[Date Prev][Date Next] [Thread Prev][Thread Next]
[Thread Index]
[Date Index]
[Author Index]
[Linux-cachefs] Re: Having trouble with the latest cachefiles
- From: "John Groves" <John groves net>
- To: "Linux filesystem caching discussion list" <linux-cachefs redhat com>
- Subject: [Linux-cachefs] Re: Having trouble with the latest cachefiles
- Date: Sat, 3 Jan 2009 09:56:42 -0600
Um, I looked again and discovered cachefilesd-0.9-1. Installing that solved
the problem related to the dump, so never mind that. Sorry I missed that
the first time.
I'm still interested in insight as to the performance overhead.
Thanks,
John
On Fri, Jan 2, 2009 at 10:44 AM, John Groves <John groves net> wrote:
> David,
>
> This is related to the lustre integration that I've done. It is necessary
> to summarize results as expediently as possible, which we hope justify
> continuation of this work.
>
> To that end, although I had a lot of stability problems with the very early
> fscache code in EL5 kernels, I was able to come up with a testable
> configuration that demonstrates that reading from fscache is faster in
> appropriate circumstances. However, I was using a cache disk subsystem that
> can be read by (iozone / ext3) at 1GB/s, but the (iozone / lustre / fscache
> /ext3 / fast disk) read performance was just under half that. In the old
> fscache code, (iozone / Lustre / fscache / ext3 / fast disk) performance =
> 0.45 (iozone / ext3 / fast disk).
>
> Bringing Lustre up to modern kernels is a separate piece of work, and it's
> just not ready yet. For analysis purposes, we decided to test the (iozone /
> nfs / fscache /ext3 / fast disk) performance on the very latest fscache code
> - if the latest code incurs less performance degradation, that would support
> the assertion that when Lustre is moved to late-model kernels (and therefore
> late model fscache code) we'll get a [hopefully large] performance gain.
>
> First, do you have any performance data that might shed light on this? I
> would be a bit surprised if you've tested with disk that's this fast...
>
> Secondly, I've built the GIT kernel with your nfs+fscache-42 patches
> applied (all went in cleanly). I'm running this kernel on an otherwise
> clean Fedora 10 system. fscache and cachefiles modules load cleanly, but I
> can't get cachefilesd to run. If I "yum install cachefilesd", I get version
> 0.7-5, and it complains that /proc/fs/cachefiles is missing. If I build
> cachefilesd-0.9, I a kernel BUG. The relevant lines from /var/log/messages
> are attached.
>
> Am I using the wrong version of cachefilesd? Is it clear what I'm doing
> wrong?
>
> Thanks,
> John Groves
>
>
>
[Date Prev][Date Next] [Thread Prev][Thread Next]
[Thread Index]
[Date Index]
[Author Index]