Gary, On Thu, Oct 22, 2009 at 6:47 PM, <gary.m...@bull.com> wrote: > > Hi Stephane > > We are preparing to move to the Fedora 12 distribution which comes with the > 2.6.31 kernel. > > This kernel already contains the perf_events code from Ingo plus you have > built libpfm4 to work with it and papi-3.7.0 has early support for this > subsystem. >
True. > But I have not seen any mention of a pfmon that will be able to use it and > perf_events does not yet support uncore events on Nehalem (events that we > think provide important information). Although pfmon is an important tool > for us, our primary use with be with hpctoolkit through papi. There is the perf tool which is part of the kernel source tree. But it does not have all the symbolic event names, just generic events. It can be connected to libpfm4 (I have toyed with that myself). But given that they chose to have the tool in the kernel, it makes it hard to link against a library that is not. I don't think libpfm4 belongs in the kernel. I think there is enough in libpfm4 to have: hpctoolkit -> papi -> libpfm4 -> perf_events. But I don't think papi has libpfm4 support just yet. > > We are now in a position where we need to decide if it is time to move to > perf_events (and hope things improve fairly rapidly) or stay on perfmon (a > much more comfortable idea) for this next release. > I will not be producing a 2.6.31 perfmon2 patch. I don't see the point anymore. Perf_events was chosen to be the interface, it is time to live with it and its current shortcomings, especially in terms of PMU support. > >From your point of view: > > Do you feel that perf_events is ready to be used by customers using x86_64 > Nehalem systems ?? > It all depends on what you are after. For most PMU events, you will be okay. But there is list of about 10 events which may return bogus values because they may be programmed in the wrong counter. You may have seen that I posted a patch to fix that. It is in 2.6.32. Note that you may be able to backport the pair of patches I posted to fix this. > Do you plan to provide a set of perfmon2 kernel patches for the 2.6.31 > kernel ?? > No. > Do you plan to produce a pfmon which will work with perf_events ?? > I think it would be useful to have more than one tool. I know many people are familiar with the pfmon cmdline. But it will take some major rewriting and thus quite some time. The good thing is that I think the code will be much simpler and a lot of the cmdline options will likely remain. > And of course that ugly thing called schedules always gets in the way so if > you have plans to do these, could you give a rough idea where it sits on > your priority list (or even better about how long before it may be > available) that would be useful. > There will not be a 2.6.31 patch for perfmon2. I will deliver the 2.6.30 patch once: - I figure out why PMU interrupt are crashing the machine in i386 - when I get my Nehalem system back for testing I am hopeful next week, but I realize I have said that several times already! ------------------------------------------------------------------------------ Come build with us! The BlackBerry(R) Developer Conference in SF, CA is the only developer event you need to attend this year. Jumpstart your developing skills, take BlackBerry mobile applications to market and stay ahead of the curve. Join us from November 9 - 12, 2009. Register now! http://p.sf.net/sfu/devconference _______________________________________________ perfmon2-devel mailing list perfmon2-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
