platform (something I expect to change in the future). To answer your question: yes, supporting but 32/64 libraries would be very useful and should be in the works. Current plans are to release an 0.6.2 beta in August or September and more complete support for that would be very useful. Everything has to be done under the Autotools framework, so it might take a little bit of thinking to get it right (or "standard"). Any feedback on how things should look (in terms of the filesystem layout, for example) would be much appreciated. I'll copy the list in case others have run into the same issues. Thanks - glad to hear you are at least partially there! Rick On Mon, 18 Jul 2005, Alex Mericas wrote: > I am running Suse on an Athlon64. I can use psrun without problems on binaries compiled for 64bit, but if I try to run 32bit binaries psrun cannot find its runtime library. I'm guessing psrun would need both a 32bit and a 64bit library to work this way. Is this possible? Do you have plans to implement? Thanks! > |
From: Rick K. <rk...@nc...> - 2005-06-28 22:38:43
|
Van, Thanks for letting me know that you figured out the problem. Sounds like a simple solution and a pointer to checking environment variables if things don't work quite right. Good thing to keep in mind! Rick On Fri, 24 Jun 2005 Va...@ma... wrote: > Hi Rick, > > It looks like certain environment variables I had previously set was causing the problems. Deleting all the environment variables I set in the .soft file seem to have solved the problem. Thanks for your input. > > Van Bui > > >On Thu, 23 Jun 2005 Va...@ma... wrote: > > > >> Hi, > >> > >> I am running the psrun command on the sample programs provided in the examples directory in perfsuite. xml files only generate for the executables provided such as psrun-raw and geneventlist but not for the sample programs for which I have to compile myself such as those in cpi and others. psrun runs the program and the output looks fine but no xml files are created. I have tried many different options in attempts to see what is going wrong but have ran out of ideas. Any help would be appreciated at this point. > >> > >> Thanks, > >> > >> Van Bui > > > >Van, > > > >It's hard to tell what is happening with your tests. As I understand it, > >you are working partially on the NCSA Altix? If so, there might be > >something specific to this site that I can exchange email with you about > >off-list. > > > >Another thing that may be happening is the use of statically-linked > >executables. Some compilers (for example, Intel when using -fast) arrange > >for static linking which defeats the library preloading that psrun uses > >and result in no XML documents. Is it possible that this is going on? > >If not, then more information about your environment, compilers, flags, > >etc would be great to have in order to get you going. > > > >Rick > > > > > |
From: Rick K. <rk...@nc...> - 2005-06-24 04:35:47
|
On Thu, 23 Jun 2005 Va...@ma... wrote: > Hi, > > I am running the psrun command on the sample programs provided in the examples directory in perfsuite. xml files only generate for the executables provided such as psrun-raw and geneventlist but not for the sample programs for which I have to compile myself such as those in cpi and others. psrun runs the program and the output looks fine but no xml files are created. I have tried many different options in attempts to see what is going wrong but have ran out of ideas. Any help would be appreciated at this point. > > Thanks, > > Van Bui Van, It's hard to tell what is happening with your tests. As I understand it, you are working partially on the NCSA Altix? If so, there might be something specific to this site that I can exchange email with you about off-list. Another thing that may be happening is the use of statically-linked executables. Some compilers (for example, Intel when using -fast) arrange for static linking which defeats the library preloading that psrun uses and result in no XML documents. Is it possible that this is going on? If not, then more information about your environment, compilers, flags, etc would be great to have in order to get you going. Rick |
From: <Va...@ma...> - 2005-06-23 17:48:47
|
Hi, I am running the psrun command on the sample programs provided in the examples directory in perfsuite. xml files only generate for the executables provided such as psrun-raw and geneventlist but not for the sample programs for which I have to compile myself such as those in cpi and others. psrun runs the program and the output looks fine but no xml files are created. I have tried many different options in attempts to see what is going wrong but have ran out of ideas. Any help would be appreciated at this point. Thanks, Van Bui |
From: Rick K. <rk...@nc...> - 2005-06-18 12:15:00
|
The current issue of Linux Journal (#135, July 2005) contains an overview article about PerfSuite and its use that might be of interest to existing users or to people looking to get started with using PerfSuite for software performance analysis. This issue is on sale now and the article will be posted at the PerfSuite web site after its run in the magazine (later this year). Rick |
From: Rick K. <rk...@nc...> - 2005-04-19 00:14:38
|
