On Mon, Jun 22, 2020 at 05:26:12PM -0400, J. Bruce Fields wrote: > On Sun, Jun 21, 2020 at 09:54:08AM +1000, Dave Chinner wrote: > > On Fri, Jun 19, 2020 at 09:56:33PM -0400, J. Bruce Fields wrote: > > > On Sat, Jun 20, 2020 at 11:49:57AM +1000, Dave Chinner wrote: > > > > However, other people have different opinions on this matter (and we > > > > know that from the people who considered XFS v4 -> v5 going slower > > > > because iversion a major regression), and so we must acknowledge > > > > those opinions even if we don't agree with them. > > > > > > Do you have any of those reports handy? Were there numbers? > > > > e.g. RH BZ #1355813 when v5 format was enabled by default in RHEL7. > > Numbers were 40-47% performance degradation for in-cache writes > > caused by the original IVERSION implementation using iozone. There > > were others I recall, all realted to similar high-IOP small random > > writes workloads typical of databases.... > > Thanks, that's an interesting bug! Though a bit tangled. This is where > you identified the change attribute as the main culprit: > > https://bugzilla.redhat.com/show_bug.cgi?id=1355813#c42 > > The test was running at 70,000 writes/s (2.2GB/s), so it was one > transaction per write() syscall: timestamp updates. On CRC > enabled filesystems, we have a change counter for NFSv4 - it > gets incremented on every write() syscall, even when the > timestamp doesn't change. That's the difference in behaviour and > hence performance in this test. > > In RHEL8, or anything post-v4.16, the frequency of change attribute > updates should be back down to that of timestamp updates on this > workload. So it'd be interesting to repeat that experiment now. Yup, which in itself has been a problem for similar workloads. There's a reason we now recommend the use of lazytime for high performance database workloads that can do hundreds of thousands of small write IOs a second... > The bug was reporting in-house testing, and doesn't show any evidence > that particular regression was encountered by users; Eric said: > > https://bugzilla.redhat.com/show_bug.cgi?id=1355813#c52 > > Root cause of this minor in-memory regression was inode > versioning behavior; as it's unlikely to have real-world effects > (and has been open for years with no customer complaints) I'm > closing this WONTFIX to get it off the radar. It's just the first I found because bugzilla has a slow, less than useful search engine. We know that real applications have hit this, and we know even the overhead of timestamp updates on writes is way too high for them. > The typical user may just skip an upgrade or otherwise work around the > problem rather than root-causing it like this, so absence of reports > isn't conclusive. I understand wanting to err on the side of caution. Yup, it's a generic problem - just because we've worked around or mitigated the most common situations it impacts performance, that doesn't mean they work for everyone.... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
