Thanks, I will definitely try disabling the allocation plugin. -Ash
-----Original Message----- From: Rohan Garg [mailto:rohg...@ccs.neu.edu] Sent: Saturday, September 24, 2016 1:23 AM To: Ashutosh Varma <ashutosh.va...@synopsys.com> Cc: dmtcp-forum@lists.sourceforge.net Subject: Re: [Dmtcp-forum] A question on performance Adding a few more thoughts/comments: > In our case, we do see some noticeable impact, and need to understand why. > After restore of checkpoint, almost normal runtime performance is seen. It's possible that the application is doing many I/O operations/system calls during its initialization. This could be the reason why the application is slowed down when running under DMTCP. And so, if you were to checkpoint-restart after the initialization phase, you would see better performance when the process becomes compute bound. Another comment I have related to initialization is related to memory allocation. If the application does many malloc's/free's during init, and if you are using DMTCP's alloc plugin (which is turned on by default), you could see a performance degradation. The wrapper functions for the alloc/free functions in the alloc plugin don't do much but do acquire and release a global rw lock to prevent certain race conditions. You could try turning it off as a quick test. $ dmtcp_launch --disable-alloc-plugin a.out Finally, I would suggest doing an strace and/or running under perf tool to figure out the bottleneck(s). It wouldn't be hard to optimize once we know the root cause. On Fri, Sep 23, 2016 at 05:06:17PM -0400, gene wrote: > Hi Ashutosh, > In most cases, there is no measurable impact or overhead when > running under dmtcp_launch. We do put wrapper functions around > certain system calls, but these are very lightweight. > If you're seeing a noticeable impact, which then goes away > after restart, then that's particularly intriguing. > > Could you give us some more details about what application you > are running? If there is a version of this that you can send us, > we'd like to try it locally. This forum is public, but you're also > welcome to write directly to us, and we could analyze this offline. > > Best, > - Gene > > On Fri, Sep 23, 2016 at 03:49:15PM +0000, Ashutosh Varma wrote: > > Hi, > > > > Has there been any data collected by users on if running an environment > > under dmtcp_launch/dmtcp_coordinator affects runtime performance or does > > not? > > In our case, we do see some noticeable impact, and need to understand why. > > After restore of checkpoint, almost normal runtime performance is seen. > > > > Thanks. > > > > -Ash > > > > > ------------------------------------------------------------------------------ > > > _______________________________________________ > > Dmtcp-forum mailing list > > Dmtcp-forum@lists.sourceforge.net > > https://lists.sourceforge.net/lists/listinfo/dmtcp-forum > > > ------------------------------------------------------------------------------ > _______________________________________________ > Dmtcp-forum mailing list > Dmtcp-forum@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/dmtcp-forum ------------------------------------------------------------------------------ _______________________________________________ Dmtcp-forum mailing list Dmtcp-forum@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/dmtcp-forum
