Date: Fri, 18 Jan 2013 10:39:25 -0500 From: Eitan Adler <lists@eitanadler.com> To: Fabian Keil <freebsd-listen@fabiankeil.de> Cc: questions@freebsd.org Subject: Re: absurd I/O perf with ZFS: hangs on zfs->cv) Message-ID: <CAF6rxgneT9=_%2BRAqNo9%2BUwSg2VCdn=wLHkpsHSdXEwYp=mPTww@mail.gmail.com> In-Reply-To: <20130118112010.43ad3edb@fabiankeil.de> References: <CAF6rxgkJr7xrmZkkp0eH_-WzuvdmCBDwPPOZFoqWUZ2jqXqneg@mail.gmail.com> <CAF6rxgkCWC_HopEWqhMKw245itPXKJbn4_DD66NKiD1di8ysWw@mail.gmail.com> <20130117135241.434f273a@fabiankeil.de> <CAF6rxgnx4CXgHBR9SEUebps7LjL%2Bk0UczbncNrSpoPpU7qCmAw@mail.gmail.com> <20130118112010.43ad3edb@fabiankeil.de>
next in thread | previous in thread | raw e-mail | index | archive | help
On 18 January 2013 05:20, Fabian Keil <freebsd-listen@fabiankeil.de> wrote: > A common recommendation is to disable atime for all datasets where > it isn't needed as it can cause lots of unnecessary write operations. Good call. I thought I had already disabled atime updating but it turns out that some datasets had this property still on. > With 8 GB of RAM ZFS enables prefetching by default and I assume > for the git use case it's not too useful and could hurt performance > by amplifying read operations. The "Data Prefetch Efficiency" is > shown by zfs-stats and if it doesn't look too impressive you might > want to disable prefetching to see if it helps. I will look at this number - thanks. > If your repository isn't fresh, you could also try "git gc". > My impression is that the automatic doesn't trigger frequently > enough for larger repositories like /usr/src. This isn't git specific. Doing anything remotely I/O related takes forever: - starting a new shell take over a minute - opening a file in vim which wasn't opened recently take a few minutes - etc. That said, I will be looking closely at the zfs-stats and see if I notice any patterns. -- Eitan Adler
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAF6rxgneT9=_%2BRAqNo9%2BUwSg2VCdn=wLHkpsHSdXEwYp=mPTww>