With glibc 2.26 there's a new per-thread cache in malloc, which improves malloc performance in general, and hopefully, specifically for the apps which each of you are most concerned about. In the event that you feel it doesn't help, or if you have other concerns about malloc performance, there's another part of my malloc performance work that's not in glibc 2.26 - the ability to capture and play back malloc activity. I created these tools to help with the malloc performance testing, since reliably recreating situations I wanted to test was difficult - especially with apps that interacted with the user, like LibreOffice. It's also handy for capturing data from apps that are difficult to configure (think, "requires other hosts to provide services" or "only happens with my /etc/passwd") and the occasional (gasp!) proprietary app that cannot be included in a bug report. Anyway, I've put the patch (suitable for rpmbuild) and tools tarball (they can run on a separate machine), along with a pointer to a COPR repo with F26 and rawhide builds of glibc-2.26.90-5 (x86 32/64 only for F26, ppc64le included for rawhide), here: http://people.redhat.com/dj/glibc/ Instructions for using the tools are also there. The tarball also includes a sample trace (of /bin/ls ;) to play with. So, if you think you have malloc-related performance issues, or want to benchmark a difficult application on different malloc implementations (the simulator is a regular app, you can LD_PRELOAD an interposed malloc), please give this a try. If you think you have an unusual app and want me to consider your malloc performance in future work, please feel free to send me (privately, they're huge) a workload to include in my collection :-) Thanks! DJ _______________________________________________ devel mailing list -- devel@xxxxxxxxxxxxxxxxxxxxxxx To unsubscribe send an email to devel-leave@xxxxxxxxxxxxxxxxxxxxxxx
