On Sat, 2021-09-04 at 15:18 +0200, Daniel Danzberger wrote: > > > > > > > +CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS=y > > > > > > CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS=y might not be a performance > > > win, > > > especially with only 1 GiB of RAM. Should be measured, otherwise > > > MADVISE should be selected instead. > > Same here, it was selected in the 5.4 kernel config and I have only > > run > > tested with this option so far. > > Let me do some benchmarking with MADVISE before we continue here > > ... > Looks like none of the processes running even gets a huge page: > -- > root@OpenWrt:~# grep AnonHugePages /proc/meminfo > AnonHugePages: 0 kB > root@OpenWrt:~# > -- > Even my test tool that just allocs one 8MB block doesn't get one. > Looks like CONFIG_TRANSPARENT_HUGEPAGE isn't working on this > platform. After further testing, huge pages only get used when memory is allocated >= 2MB with mmap() on arm64. Unlike on x86 where processes get a huge page when using malloc() or mmap() >= 2MB. So I think we can safely keep CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS=y like in 5.4
-- Regards Daniel Danzberger embeDD GmbH, Alter Postplatz 2, CH-6370 Stans _______________________________________________ openwrt-devel mailing list openwrt-devel@lists.openwrt.org https://lists.openwrt.org/mailman/listinfo/openwrt-devel
