Re: [PerfSuite-users] Periodic measurement

SourceForge Headquarters 225 Broadway Suite 1600 San Diego, CA 92101 +1 (858) 422-6466

Hello Sir
 
 Progress about MPI : I am able to get output with MPI application.
 
 But one of my MPI application is using "charm++", On that if I insert the routines
 
 
 The application runs for given sampling seconds, then inside the
 sampler subroutine, after calling ps_hwpc_read(), the program is exiting. (ret value is not equal to success)
 
 The return value from ps_hwpc_read() is 1012.
 
 The error it is giving while exiting is: "calling sequence not allowed"
 
 Any suggestion on this. Is this because of charm++?
 
 Thanx
 Regards
 
 sanjay
 <blockquote style='border-top: 1px solid rgb(204, 204, 204); margin: 0.8ex 0pt 0pt 0pt; padding-bottom: 1ex;'>
 ---------[ Received Mail Content ]---------- 
 Subject : Re: [PerfSuite-users] Periodic measurement 
 Date : Wed, 30 Jan 2008 11:10:04 -0600 
 From : Rick Kufrin &lt;rk...@nc...&gt; 
 To : Sanjay H A &lt;san...@ly...&gt; 
 Cc : per...@li... 
 
 Sanjay,
 
 The example programs like sampler that are distributed with PerfSuite 
 are not meant to be production-quality tools, but small cases that might 
 be useful for people to see how things work and demonstrate some basic 
 capabilities. However, you might want to think about what's happening 
 underneath if you base something on an example like sampler. It might 
 be helpful for you to take a look at the manual pages for setitimer() 
 and sigaction(), which are used in sampler as building blocks.
 
 Here are a few points that may be relevant:
 
 1. For performance data to be printed within sampler, the signal handler 
 that is installed must be triggered by delivery of the signal. The 
 handler itself is the function "sampler". As an aside, this is not 
 typically a good idea to be doing output from within a signal handler 
 but again this is just a toy example program.
 
 2. The sampler program sets interval timers using the ITIMER_VIRTUAL 
 time domain, which is decremented only when the process is executing. 
 If a process is sleeping, it's not accumulating any execution time, and 
 so this timer would likely be decrementing at a much slower rate.
 
 3. The sampler program, in its use of sigaction, does not request that 
 system calls are restarted across signals (see the manual page for 
 sigaction()).
 
 I ran some tests and found that even with calls to printf() or sleep() 
 within the busyloop function, output from the signal handler will occur, 
 but not necessarily at the same rate one would expect when running the 
 example unchanged. You might want to play around with it some more to 
 see if you can produce similar results.
 
 One other point to raise in case you make some changes is that PAPI, 
 when multiplexing, also uses signals to trigger the timeslicing of 
 counters (in most cases, I believe it uses SIGPROF and ITIMER_PROF). 
 That means some care has to be taken if changing signals and interval 
 timers within sampler.
 
 Not trying to be evasive here, but rather, trying to point you in a 
 different direction in terms of how you are thinking things are working 
 with the sampler example.
 
 Rick
 
 Sanjay H A wrote:
 > Hello Sir
 > I got the point. Then I will insert persuite routines in my application to measure performance periodically.
 > But I am afraid about the test I did with perfsuite sampler program.The thing I noticed is, inside busyloop(), if I insert printf statement or some sleep statement , the performance data will not 
 > be printed on standard output periodically. Any solutions for that?
 >
 > Thanx
 > Regards
 > Sanjay
 > 
 >
 >
 > 
 </BLOCKQUOTE>