Tyler - well, it seems that you've already made progress which is great news. Thanks again for the patience. I just have a couple of comments that might help clarify a thing or two: On Mon, 18 Apr 2005 ol...@ec... wrote: > In fact, tDOM was > installed correctly the entire time but I had set the --with-tdom=TDOM_PATH > option in ./configure incorrectly which was causing PerfSuite to not find the > files it needed. I see. This happens pretty frequently, usually it's the trailing "/lib" portion of the path that people leave out when configuring PerfSuite. This might be something that should be changed within the configure itself if it continues to trip people up. I'm glad to hear you got it working. > > psrun does work fine with profil.xml, but it doesn't report anything useful. > The only thing it tells us is the processor, it's cache configuration, etc. > and nothing about instruction counts, cache miss rates, or other profiling > information. Without kernel support for accessing the performance counters, it's not possible to get things like cache miss counts, etc. On x86/Linux, the kernel support comes from the perfctr patch, which is in turn used by PAPI as you know. So I would never expect an unpatched x86 kernel to provide any of the counter stuff. In that case, about the best we can do is time-based, gprof-style, profiling, which is what psrun should produce in your case when using the profil.xml configuration file. The executable has to run long enough for an interrupt/sample to occur, though... a few seconds of CPU time should be plenty. The output XML document should contain addresses and sample counts, which can be post-processed with psprocess. > Anyway with tDOM installed correctly I enthusiasticly went to run > psconfig to create our own configuration, when it spit out an error message: > > > ERROR: Unable to find PAPI > > Which I thought was completely optional. But I installed PAPI on my machine > anyway (without patching the kernel, because I don't have the priveledges to > do so) and re-configured and installed PerfSuite with the --with-papi=PAPI_DIR > option, but it still did not work. psconfig is the "least important" of the pieces, and so is not regularly updated. I think it's still a nice and useful little tool, and probably down the road will be updated so that it'll do something if hardware counter support is not present but unfortunately what you report is what you can expect from 0.6.1 (and 0.6.2). Bad documentation on this side, I'm afraid :( > > Anyway for the time being this project is being due to being a busy grad- > student, but I'll probably pick it back up sometime this summer when I have > more free-time. In regards for making distrobution specific packages (RPMs, > DEB, etc) I'd love to see that happen. I know on SourceForge you can place > help wanted ads (look in the Admin section and it should be there somewher) > and see if there are any OSS developers who can help you package it. (I'm > assuming you don't have the time to do it yourself). I'd like to help, but I'm > far too busy right now and I've never created packages for any of the programs > I've written, although I'll need to learn how to soon for another OSS project > I've been working on for over 9 months. :) > > Thanks once again for all your help. Sounds good - any and all help is welcome and appreciated, and your feedback qualifies as help already (hopefully your pain is someone else's gain!). I think you probably will see at a minimum an RPM-based release in addition to the usual .tar.gz format later on, probably for 0.6.2 final. Most likely, there will be a tDOM RPM and a PerfSuite one, and ideally you would install them in that order. I'd guess that there will be with- and without-PAPI versions, since PAPI is optional and has that kernel path dependency currently on x86 systems. Thanks again - Rick |
From: <ol...@ec...> - 2005-04-18 17:29:04
|
Hey no problem, I appreciate your willingness to help. Well I managed to make some good progress and I got tDOM installed correctly. In fact, tDOM was installed correctly the entire time but I had set the --with-tdom=TDOM_PATH option in ./configure incorrectly which was causing PerfSuite to not find the files it needed. psrun does work fine with profil.xml, but it doesn't report anything useful. The only thing it tells us is the processor, it's cache configuration, etc. and nothing about instruction counts, cache miss rates, or other profiling information. Anyway with tDOM installed correctly I enthusiasticly went to run psconfig to create our own configuration, when it spit out an error message: > ERROR: Unable to find PAPI Which I thought was completely optional. But I installed PAPI on my machine anyway (without patching the kernel, because I don't have the priveledges to do so) and re-configured and installed PerfSuite with the --with-papi=PAPI_DIR option, but it still did not work. Anyway for the time being this project is being due to being a busy grad- student, but I'll probably pick it back up sometime this summer when I have more free-time. In regards for making distrobution specific packages (RPMs, DEB, etc) I'd love to see that happen. I know on SourceForge you can place help wanted ads (look in the Admin section and it should be there somewher) and see if there are any OSS developers who can help you package it. (I'm assuming you don't have the time to do it yourself). I'd like to help, but I'm far too busy right now and I've never created packages for any of the programs I've written, although I'll need to learn how to soon for another OSS project I've been working on for over 9 months. :) Thanks once again for all your help. - Tyler Quoting Rick Kufrin <rk...@nc...>: > Tyler, > > First of all, apologies for delay in my reply. I'm currently on travel > and not able to read or respond as quickly as I'd prefer, so thanks > for your patience. > > It's possible that there are a couple of things going on here, or rather > *not* going on. It does seem as if tDOM is a roadblock from what you > report. An improper tDOM install will prevent psprocess (or psconfig, > as you note from your original message) from being able to work with > either the input (XML config) or output (perf measurement) documents > for psrun. > > Although I'm not certain, it seems as though the "profil.xml" file as you > show does not result in a psrun runtime error, although all others do, > since they rely on PAPI. If you receive an XML output file when using > profil.xml and psrun on your program, then that likely means that psrun > itself is OK - the output document should contain profiling samples if > so. Unfortunately, psprocess (and therefore tDOM) would be required > to map the samples in the XML doc to source code lines for the output > to be of much use. > > Can you verify that XML is output when you run psrun with the profil.xml > config? > > If the psrun/profil.xml combination does produce output, then I think > that points again to tDOM being the missing part in your install. > I'd be happy to help you with the tDOM build, probably off-list is > best. Let me know if that might help (I am still on travel, although > will check more frequently). > > I have thought recently that providing RPMs for x86/x86-64/ia64 tDOM > might make it easier for people to install and use PerfSuite and avoid > this type of problem, any feedback is welcome. > > Rick > > On Tue, 12 Apr 2005 ol...