[Gluster-users] Gluster / KVM Filesystem Benchmarks

[Joe Topjian]

Hello,

These benchmarks were not testing the Gluster / KVM native integration
patch. The benchmarks were testing performance of qcow2 files running on
Gluster volumes.

I just wanted to clear that up.

Thanks,
Joe


On Sun, Sep 9, 2012 at 4:32 AM, Andrei Mikhailovsky <andrei at arhont.com>wrote:

> Guys, do you know if this patch supports glusterfs over rdma?
>
> I am running glusterfs over infiniband and kvm performance is so nasty. As
> an example, I get around 600-800 mb/s for read/write on the glusterfs
> partition mounted on kvm server. However, vms stored on this partition can
> only read/write around 40-50mb/s. I would love to try this patch if there
> is rdma support.
>
> Thanks
>
>
>
>
> ------------------------------
> *From: *"Joe Topjian" <joe at topjian.net>
> *To: *gluster-users at gluster.org
> *Sent: *Wednesday, 5 September, 2012 6:47:09 AM
> *Subject: *[Gluster-users] Gluster / KVM Filesystem Benchmarks
>
>
> Hello,
>
> I did a few filesystem benchmarks with Gluster (3.3) and KVM using iozone
> and have compiled a spreadsheet with the results:
>
> https://docs.google.com/open?id=0B6GzZufzohwFZmozTFRSSHk5T0E
>
> Just a heads up: It is an Excel spreadsheet.
>
> All of the details that were used to generate the results are described in
> the spreadsheet. Of most interest would be the second tab titled "Gluster".
> The results that do not have "vm" in the description were iozone procs
> running directly on a mounted replicated Gluster volume (2 bricks). The
> "vm" results are iozone procs running in KVM virtual machines stored in
> qcow2 files.
>
> The first tab, General, is just some simple non-Gluster benchmarks that I
> ran for comparison.
>
> The third tab, Gluster old, was me doing iozone benchmarks on files with
> sizes ranging from 8mb to 512mb. I noticed that there was very little
> difference in the results so I decided to work with only 128mb and 256mb
> sized files.
>
> If you do not have access to Excel or something compatible, you can still
> view most of the information in the Google Doc. Here is a jpeg image of the
> main graph that was generated:
>
> https://docs.google.com/open?id=0B6GzZufzohwFWGtFS3I5UEllTkU
>
> Questions I have:
>
> * The "optimized settings" that I used were pulled from a Gluster
> Performance Tuning presentation. It doesn't look like the settings did very
> much in terms of optimization. Can someone comment on these settings? Are
> there better settings to use?
>
> * I'm a bit confused at how the KVM / qcow2 reads are much higher than the
> reads directly on the Gluster volume. Any idea why that is?
>
> * I ran all tests with the cache-io translator on and off. Like the
> "optimized settings", it wasn't of much use. Did I use this incorrectly?
>
> * The reason I did all tests with 128mb and 256mb sized files was to
> highlight the very bizarre trait where certain increments (16, 64, 256)
> gave very poor results while increments such as 8, 32, and 128 had good
> results. Any idea why that is?
>
> * Can anyone comment on if these results are of any use? Or are the stats
> I collect and the way I collected them incorrect?
>
> Please let me know if anyone has any questions or needs anything clarified.
>
> Thanks,
> Joe
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://supercolony.gluster.org/pipermail/gluster-users/attachments/20120909/92445a5a/attachment.html>