Menu
▾
▴
perfsuite-users
|
From: Rick K. <rk...@il...> - 2009-07-22 12:30:36
|
Cheng, Thanks for the followup and further info. I'm glad to hear that you are getting better results with the itimer.xml configuration versus profil.xml. I have seen problems with profil.xml before, which uses the C library profil() functionality, but they had seemed to be isolated to Itanium platforms. Regardless, this is something we will note and see if we can reproduce. Thanks again for the info - I am copying the SourceForge list for archival/status purposes. Rick ----- Original Message ----- From: "Cheng Liao" <cyl...@gm...> To: "Rick Kufrin" <rk...@il...> Sent: Tuesday, July 21, 2009 11:30:08 PM GMT -06:00 US/Canada Central Subject: Re: [PerfSuite-users] psrun profile doesn't show any sampling point Rick, Thought I would follow up. Running longer jobs helps, since not all CPU time would be reported by profiling samples. However, my real problem is I need to use itimer.xml instead of profil.xml for psrun. For some reason profil.xml would show empty. Maybe I have messed up my build somewhere? Now, everything looks great, except sometime a significant chuck of the sampling points cannot find the function names and would show ?? on the output. Thanks again, Cheng On Mon, Jul 20, 2009 at 1:28 PM, Cheng Liao < cyl...@gm... > wrote: Hi Rick, Thanks for your quick reply. Indeed the sample output I showed was from a Nehalem machine. However, I saw the same problem on some wolfdale and harpertown systems too. So far, I have tried only a 0 second and a 6 second jobs. Will make some longer runs. Cheng On Mon, Jul 20, 2009 at 12:50 PM, Rick Kufrin < rk...@il... > wrote: Cheng, There are two issues that I can think of that you might be experiencing. The first is simply that the program may not run long enough to generate profiling samples. However, I am also wondering if (based on the processor brand string listed in the output) if you are working with an Intel Nehalem CPU. If so, there are known problems with the most recent release of PerfSuite on that processor. We have finished adding the support for that system, but have not yet placed a new software release on SourceForge or at NCSA. We hope to do so in the very near future, though. Sorry for the difficulties, please let me know if you also suspect this may be what is affecting you. Rick Cheng Liao wrote: I just installed perfsuite on some linux x86_64 machines (only one with papi and perfctr). However, even though 'make -s check' outputs looked fine, all the psprocess outputs would be 'empty' and show no profile timing info. Is this something that can be easily fixed? Thanks, Cheng PerfSuite Hardware Performance Summary Report Version : 1.0 Created : Mon Jul 20 11:20:26 AM PDT 2009 Generator : psprocess 0.3 XML Source : ls.15381.em64td.xml Execution Information ============================================================================================ Collector : libpshwpc Date : Sat Jul 18 17:31:54 2009 Host : em64td User : cliao Command : ls Processor and System Information ============================================================================================ Node CPUs : 8 Vendor : Intel Family : Pentium Pro (P6) Brand : Intel(R) Xeon(R) CPU X5570 @ 2.93GHz CPU Revision : 4 Clock (MHz) : 2933.589 Memory (MB) : 24096.66 Pagesize (KB) : 4 Cache Information ============================================================================================ Cache levels : 0 Profile Information ============================================================================================ Class : profil Version : 2.5 Event : milliseconds Period : 10 Samples : 0 Domain : user Run Time : 0.00 (seconds) Min Self % : (all) Module Summary -------------------------------------------------------------------------------- Samples Self % Total % Module File Summary -------------------------------------------------------------------------------- Samples Self % Total % File Function Summary -------------------------------------------------------------------------------- Samples Self % Total % Function Function:File:Line Summary -------------------------------------------------------------------------------- Samples Self % Total % Function:File:Line ------------------------------------------------------------------------ ------------------------------------------------------------------------------ Enter the BlackBerry Developer Challenge This is your chance to win up to $100,000 in prizes! For a limited time, vendors submitting new applications to BlackBerry App World(TM) will have the opportunity to enter the BlackBerry Developer Challenge. See full prize details at: http://p.sf.net/sfu/Challenge ------------------------------------------------------------------------ _______________________________________________ PerfSuite-users mailing list Per...@li... https://lists.sourceforge.net/lists/listinfo/perfsuite-users |
|
From: Cheng L. <cyl...@gm...> - 2009-08-19 16:24:58
|
Rick,
The large ?? entry in my profile is caused by an incorrect offset for
libc.so.6 base
address. I need to manually set the libc offset to zero in the xml file,
then run psprocess, to
fix the problem.
