While I don't understand your 'each brick system also being a client' setup - you mean that each gluster brick is a native gluster client as well? And that is where much of your gluster access is coming from? That seems .. suboptimal if that's the setup. Is there a reason for that setup? We have a distributed-only glusterfs feeding a medium cluster over a similar same setup QDR IPoIB with 4 servers with 2 bricks each. On a fairly busy system (~80MB/s background), I can get about 100-300MB/s writes to the gluster fs on a large 1.7GB file. (With tiny writes, the perf decreases dramatically). Here is my config: (if anyone spies something that I should change to increase my perf, please feel free to point out my mistake) gluster: Volume Name: gl Type: Distribute Volume ID: 21f480f7-fc5a-4fd8-a084-3964634a9332 Status: Started Number of Bricks: 8 Transport-type: tcp,rdma Bricks: Brick1: bs2:/raid1 Brick2: bs2:/raid2 Brick3: bs3:/raid1 Brick4: bs3:/raid2 Brick5: bs4:/raid1 Brick6: bs4:/raid2 Brick7: bs1:/raid1 Brick8: bs1:/raid2 Options Reconfigured: performance.write-behind-window-size: 1024MB performance.flush-behind: on performance.cache-size: 268435456 nfs.disable: on performance.io-cache: on performance.quick-read: on performance.io-thread-count: 64 auth.allow: 10.2.*.*,10.1.*.* my RAID6s (via 3ware 9750s) are mounted with the following options /dev/sdc /raid1 xfs rw,noatime,sunit=512,swidth=8192,allocsize=32m 0 0 /dev/sdd /raid2 xfs rw,noatime,sunit=512,swidth=7680,allocsize=32m 0 0 (and should probably be using 'nobarrier,inode64' as well. - testing this now) There are some good refs on prepping XFS fs for max perf here: <http://www.mythtv.org/wiki/Optimizing_Performance#XFS-Specific_Tips> The script at: <http://www.mythtv.org/wiki/Optimizing_Performance#Further_Information> can help to setup the sunit/swidth options. <http://www.mysqlperformanceblog.com/2011/12/16/setting-up-xfs-the-simple- edition/> Your ib interfaces should be using large mtus (65536) hjm On Wednesday, February 13, 2013 10:35:12 PM Michael Colonno wrote: > More data: I got the Infiniband network (QDR) working well and > switched my gluster volume to the Infiniband fabric (IPoIB, not RDMA since > it doesn't seem to be supported yet for 3.x). The filesystem was slightly > faster but still well short of what I would expect by a wide margin. Via an > informal test (timing the movement of a large file) I'm getting several MB/s > - well short of even a standard Gb network copy. With the faster network > the CPU load on the brick systems increased dramatically: now I'm seeing > 200%-250% usage by glusterfsd and glusterfs. > > This leads me to believe that gluster is really not enjoying my > eight-brick, 2x replication volume with each brick system also being a > client. I tried a rebalance but no measurable effect. Any suggestions for > improving the performance? Having each brick be a client of itself seemed > the most logical choice to remove interdependencies but now I'm doubting the > setup. > > > > Thanks, > > ~Mike C. > > > > From: gluster-users-bounces at gluster.org > [mailto:gluster-users-bounces at gluster.org] On Behalf Of Joe Julian > Sent: Sunday, February 03, 2013 11:47 AM > To: gluster-users at gluster.org > Subject: Re: high CPU load on all bricks > > > > On 02/03/2013 11:22 AM, Michael Colonno wrote: > > > > Having taken a lot more data it does seem the glusterfsd and > glusterd processes (along with several ksoftirqd) spike up to near 100% on > both client and brick servers during any file transport across the mount. > Thankfully this is short-lived for the most part but I'm wondering if this > is expected behavior or what others have experienced(?) I'm a little > surprised such a large CPU load would be required to move small files and / > or use an application within a Gluster mount point. > > > If you're getting ksoftirqd spikes, that sounds like a hardware issue to me. > I never see huge spikes like that on my servers nor clients. > > > > > > > I wanted to test this against an NFS mount of the same Gluster > volume. I managed to get rstatd installed and running but my attempts to > mount the volume via NFS are met with: > > > > mount.nfs: requested NFS version or transport protocol is not > supported > > > > Relevant line in /etc/fstab: > > > > node1:/volume /volume nfs > defaults,_netdev,vers=3,mountproto=tcp 0 0 > > > > It looks like CentOS 6.x has NFS version 