@ec... wrote: > > > Thanks for replying. Unfortunately I have already tried what you mentioned > and > > had no luck. First of all, the "profil.xml" file contains the following: > > > > <?xml version="1.0" encoding="UTF-8" ?> > > <ps_hwpc_profile class="null"> > > <!-- ========================================================== > > This is a configuration file that can be used to ask for > > profiling, based on profil(). > > > > $Id: profil.xml,v 1.3 2003/12/27 14:23:51 rkufrin Exp $ > > ========================================================== --> > > </ps_hwpc_profile> > > > > > > It is exactly the same as the "null.xml" file (using Perfsuite-0.6.1 by > the > > way). And if we try to use any of the other config files (all prefixed > with > > papi) psrun will refuse to execute because the kernel we are running does > not > > have PAPI patched to it. > > > > > > So really, we are in a situation where we have no valid config files to use > and > > are unable to build our own config files. I've spent nearly 2 days trying > to > > build tDOM and I have read all the documentation provided with it, but no > matter > > what I've tried it still builds incorrectly. If there are any other > general > > config files that may be laying around somewhere (especially for Xeon) that > do > > *not* use PAPI or a guide somewhere to explain the format of the config > file so > > we could build one by hand, it would really help us out. Thanks once > again. > > > > > > - Tyler > > > > ------------------------------------------------------- > SF email is sponsored by - The IT Product Guide > Read honest & candid reviews on hundreds of IT Products from real users. > Discover which products truly live up to the hype. Start reading now. > http://ads.osdn.com/?ad_id=6595&alloc_id=14396&op=click > _______________________________________________ > PerfSuite-users mailing list > Per...@li... > https://lists.sourceforge.net/lists/listinfo/perfsuite-users > ------------------------------------------------- This mail sent through IMP: http://horde.org/imp/ |
From: Rick K. <rk...@nc...> - 2005-04-18 14:57:54
|
Tyler, First of all, apologies for delay in my reply. I'm currently on travel and not able to read or respond as quickly as I'd prefer, so thanks for your patience. It's possible that there are a couple of things going on here, or rather *not* going on. It does seem as if tDOM is a roadblock from what you report. An improper tDOM install will prevent psprocess (or psconfig, as you note from your original message) from being able to work with either the input (XML config) or output (perf measurement) documents for psrun. Although I'm not certain, it seems as though the "profil.xml" file as you show does not result in a psrun runtime error, although all others do, since they rely on PAPI. If you receive an XML output file when using profil.xml and psrun on your program, then that likely means that psrun itself is OK - the output document should contain profiling samples if so. Unfortunately, psprocess (and therefore tDOM) would be required to map the samples in the XML doc to source code lines for the output to be of much use. Can you verify that XML is output when you run psrun with the profil.xml config? If the psrun/profil.xml combination does produce output, then I think that points again to tDOM being the missing part in your install. I'd be happy to help you with the tDOM build, probably off-list is best. Let me know if that might help (I am still on travel, although will check more frequently). I have thought recently that providing RPMs for x86/x86-64/ia64 tDOM might make it easier for people to install and use PerfSuite and avoid this type of problem, any feedback is welcome. Rick On Tue, 12 Apr 2005 ol...@ec... wrote: > Thanks for replying. Unfortunately I have already tried what you mentioned and > had no luck. First of all, the "profil.xml" file contains the following: > > <?xml version="1.0" encoding="UTF-8" ?> > <ps_hwpc_profile class="null"> > <!-- ========================================================== > This is a configuration file that can be used to ask for > profiling, based on profil(). > > $Id: profil.xml,v 1.3 2003/12/27 14:23:51 rkufrin Exp $ > ========================================================== --> > </ps_hwpc_profile> > > > It is exactly the same as the "null.xml" file (using Perfsuite-0.6.1 by the > way). And if we try to use any of the other config files (all prefixed with > papi) psrun will refuse to execute because the kernel we are running does not > have PAPI patched to it. > > > So really, we are in a situation where we have no valid config files to use and > are unable to build our own config files. I've spent nearly 2 days trying to > build tDOM and I have read all the documentation provided with it, but no matter > what I've tried it still builds incorrectly. If there are any other general > config files that may be laying around somewhere (especially for Xeon) that do > *not* use PAPI or a guide somewhere to explain the format of the config file so > we could build one by hand, it would really help us out. Thanks once again. > > > - Tyler |
From: <ol...@ec...> - 2005-04-12 20:52:19
