On 09/03/14 21:18, Steven Hartland wrote: > > ----- Original Message ----- From: "Andriy Gapon" <a...@freebsd.org> > > >> on 03/09/2014 23:22 Nikolai Lifanov said the following: >>> On 09/03/14 15:22, John Baldwin wrote: >>>> On Wednesday, September 03, 2014 11:05:04 AM Nikolai Lifanov wrote: >>>>> On 09/03/14 04:09, Steven Hartland wrote: >>>>>> I'm looking to MFC this change so wanted to check if >>>>>> anyone had an final feedback / objections? >>>>>> >>>>>> I know we currently have Alan's feedback on changing >>>>>> the #ifdef __i386__ to #ifndef UMA_MD_SMALL_ALLOC >>>>>> which sounds sensible but waiting Peter to comment on. >>>>>> >>>>>> Regards >>>>>> Steve >>>>> >>>>> I have no technical input, but this change improves ARC usefulness for >>>>> me quite a bit. I would like to see the improvement in 10-STABLE. >>>> >>>> Can you verify that the current 10-STABLE (as of today) with all the >>>> various pagedaemon fixes still has ARC issues for your workload? >>>> >>> >>> It doesn't have any issues, but I noticed the improvement on CURRENT. I >>> observed that just after this change, my package builder is much more >>> likely to retain MFU and not evict useful things from there (the port >>> tree) after large builds. >>> However, I run a lot more 10.0-RELEASE than CURRENT and I would like to >>> see this improvement release-bound. >>> >>> I would be happy to test this on 10-STABLE if you think that this is >>> relevant. >> >> >> As noted before, unfortunately, this commit (plus its fixups) contains >> at least >> two related but distinct changes. So, to separate the wheat from the >> chaff, >> could you please try to comment out the following block in >> sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c, function >> arc_reclaim_needed: >> >> if (kmem_free_count() < zfs_arc_free_target) { >> DTRACE_PROBE2(arc__reclaim_freetarget, uint64_t, >> kmem_free_count(), uint64_t, zfs_arc_free_target); >> return (1); >> } >> >> Alternatively, I think that the same effect can be achieved by setting >> sysctl >> vfs.zfs.arc_free_target to the same value as vm.stats.vm.v_free_min. > > Thats correct that would achieve the same thing. > >> It's interesting to me whether you would still see the better >> performance or if >> that improvement would be undone. > > Indeed that would be interesting, but we might find that its quite > memory size > dependent given the scaling so confirming HW details would be nice too. > > I'd also be interested to know who wins the free race between the VM and > ARC > when using that value. > > For those following this thread but not the review, I've added some > additional > information there which you might be interested in: > https://reviews.freebsd.org/D702 > > Regards > Steve
Just an update: I'm in the middle of testing this. I have to finish a large bulk build to observe the behavior one way or another. I have 32G of physical memory and 2x16G dedicated swap SSDs (L2ARC wasn't very useful, but I should probably retest this) on this machine. My ARC is usually at 14G with ~5G of MFU full of things I benefit from keeping there (port trees, base jails). Builds themselves happen in tmpfs and I usually have around 1.5G - 4G "Free" memory (unless building something like pypy). - Nikolai Lifanov _______________________________________________ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"