A new problem I have encountered is for large jobs, the bin for the number
of sampling points would overflow:
<sample pc="59d770">27665</sample>
<sample pc="59d772">26940</sample>
<sample pc="59d776">27564</sample>
<sample pc="59d77a">-31502</sample>
<---------- negative
<sample pc="59d77e">11502</sample>
<sample pc="59d782">5323</sample>
<sample pc="59d78a">21065</sample>
<sample pc="59d78e">32422</sample>
<sample pc="59d790">27709</sample>
<sample pc="59d794">-30384</sample>
<------------- negative
<sample pc="59d798">24312</sample>
Does perfsuite support a large sampling point bin, or using a higher
profiling overflow threhold would be the only solution?
Thanks,
Cheng
On Wed, Jul 22, 2009 at 5:30 AM, Rick Kufrin <rk...@il...> wrote:
> Cheng,
>
> Thanks for the followup and further info. I'm glad to hear that you are
> getting better results with the itimer.xml configuration versus profil.xml.
> I have seen problems with profil.xml before, which uses the C library
> profil() functionality, but they had seemed to be isolated to Itanium
> platforms. Regardless, this is something we will note and see if we can
> reproduce.
>
> Thanks again for the info - I am copying the SourceForge list for
> archival/status purposes.
>
> Rick
>
> ----- Original Message -----
> From: "Cheng Liao" <cyl...@gm...>
> To: "Rick Kufrin" <rk...@il...>
> Sent: Tuesday, July 21, 2009 11:30:08 PM GMT -06:00 US/Canada Central
> Subject: Re: [PerfSuite-users] psrun profile doesn't show any sampling
> point
>
> Rick,
>
> Thought I would follow up.
>
> Running longer jobs helps, since not all CPU time would be reported by
> profiling
> samples. However, my real problem is I need to use itimer.xml instead of
> profil.xml for psrun. For some reason profil.xml would show empty. Maybe I
> have
> messed up my build somewhere?
>
> Now, everything looks great, except sometime a significant chuck of the
> sampling points cannot find the function names and would show ?? on
> the output.
>
> Thanks again,
> Cheng
>
>
> On Mon, Jul 20, 2009 at 1:28 PM, Cheng Liao < cyl...@gm... > wrote:
>
>
>
> Hi Rick,
>
> Thanks for your quick reply. Indeed the sample output I showed was from a
> Nehalem machine. However, I saw the same problem on some wolfdale and
> harpertown systems too.
>
> So far, I have tried only a 0 second and a 6 second jobs. Will make some
> longer runs.
>
> Cheng
>
>
>
>
>
> On Mon, Jul 20, 2009 at 12:50 PM, Rick Kufrin < rk...@il... >
> wrote:
>
>
> Cheng,
>
> There are two issues that I can think of that you might be experiencing.
> The first is simply that the program may not run long enough to generate
> profiling samples. However, I am also wondering if (based on the processor
> brand string listed in the output) if you are working with an Intel Nehalem
> CPU. If so, there are known problems with the most recent release of
> PerfSuite on that processor. We have finished adding the support for that
> system, but have not yet placed a new software release on SourceForge or at
> NCSA. We hope to do so in the very near future, though.
>
> Sorry for the difficulties, please let me know if you also suspect this may
> be what is affecting you.
>
> Rick
>
> Cheng Liao wrote:
>
>
>
>
>
> I just installed perfsuite on some linux x86_64 machines (only one with
> papi and perfctr).
> However, even though 'make -s check' outputs looked fine, all the psprocess
> outputs would
> be 'empty' and show no profile timing info. Is this something that can be
> easily fixed?