|
Thanks for replying. Unfortunately I have already tried what you mentioned and had no luck. First of all, the "profil.xml" file contains the following: <?xml version="1.0" encoding="UTF-8" ?> <ps_hwpc_profile class="null"> <!-- ========================================================== This is a configuration file that can be used to ask for profiling, based on profil(). $Id: profil.xml,v 1.3 2003/12/27 14:23:51 rkufrin Exp $ ========================================================== --> </ps_hwpc_profile> It is exactly the same as the "null.xml" file (using Perfsuite-0.6.1 by the way). And if we try to use any of the other config files (all prefixed with papi) psrun will refuse to execute because the kernel we are running does not have PAPI patched to it. So really, we are in a situation where we have no valid config files to use and are unable to build our own config files. I've spent nearly 2 days trying to build tDOM and I have read all the documentation provided with it, but no matter what I've tried it still builds incorrectly. If there are any other general config files that may be laying around somewhere (especially for Xeon) that do *not* use PAPI or a guide somewhere to explain the format of the config file so we could build one by hand, it would really help us out. Thanks once again. - Tyler Quoting Rick Kufrin <rk...@nc...>: > Tyler - hopefully it is simple to get you past this hurdle, although > as you point out, documentation for psconfig is lacking (sorry). > > In general, psconfig isn't really necessary to do performance measurement > with psrun or the API, whether you have PAPI or not. If the installation > completes successfully, "psrun <somecommand>" should run without errors, > and you should be able to use psprocess on the resulting XML document > to view the data. If the XML doc contains profiling data, you'll > also have to supply the name of the executable to psprocess, with the > option -e. > > Alternate XML config files, including samples without PAPI, should > be installed in the directory PREFIX/share/perfsuite/xml/pshwpc, > and there is a sample config file called "profil.xml" that might > help out here. You might try: > > psrun -c profil.xml a.out > psprocess -e a.out yourxmlfile.xml > > You might need to supply the full path to profil.xml for things to > work properly with your install. > > Again, use of psconfig is completely optional and not necessary. > However, there isn't enough info here to see if in fact your tDOM > install isn't correct. If it isn't, then psprocess will have problems > (probably will not be able to find/load tDOM). psrun doesn't use > tDOM, so the use of an alternate XML config is independent of tDOM. > > The INSTALL file has some information about install/build of tDOM and > use of "configure" that might help see what might be the problem as > well. If still stuck, then some sample error output might help and > also the exact configure line you used (it should be stored in > config.log). > > Hope this helps track things down, > > Rick > > On Thu, 7 Apr 2005 ol...@ec... wrote: > > > I'm trying to do some performance analysis on a quad-processor Intel Xeon > > 1.80Ghz machine. I've spent nearly the last two weeks trying to build a > > configuration input file for perfsuite for this machine and psconfig > refuses to > > run because it detects a problem with the tDOM installation. I've done > > absolutely everything I can to try and install tDOM properly, but it just > will > > not work correctly and the documentation for that software package is > absolutely > > horrible, so I have no idea how to fix it. > > > > > > What I would like to know is, is there an easy way to build a configuration > file > > *without* using psconfig? If not, is there some documentation about how > the > > configuration files are structured, because at this point it would be > much > > easier/faster to just write the XML file myself by hand. (An example > > configuration file *without* PAPI would be wonderful to look at). Thank you > for > > any help you can provide. > > > > > > - Tyler Olsen > > > > ------------------------------------------------------- > SF email is sponsored by - The IT Product Guide > Read honest & candid reviews on hundreds of IT Products from real users. > Discover which products truly live up to the hype. Start reading now. > http://ads.osdn.com/?ad_id=6595&alloc_id=14396&op=click > _______________________________________________ > PerfSuite-users mailing list > Per...@li... > https://lists.sourceforge.net/lists/listinfo/perfsuite-users > ------------------------------------------------- This mail sent through IMP: http://horde.org/imp/ |
From: Rick K. <rk...@nc...> - 2005-04-09 15:29:46
|
Tyler - hopefully it is simple to get you past this hurdle, although as you point out, documentation for psconfig is lacking (sorry). In general, psconfig isn't really necessary to do performance measurement with psrun or the API, whether you have PAPI or not. If the installation completes successfully, "psrun <somecommand>" should run without errors, and you should be able to use psprocess on the resulting XML document to view the data. If the XML doc contains profiling data, you'll also have to supply the name of the executable to psprocess, with the option -e. Alternate XML config files, including samples without PAPI, should be installed in the directory PREFIX/share/perfsuite/xml/pshwpc, and there is a sample config file called "profil.xml" that might help out here. You might try: psrun -c profil.xml a.out psprocess -e a.out yourxmlfile.xml You might need to supply the full path to profil.xml for things to work properly with your install. Again, use of psconfig is completely optional and not necessary. However, there isn't enough info here to see if in fact your tDOM install isn't correct. If it isn't, then psprocess will have problems (probably will not be able to find/load tDOM). psrun doesn't use tDOM, so the use of an alternate XML config is independent of tDOM. The INSTALL file has some information about install/build of tDOM and use of "configure" that might help see what might be the problem as well. If still stuck, then some sample error output might help and also the exact configure line you used (it should be stored in config.log). Hope this helps track things down, Rick On Thu, 7 Apr 2005 ol...