4 built into everything. So a few > questions: > > > > - Has anyone else noted significant performance differences between a > glusterfs mount and NFS mount for volumes of 8+ bricks? > > - Is there a straightforward way to make the newer versions of CentOS > play nice with NFS version 3 + Gluster? > > - Are there any general performance tuning guidelines I can follow to > improve CPU performance? I found a few references to the cache settings but > nothing solid. > > > > If the consensus is that NFS will not gain anything then I won't waste the > time setting it all up. > > > NFS gains you the use of FSCache to cache directories and file stats making > directory listings faster, but it adds overhead decreasing the overall > throughput (from all the reports I've seen). > > I would suspect that you have the kernel nfs server running on your servers. > Make sure it's disabled. > > > > > > > Thanks, > > ~Mike C. > > > > > > From: gluster-users-bounces at gluster.org > [mailto:gluster-users-bounces at gluster.org] On Behalf Of Michael Colonno > Sent: Friday, February 01, 2013 4:46 PM > To: gluster-users at gluster.org > Subject: Re: high CPU load on all bricks > > > > Update: after a few hours the CPU usage seems to have dropped > down to a small value. I did not change anything with respect to the > configuration or unmount / stop anything as I wanted to see if this would > persist for a long period of time. Both the client and the self-mounted > bricks are now showing CPU < 1% (as reported by top). Prior to the larger > CPU loads I installed a bunch of software into the volume (~ 5 GB total). Is > this kind a transient behavior - by which I mean larger CPU loads after a > lot of filesystem activity in short time - typical? This is not a problem > in my deployment; I just want to know what to expect in the future and to > complete this thread for future users. If this is expected behavior we can > wrap up this thread. If not then I'll do more digging into the logs on the > client and brick sides. > > > > Thanks, > > ~Mike C. > > > > From: Joe Julian [mailto:joe at julianfamily.org] > Sent: Friday, February 01, 2013 2:08 PM > To: Michael Colonno; gluster-users at gluster.org > Subject: Re: high CPU load on all bricks > > > > Check the client log(s). > > Michael Colonno <mcolonno at stanford.edu> wrote: > > Forgot to mention: on a client system (not a brick) the > glusterfs process is consuming ~ 68% CPU continuously. This is a much less > powerful desktop system so the CPU load can't be compared 1:1 with the > systems comprising the bricks but still very high. So the issue seems to > exist with both glusterfsd and glusterfs processes. > > > > Thanks, > > ~Mike C. > > > > From: gluster-users-bounces at gluster.org > [mailto:gluster-users-bounces at gluster.org] On Behalf Of Michael Colonno > Sent: Friday, February 01, 2013 12:46 PM > To: gluster-users at gluster.org > Subject: high CPU load on all bricks > > > > Gluster gurus ~ > > > > I've deployed and 8-brick (2x replicate) Gluster 3.3.1 volume on > CentOS 6.3 with tcp transport. I was able to build, start, mount, and use > the volume. On each system contributing a brick, however, my CPU usage > (glusterfsd) is hovering around 20% (virtually zero memory usage > thankfully). These are brand new, fairly beefy servers so 20% CPU load is > quite a bit. The deployment is pretty plain with each brick mounting the > volume to itself via a glusterfs mount. I assume this type of CPU usage is > atypically high; is there anything I can do to investigate what's soaking up > CPU and minimize it? Total usable volume size is only about 22 TB (about 45 > TB total with 2x replicate). > > > > Thanks, > > ~Mike C. > > > > > _____ > > > Gluster-users mailing list > Gluster-users at gluster.org > http://supercolony.gluster.org/mailman/listinfo/gluster-users > > > > > > > _______________________________________________ > Gluster-users mailing list > Gluster-users at gluster.org > http://supercolony.gluster.org/mailman/listinfo/gluster-users --- Harry Mangalam - Research Computing, OIT, Rm 225 MSTB, UC Irvine [m/c 2225] / 92697 Google Voice Multiplexer: (949) 478-4487 415 South Circle View Dr, Irvine, CA, 92697 [shipping] MSTB Lat/Long: (33.642025,-117.844414) (paste into Google Maps) --- "Something must be done. [X] is something. Therefore, we must do it." Bruce Schneier, on American response to just about anything.