> Thanks,
> Cheng
> PerfSuite Hardware Performance Summary Report
>
> Version : 1.0
> Created : Mon Jul 20 11:20:26 AM PDT 2009
> Generator : psprocess 0.3
> XML Source : ls.15381.em64td.xml
>
> Execution Information
>
> ============================================================================================
> Collector : libpshwpc
> Date : Sat Jul 18 17:31:54 2009
> Host : em64td
> User : cliao
> Command : ls
>
> Processor and System Information
>
> ============================================================================================
> Node CPUs : 8
> Vendor : Intel
> Family : Pentium Pro (P6)
> Brand : Intel(R) Xeon(R) CPU X5570 @ 2.93GHz
> CPU Revision : 4
> Clock (MHz) : 2933.589
> Memory (MB) : 24096.66
> Pagesize (KB) : 4
>
> Cache Information
>
> ============================================================================================
> Cache levels : 0
>
> Profile Information
>
> ============================================================================================
> Class : profil
> Version : 2.5
> Event : milliseconds
> Period : 10
> Samples : 0
> Domain : user
> Run Time : 0.00 (seconds)
> Min Self % : (all)
>
> Module Summary
>
> --------------------------------------------------------------------------------
> Samples Self % Total % Module
>
>
> File Summary
>
> --------------------------------------------------------------------------------
> Samples Self % Total % File
>
>
> Function Summary
>
> --------------------------------------------------------------------------------
> Samples Self % Total % Function
>
>
> Function:File:Line Summary
>
> --------------------------------------------------------------------------------
> Samples Self % Total % Function:File:Line
>
> ------------------------------------------------------------------------
>
>
> ------------------------------------------------------------------------------
> Enter the BlackBerry Developer Challenge This is your chance to win up to
> $100,000 in prizes! For a limited time, vendors submitting new applications
> to BlackBerry App World(TM) will have
> the opportunity to enter the BlackBerry Developer Challenge. See full prize
> details at: http://p.sf.net/sfu/Challenge
> ------------------------------------------------------------------------
>
> _______________________________________________
> PerfSuite-users mailing list
> Per...@li...
> https://lists.sourceforge.net/lists/listinfo/perfsuite-users
>
>
>
>
>
|
|
From: Rick K. <rk...@il...> - 2009-08-19 19:26:33
|
Cheng, These are some interesting observations and questions. I have a few comments, but first one question: which release of PerfSuite are you running? To find out, do "psinv -V", and it will appear at the top. The reason why I ask this is that the appearance of negative sample counts is a problem that I believe was addressed already, a fix checked in quite a while ago. If the release you are using is after that fix, then apparently we still have an issue (and I thank you for your report in advance). Second, I am glad you resolved (apparently) an incorrect offset, but I am curious to know how you verify that this provided a fix. It is not unusual to see question marks (PCs that cannot be mapped to source locations). Regarding the sampling bin size: that is a fixed size, and it is currently the size of an unsigned short, which should allow for up to 64K samples to map to any one bin. After that, overflow would occur and you would effectively lose samples and get an incorrect profile. The only solution at present, unfortunately and as you suggest, is to increase the sampling period to reduce overall sample counts. Rick Cheng Liao wrote: > Rick, > > The large ?? entry in my profile is caused by an incorrect offset for > libc.so.6 base > address. I need to manually set the libc offset to zero in the xml > file, then run psprocess, to > fix the problem. > > A new problem I have encountered is for large jobs, the bin for the > number of sampling points would overflow: > > <sample pc="59d770">27665</sample> > <sample pc="59d772">26940</sample> > <sample pc="59d776">27564</sample> > <sample pc="59d77a">-31502</sample> > <---------- negative > <sample pc="59d77e">11502</sample> > <sample pc="59d782">5323</sample> > <sample pc="59d78a">21065</sample> > <sample pc="59d78e">32422</sample> > <sample pc="59d790">27709</sample> > <sample pc="59d794">-30384</sample> > <------------- negative > <sample pc="59d798">24312</sample> > Does perfsuite support a large sampling point bin, or using a higher > profiling overflow threhold would be the only solution? > > Thanks, > Cheng > |
|
From: Cheng L. <cyl...@gm...> - 2009-08-19 20:43:52
|
Rick,
Here is my 'psinv -V' output for the particular machine on which I am seeing
the negative numbers of
sampling points.
[cliao@maui1 bin]$ psinv -V
--------------------------------------------
PerfSuite 0.6.2a6
psinv 0.6
University of Illinois/NCSA
Open Source License
http://perfsuite.ncsa.uiuc.edu/
http://perfsuite.sourceforge.net/
--------------------------------------------
I also have PerfSuite 0.6.2 installed on different machines, but haven't run
any large job on those machines yet.