@ec... wrote: > I'm trying to do some performance analysis on a quad-processor Intel Xeon > 1.80Ghz machine. I've spent nearly the last two weeks trying to build a > configuration input file for perfsuite for this machine and psconfig refuses to > run because it detects a problem with the tDOM installation. I've done > absolutely everything I can to try and install tDOM properly, but it just will > not work correctly and the documentation for that software package is absolutely > horrible, so I have no idea how to fix it. > > > What I would like to know is, is there an easy way to build a configuration file > *without* using psconfig? If not, is there some documentation about how the > configuration files are structured, because at this point it would be much > easier/faster to just write the XML file myself by hand. (An example > configuration file *without* PAPI would be wonderful to look at). Thank you for > any help you can provide. > > > - Tyler Olsen |
From: <ol...@ec...> - 2005-04-07 22:06:59
|
I'm trying to do some performance analysis on a quad-processor Intel Xeon 1.80Ghz machine. I've spent nearly the last two weeks trying to build a configuration input file for perfsuite for this machine and psconfig refuses to run because it detects a problem with the tDOM installation. I've done absolutely everything I can to try and install tDOM properly, but it just will not work correctly and the documentation for that software package is absolutely horrible, so I have no idea how to fix it. What I would like to know is, is there an easy way to build a configuration file *without* using psconfig? If not, is there some documentation about how the configuration files are structured, because at this point it would be much easier/faster to just write the XML file myself by hand. (An example configuration file *without* PAPI would be wonderful to look at). Thank you for any help you can provide. - Tyler Olsen ------------------------------------------------- This mail sent through IMP: http://horde.org/imp/ |
From: Rick K. <rk...@nc...> - 2005-02-04 15:39:51
|
> Hi All, > > I am planning to use PSPMPI to profile some MPI application and have > downloaded > http://perfsuite.ncsa.uiuc.edu/download/perfsuite-0.6.1.tar.bz2. > > I am not able to find libpspmpi (the other companions, libperfsuite and > libpshwpc). > > It would be nice if someone could tell me how to obtain libpspmpi > > thanks and regards > santhosh > Santhosh, That's an understandable confusion, based on the material posted at the perfsuite.ncsa web site. There is some background information about development and release plans in a message from the perfsuite bugs list here: http://sourceforge.net/mailarchive/forum.php?thread_id=6310500&forum_id=42219 The short story is that versions 0.6.x focus on hardware counting and profiling, not MPI (although PerfSuite does have a PMPI-based interface for performance measurement, but that for hardware counters still, not communication profiling). PSPMPI has existed for a while now, but needs some work, testing, documentation before it can be released. In the meantime, I would recommend a library called mpiP from LLNL and ORNL, which has some intersection in functionality with PSPMPI. You can find a URL here: http://perfsuite.ncsa.uiuc.edu/perftools/ There will be a 0.6.2 alpha release of PerfSuite very shortly, but that will also continue on the hardware counter focus. Sorry for the confusion, Rick |
From: V S. K. <gv...@ri...> - 2005-02-04 14:47:26
|
Hi All, I am planning to use PSPMPI to profile some MPI application and have downloaded http://perfsuite.ncsa.uiuc.edu/download/perfsuite-0.6.1.tar.bz2. I am not able to find libpspmpi (the other companions, libperfsuite and libpshwpc). It would be nice if someone could tell me how to obtain libpspmpi thanks and regards santhosh |
From: Rick K. <rk...@nc...> - 2005-01-18 01:03:08
|
Vangal, I'll copy the mailing list on this one because I think it's a good question and a little hard to try to figure out based on the error message you got, which is not especially descriptive. I believe what is going on here is that you are trying to profile using a derived PAPI event. On Itanium 2, PAPI defines PAPI_TLB_TL as the sum of the two native events ITLB_MISSES_FETCH_L2ITLB and L2DTLB_MISSES. Although recent releases of PAPI (I think) support profiling based on derived events, this is not possible when profiling through PerfSuite, and that's why you're getting that error. There is nothing wrong in principle with the XML configuration you're using, though (just the event). To double check on this, you might try profiling on either PAPI_TLB_DM or PAPI_TLB_IM and see if you get past the error. Those two are not derived events (they're just one of the events above). Rick On Mon, 17 Jan 2005, Venkatesh, Vangal wrote: > Rick, > Do you know if I can do a profile of TLB misses and correspond to > source? I modified the papi_profile_cycles.xml files for TLB misses: > > I tried > <?xml version="1.0" encoding="UTF-8" ?> > <ps_hwpc_profile class="PAPI"> > <!-- ========================================================== > This is a configuration file that can be used to ask for > profiling, based on total cycles with a period of 100,000. > You will probably want to adjust the period based on your > event selection and also on your processor. There are no > particular recommended values. > > $Id: papi_profile_cycles.xml,v 1.1 2004/01/12 01:01:43 rkufrin > Exp $ > ========================================================== --> > <ps_hwpc_event type="preset" name="PAPI_TLB_TL" threshold="100000"/> > </ps_hwpc_profile> > > And got > psrun fatal error: error reported by PAPI > > Vangal > |
