From: George M. <ge...@ma...> - 2009-06-15 14:31:53
|
Dear all, sorry if my question is simple but I didn't find any info about it. I have a fortran code, I have profiled it either with psrun either by PSF_ commands. My question is if I can profile two (or more) sections of my code and take as output two values about Mflops for example I would like to see Mflops per function and not total Mflops for all the program. I tried with PSF commands to start profiling, write the file and start again profiling but there was no second file. Thanks a lot, Best regards, George Markomanolis |
From: Rick K. <rk...@il...> - 2009-06-15 21:36:40
Attachments:
select.f
|
George, No need to be sorry, your question is reasonable. Unfortunately, although I understand conceptually what you are trying to accomplish, that is: 1. Bracket multiple portions of your application and collect counter data independently for each 2. Output separate XML documents for each bracketed portion 3. Use psprocess to generate derived metrics for each XML document ... this mode of operation is not well-supported in PerfSuite. One could arrange to mimic this behavior, but it would be a lot of effort and code changes within the application to manage the data. The simplest way of achieving this with the PS API is to do multiple runs of your application, each run collecting the performance data from a different region of the application, rather than attempting to collect data for multiple regions simultaneously. It is still possible to get the information you are looking for through the PerfSuite API, and I am attaching a small example program written today to show the basic idea. The example has very few comments, but hopefully it is not too difficult to understand. Note that XML is not used at all in this way of measurement. Regarding your second question today (about profiling), it is currently not possible through PerfSuite to profile based on more than one hardware event. I should emphasize this is a limitation of PerfSuite, and not the hardware capabilities or limitations of PAPI. Rick George Markomanolis wrote: > Dear all, > > sorry if my question is simple but I didn't find any info about it. I > have a fortran code, I have profiled it either with psrun either by PSF_ > commands. My question is if I can profile two (or more) sections of my > code and take as output two values about Mflops for example I would like > to see Mflops per function and not total Mflops for all the program. I > tried with PSF commands to start profiling, write the file and start > again profiling but there was no second file. > > Thanks a lot, > Best regards, > George Markomanolis > > ------------------------------------------------------------------------------ > Crystal Reports - New Free Runtime and 30 Day Trial > Check out the new simplified licensing option that enables unlimited > royalty-free distribution of the report engine for externally facing > server and web deployment. > http://p.sf.net/sfu/businessobjects > _______________________________________________ > PerfSuite-users mailing list > Per...@li... > https://lists.sourceforge.net/lists/listinfo/perfsuite-users > > |
From: George M. <ge...@ma...> - 2009-06-15 22:19:55
|
Dear Rick, Thanks for the answer. At first I should say that I don't want to run the program many times in order to profile all the code as I believe it's not so right specially if I use a lot of processors and I have to run it again and again. Although your code (I am not expert in fortran) seems very nice and it does whatever I want. I don't care about XML file as I worry only for the result's accuracy about Mflops. Thanks a lot for the tip as I can understand it reads the hardware counters so maybe it's ok. Thanks a lot, Best regards, George Markomanolis On Mon, Jun 15, 2009 at 11:36 PM, Rick Kufrin <rk...@il...> wrote: > George, > > No need to be sorry, your question is reasonable. Unfortunately, although > I understand conceptually what you are trying to accomplish, that is: > > 1. Bracket multiple portions of your application and collect counter data > independently for each > 2. Output separate XML documents for each bracketed portion > 3. Use psprocess to generate derived metrics for each XML document > > ... this mode of operation is not well-supported in PerfSuite. One could > arrange to mimic this behavior, but it would be a lot of effort and code > changes within the application to manage the data. > > The simplest way of achieving this with the PS API is to do multiple runs > of your application, each run collecting the performance data from a > different region of the application, rather than attempting to collect data > for multiple regions simultaneously. > > It is still possible to get the information you are looking for through the > PerfSuite API, and I am attaching a small example program written today to > show the basic idea. The example has very few comments, but hopefully it is > not too difficult to understand. Note that XML is not used at all in this > way of measurement. > > Regarding your second question today (about profiling), it is currently not > possible through PerfSuite to profile based on more than one hardware event. > I should emphasize this is a limitation of PerfSuite, and not the hardware > capabilities or limitations of PAPI. > > Rick > > George Markomanolis wrote: > >> Dear all, >> >> sorry if my question is simple but I didn't find any info about it. I have >> a fortran code, I have profiled it either with psrun either by PSF_ >> commands. My question is if I can profile two (or more) sections of my code >> and take as output two values about Mflops for example I would like to see >> Mflops per function and not total Mflops for all the program. I tried with >> PSF commands to start profiling, write the file and start again profiling >> but there was no second file. >> >> Thanks a lot, >> Best regards, >> George Markomanolis >> >> >> ------------------------------------------------------------------------------ >> Crystal Reports - New Free Runtime and 30 Day Trial >> Check out the new simplified licensing option that enables unlimited >> royalty-free distribution of the report engine for externally facing >> server and web deployment. >> http://p.sf.net/sfu/businessobjects >> _______________________________________________ >> PerfSuite-users mailing list >> Per...@li... >> https://lists.sourceforge.net/lists/listinfo/perfsuite-users >> >> >> > > |