To track down the incorrect libc offset on Linux, I simply pored throu the
xml file and ran addr2line manually
for the samples that were pertinent to the executable to see if I got a lot
of ??. No luck there, but I was able to
conclude the libc offset should be zero after I did a 'nm libc.so.6' and saw
pretty similiar addresses from the nm
output and the xml file.
The software company I am working for also has a large PC/Windows customer
base...
Thanks,
Cheng
On Wed, Aug 19, 2009 at 12:25 PM, Rick Kufrin <rk...@il...> wrote:
> Cheng,
>
> These are some interesting observations and questions. I have a few
> comments, but first one question: which release of PerfSuite are you
> running? To find out, do "psinv -V", and it will appear at the top. The
> reason why I ask this is that the appearance of negative sample counts is a
> problem that I believe was addressed already, a fix checked in quite a while
> ago. If the release you are using is after that fix, then apparently we
> still have an issue (and I thank you for your report in advance).
>
> Second, I am glad you resolved (apparently) an incorrect offset, but I am
> curious to know how you verify that this provided a fix. It is not unusual
> to see question marks (PCs that cannot be mapped to source locations).
>
> Regarding the sampling bin size: that is a fixed size, and it is currently
> the size of an unsigned short, which should allow for up to 64K samples to
> map to any one bin. After that, overflow would occur and you would
> effectively lose samples and get an incorrect profile. The only solution at
> present, unfortunately and as you suggest, is to increase the sampling
> period to reduce overall sample counts.
>
> Rick
>
>
> Cheng Liao wrote:
>
>> Rick,
>> The large ?? entry in my profile is caused by an incorrect offset for
>> libc.so.6 base
>> address. I need to manually set the libc offset to zero in the xml file,
>> then run psprocess, to
>> fix the problem.
>> A new problem I have encountered is for large jobs, the bin for the
>> number of sampling points would overflow:
>> <sample pc="59d770">27665</sample>
>> <sample pc="59d772">26940</sample>
>> <sample pc="59d776">27564</sample>
>> <sample pc="59d77a">-31502</sample>
>> <---------- negative
>> <sample pc="59d77e">11502</sample>
>> <sample pc="59d782">5323</sample>
>> <sample pc="59d78a">21065</sample>
>> <sample pc="59d78e">32422</sample>
>> <sample pc="59d790">27709</sample>
>> <sample pc="59d794">-30384</sample>
>> <------------- negative
>> <sample pc="59d798">24312</sample>
>> Does perfsuite support a large sampling point bin, or using a higher
>> profiling overflow threhold would be the only solution?
>> Thanks,
>> Cheng
>>
>>
>
|
|
From: Rick K. <rk...@il...> - 2009-08-19 21:11:45
|
Cheng, Well, I think the versioning number explains the negative sample count. Version 0.6.2a6 dates from 2006, and the fix I mentioned earlier was checked in in September 2007, so you would need a release later than that to pick up the fix (but the overflow issue would still be present). I am not sure what to make of the change of the offset to libc, the offset is computed from what is present in the load map for the process. I will think about it, and if I have any insight will follow up. Rick Cheng Liao wrote: > Rick, > > Here is my 'psinv -V' output for the particular machine on which I am > seeing the negative numbers of > sampling points. > > [cliao@maui1 bin]$ psinv -V > -------------------------------------------- > PerfSuite 0.6.2a6 > psinv 0.6 > University of Illinois/NCSA > Open Source License > http://perfsuite.ncsa.uiuc.edu/ > http://perfsuite.sourceforge.net/ > -------------------------------------------- > I also have PerfSuite 0.6.2 installed on different machines, but > haven't run any large job on those machines yet. > > To track down the incorrect libc offset on Linux, I simply pored throu > the xml file and ran addr2line manually > for the samples that were pertinent to the executable to see if I got > a lot of ??. No luck there, but I was able to > conclude the libc offset should be zero after I did a 'nm libc.so.6' > and saw pretty similiar addresses from the nm > output and the xml file. > > The software company I am working for also has a large PC/Windows > customer base... > > Thanks, > Cheng > > |
×
Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.
Thanks for helping keep SourceForge clean.
X