From: Rick K. <rk...@nc...> - 2005-01-07 02:31:20
|
PerfSuite 0.6.1 final is now available. This release contains all the features, enhancements, and bug fixes released to date. There are just a few changes from the 0.6.1rc1 release, most notably a change in handling of existing LD_PRELOAD settings for psrun, and a few adjustments in the configure/build process. Please see the ChangeLog for details. URL: http://perfsuite.ncsa.uiuc.edu/download/ http://www.sourceforge.net/projects/perfsuite/ This is the same version that is in production on NCSA x86/ia64 Linux systems and now has slightly under two years of testing in a production environment, so it's moved from beta status to stable. Work will be ongoing to move to a new 0.6.2 version that will start in a feature addition (alpha) mode, which will be a good opportunity to provide feedback as well as any requests for new features. Mailing list announcements for 0.6.2 releases will be posted to perfsuite-announce and to SourceForge project news. Rick |
From: Rick K. <rk...@nc...> - 2004-12-15 22:16:07
|
This is a heads-up note to alert you that there will be test suite runtime errors in the recent release of PerfSuite (0.6.1rc1) if built with PAPI 3.0.7 using the Pentium III or Itanium PAPI substrates. These are minor errors that arise from the way that PAPI sets the default counting domain (user/kernel/all) for event sets. They do not affect the overall use of psrun or libpshwpc. You can safely ignore these errors and run the entire PerfSuite test suite to completion with "make -k check". A similar problem was identified and corrected in PAPI with the Pentium 4 PAPI substrate and has been entered in the PAPI 3 bug tracking system for these other platforms so it's not expected to cause any problems in the future (but might be a bit disconcerting for anyone building today). Rick |
From: Rick K. <rk...@nc...> - 2004-12-11 17:00:26
|
PerfSuite 0.6.1 release candidate 1 is now available. This release contains all the features, enhancements, and bug fixes released to date. Barring serious problems uncovered in 0.6.1rc1, this will be the version released as PerfSuite 0.6.1 final. Most of the changes in this release address bugs within the PerfSuite libraries or tools, and there are few user-visible changes in the tools. However, it is highly recommended that you update to this release for increased stability and compatibility with the latest versions of PAPI 3. There are some internal changes in library data structures that require recompilation/relink for those using the PerfSuite API directly. These should require no source code changes to the calling code. Please see the ChangeLog for details. URL: http://perfsuite.ncsa.uiuc.edu/download/ http://www.sourceforge.net/projects/perfsuite/ I'm cross-posting this to both perfsuite-announce and perfsuite-users in order to reach as many people as possible. The next version (0.6.2) of PerfSuite, which will see releases in early 2005, will return to postings of alpha releases to only the perfsuite-announce mailing list. Rick |
From: Rick K. <rk...@nc...> - 2004-10-02 16:04:43
|
PerfSuite version 0.6.1 beta 6 is now available. This release adds several new features, improvements, and bug fixes. The most visible changes from the user's perspective include: - psrun output XML documents are now created using the name of the program that was monitored instead of the prefix "psrun" - Fortran support for PMPI-based hardware performance measurements of MPI programs without modifying your source code - improved support for the SGI Altix platform by using the mmtimer interface for wall-clock time measurement (enabled during package configuration) - improved support for generating VProf-compatible profiles with psprocess - several changes in the default events used on Pentium 4 platforms as current available through PAPI 3 Additionally, this release restores compatibility with all versions of PAPI 3 beta available to date, including the current CVS version (changes in PAPI 3 since PerfSuite 0.6.1 beta 5 was released made them incompatible) Please see the ChangeLog for details. URL: http://perfsuite.ncsa.uiuc.edu/download/ http://www.sourceforge.net/projects/perfsuite/ Rick |
From: Rick K. <rk...@nc...> - 2004-08-24 18:24:55
|
On Tue, 24 Aug 2004, Venkatesh, Vangal wrote: > Rick, > May I ask how to build perfsuite with intel compilers? I think I am > going to get lot of incompatibilities since Fortran compilers are not > compatible. > > Vangal > Vangal, Certainly, that's a good question. We use Intel compilers all the time here (almost exclusively on IA-64 systems). The issue is not so much whether PerfSuite is compiled by the Intel compiler (as far as I know, it can't be because of asm statements), but whether it follows the same name mangling convention (C/Fortran, I mean). That's necessary to be able to link if you're using the Fortran API. The thing to do is to set F77=ifc (or efc) on the "configure" line when you build PerfSuite - it should then follow the ifc naming convention for the versions of the API routines that are for Fortran. We generally build both g77 and Intel versions of the libraries and separate the two by using the configure option "--libdir=<dir>" option to specify a particular destination directory for the Intel-compatible version of the library (e.g., /opt/perfsuite/lib/intel). That takes two configures and builds - one with g77, the other with ifc/efc. For both versions, gcc is the primary compiler for PerfSuite. Please let me know if this is not clear or if you run into any problems. Rick p.s. I have been looking into the ch_p4 issue you mentioned and it does seem that the problem is that the commands constructed and executed when using MPICH ch_p4 are different that what is supplied to mpirun. MPI_Init sets things back in order, but that's after psrun has had its chance to do its job (the command line given to psrun does not appear to contain the name of the executable that's to be measured). Still looking into possible solutions. |
From: Rick K. <rk...@nc...> - 2004-08-24 00:34:17
|
More info on psrun and MPICH/ch_p4... has anyone had success with this? Rick ---------- Forwarded message ---------- Date: Mon, 23 Aug 2004 19:31:01 -0500 (CDT) From: Rick Kufrin <rk...@os...> To: "Venkatesh, Vangal" <van...@in...> Subject: RE: Psrun with argonne mpich Vangal, Yes, I see that I can reproduce the same behavior here, using a ch_p4 version of MPICH. I'm guessing it's due to the way that the commands are constructed and eventually launched on the remote nodes, but unfortunately I don't know an easy way to invoke mpirun or mpirun.ch_p4 to get around it. We typically use a different MPICH (MPICH-GM) here, so probably I was confusing with that? One thing you can do is to change your source code to include a call to the library that psrun uses... not sure if you're willing to do that. If you are, then what you'd have to do is include a call to the following routine, probably right after MPI_Init is best: #include <pshwpc.h> int ps_hwpc_psrun(void); (C) or include 'fperfsuite.h' subroutine psf_hwpc_psrun(ierr) (Fortran) Successful returns from either of these two routines are PS_SUCCESS (or 0). That should be the only change to your source, but you would have to also arrange for the -I and -L options at link time. The libraries you'd want to include are -lpshwpc and possibly -lperfsuite and -lpapi, depending on your system/environment. With the library linked in, you can forget about psrun and just run the executable as normal. All the environment variables that psrun recognizes should work as usual. If you try this and still have problems, let me know (if it works for you, that would be good to know too!) In the meantime, I'll see if I can figure out an alternate way to use psrun rather than having to go through changing source code and let you know if something comes up. Rick On Mon, 23 Aug 2004, Venkatesh, Vangal wrote: > Rick, > I am getting the following error message: > [vvenkat1@spd206-6 run_dir]$ ../../mpich-ia32e/mpich-1.2.5.2/bin/mpirun > -np 2 / > usr/local/bin/psrun /home/spd/vvenkat1/pop1.4/run_dir/pop2 > spd206-6: No such file or directory > p0_10443: p4_error: Child process exited while making connection to > remote proc > ess on spd206-6: 0 > > Some how it thinks spd206-6 is a file (it is a machine in my > machinefile). I also tried putting in the -machinefile option. > > Vangal > |
From: Rick K. <rk...@nc...> - 2004-08-23 23:08:55
|
Vangal, Yes, it should be possible to use psrun with MPICH. Some of the things that may be required are to use the full path name to the psrun executable and also to make sure that the environment settings such as LD_LIBRARY_PATH are set properly, especially if you're using a variant of MPICH like ch_p4. One way to set that up is to add it to your dot-files like .cshrc or .profile, that way you'll be sure that the ssh to the remote nodes picks up the right values. Just to be clear, the command line might be something like: mpirun -np 8 -machinefile machines /usr/local/bin/psrun mpi_program If these suggestions don't work, then please send any error output you can get from the run and hopefully can figure it out... Rick On Mon, 23 Aug 2004, Venkatesh, Vangal wrote: > Hi, > Is it possible to use psrun with argonne mpich. I can't seem to > figure out the syntax. > > Vangal > |
From: Rick K. <rk...@nc...> - 2004-07-27 14:16:33
|
Stelios - just another comment as I realized I might have said something wrong. Actually you can get a mismatch error message from within the PerfSuite Tcl C extension for PAPI (which is what psprocess uses). There is a simple way to see if this is the case, by compiling a test program papitest.c something like this: #include <stdio.h> #include <papi.h> main() { int ret; ret = PAPI_library_init(PAPI_VER_CURRENT); printf("ret = %d, PAPI_VER_CURRENT = %d\n", ret, PAPI_VER_CURRENT); } cc -o papitest papitest.c -I<PAPIDIR>/include -L<PAPIDIR>/lib -lpapi (<PAPIDIR> is the top-level of your PAPI installation) If the two numbers don't come out the same, this could be the source of the error. If that's the case, I think it can be fixed relatively easily. If you try this, please let me know the result... Rick On Mon, 26 Jul 2004, Stelios Kyriacou wrote: > > Hi > > I am a new user to this group. We are trying to get latest perfsuite with > the latest papi. I know that this is still alpha but was wondering if > anybody has got it to work and what would be requierd. Right now the > error i get is that there is a "PAPI version mismatch" error that does not > allow pspapi to be loaded in tcl. > > Also: is there a CVS available for perfsuite? I would like to try the > latest version possible just in case there have been any improvements to > the PAPI interface. > > The reason we would like to stick with the latest papi is that it seems to > to work with the latest version of perfctr which in turn seems to work > with our specific kernel (earlier version of perfctr do not seem to patch > our kernel correctly). > > Thanks > > Stelios Kyriacou > Systems Admin > Ohio Supercomputing Center > > > > ------------------------------------------------------- > This SF.Net email is sponsored by BEA Weblogic Workshop > FREE Java Enterprise J2EE developer tools! > Get your free copy of BEA WebLogic Workshop 8.1 today. > http://ads.osdn.com/?ad_id=4721&alloc_id=10040&op=click > _______________________________________________ > PerfSuite-users mailing list > Per...@li... > https://lists.sourceforge.net/lists/listinfo/perfsuite-users > |
From: Rick K. <rk...@nc...> - 2004-07-27 14:05:05
|
Stelios, Sorry to hear you're having difficulties with getting things going on your system. The "PAPI version mismatch" comes from a call to initialize PAPI within the "libpshwpc" library in PerfSuite. It is meant to detect a conflict between PAPI versions 2 and 3 but that message will also come out if PAPI/perfctr fail to initialize. I'm not sure what specific version of perfctr and PAPI you are using, I can tell you that we run PAPI 3 with perfctr 2.6.5 on our Xeon systems and also that I've been working with perfctr 2.6.8 on a different (PIII) machine. We have not tried with any of the perfctr 2.7 things that have been released lately. Also, we use both the PAPI beta 3 and a CVS version of PAPI from June. You mention that pspapi being loaded into Tcl is the symptom: does that mean that psrun with hardware counting is working (that is, do you get XML documents with counts)? Another possibility may be an unintended libpapi.so being loaded by the Tcl extension that psprocess uses - but that seems pretty remote. As it happens, I have access to OSC systems thanks to the folks there and so possibly could help you out directly, but it would be a few weeks as I am away for a while (can take this off-list). Re: CVS - no, there is no CVS repository available. The 0.6.1b5 version of PerfSuite is quite recent and hopefully the last version before a final 0.6.1 targeted for the fall. The reason why no CVS is mainly to avoid having people update frequently, which can be pretty time-consuming. So the effort is to try to make reasonably well-tested releases available on a less-frequent basis. Rick p.s. Perfctr's author is very good/prompt at providing support for patching a number of different kernel versions so it's possible that if the patch procedure didn't work previously, that there may be a more recent version that would. The perfctr mailing list has announcements that are archived at http://www.sf.net/projects/perfctr/ On Mon, 26 Jul 2004, Stelios Kyriacou wrote: > > Hi > > I am a new user to this group. We are trying to get latest perfsuite with > the latest papi. I know that this is still alpha but was wondering if > anybody has got it to work and what would be requierd. Right now the > error i get is that there is a "PAPI version mismatch" error that does not > allow pspapi to be loaded in tcl. > > Also: is there a CVS available for perfsuite? I would like to try the > latest version possible just in case there have been any improvements to > the PAPI interface. > > The reason we would like to stick with the latest papi is that it seems to > to work with the latest version of perfctr which in turn seems to work > with our specific kernel (earlier version of perfctr do not seem to patch > our kernel correctly). > > Thanks > > Stelios Kyriacou > Systems Admin > Ohio Supercomputing Center > > > > ------------------------------------------------------- > This SF.Net email is sponsored by BEA Weblogic Workshop > FREE Java Enterprise J2EE developer tools! > Get your free copy of BEA WebLogic Workshop 8.1 today. > http://ads.osdn.com/?ad_id=4721&alloc_id=10040&op=click > _______________________________________________ > PerfSuite-users mailing list > Per...@li... > https://lists.sourceforge.net/lists/listinfo/perfsuite-users > |
From: Stelios K. <kyr...@os...> - 2004-07-26 18:44:08
|
Hi I am a new user to this group. We are trying to get latest perfsuite with the latest papi. I know that this is still alpha but was wondering if anybody has got it to work and what would be requierd. Right now the error i get is that there is a "PAPI version mismatch" error that does not allow pspapi to be loaded in tcl. Also: is there a CVS available for perfsuite? I would like to try the latest version possible just in case there have been any improvements to the PAPI interface. The reason we would like to stick with the latest papi is that it seems to to work with the latest version of perfctr which in turn seems to work with our specific kernel (earlier version of perfctr do not seem to patch our kernel correctly). Thanks Stelios Kyriacou Systems Admin Ohio Supercomputing Center |
From: <rk...@nc...> - 2004-07-05 15:48:14
|
Kursad, Thanks for the update on how things are going. It sounds like you're making some progress, I hope the comments were helpful. I'd like to make sure I'm understanding your experience because as you know, PerfSuite is still in a beta release. It would be nice to adjust things/problems before going final later this year. So to recap: you are saying that PIII configs that do not include derived events work for you, is that right? It seems to me that if counter resources are not available, that psrun shouldn't even write out an XML document, but instead report an error. Does that seem like more reasonable behavior? P4 - same thing? If you could send me an example config that shows this behavior, it would help me test/verify here on our systems. The Perl script idea sounds good: there's no Perl anything currently in the PerfSuite distribution but if you develop a post-processor that might be interesting to others, I'd love to hear of it. Also - if you're finding there are particular metrics you're using frequently, you may be able to add them to the default calculations that psprocess does by adding them into its "database". That's an XML document that gets installed here: $PREFIX/share/perfsuite/xml/pshpwc/PAPI_metrics.xml Thanks again for the update, Rick > Hi Rick, > > Looks like this was an issue that you were aware of. When I don't pass any > XML configuration file to psrun, it works perfectly. Turns out that the PAPI > 3 standard config file for Pentium 3 (P6) includes some comments from you on > why no derived events are included in the XML file. This config file works > for us, we will just have to derive the derived events ourselves by writing > a Perl script. > > What I tried to do was write my own config file that included all standard > and derived metrics. PAPI 3 still fails and gives me all zeros if I use such > a file. > > Regards, |