Hi, On Tue, 2013-09-24 at 08:57 -0400, Thom Gardner wrote: > OK, I don't know much about increasing NFS performance, but I do have > some things for you to consider that may actually help anyway: > > In general we (the cluster support group at RedHat) have started > recommending that you just not even use GFS or GFS2 for use as an > exported FS from a cluster. The reason being that if you follow > all the rules, these filesystems don't usually buy you anything. > While a GFS or GFS2 filesystem is exported, you shouldn't have any > other type of access to it, and then only from one node at a time. Well thats not entirely true - you should get a faster failover time doing things this way. > We recommend that you don't even mount on more than one node at > a time. Thats news to me... there is no problem at all with mounting a GFS(2) on multiple nodes in this kind of configuration, and if there is no i/o from the second node, it will make no difference to overall performance either. > You certainly should not NFS export it from more than one > node at a time. You should not even export it as both NFS and CIFS > at the same time from the same node. If memory serves correctly, > you shouldn't even use the FS for, say, a database on the same node > as the one you're exporting it from (I may be remembering that last > bit wrong, if it's important I can check, but I don't think it's > important to this discussion). > > The point here is that because you can't use it from more than one > node at a time when you are exporting it, there's just no point in > using GFS or GFS2 at all. This is not a high performance filesystem. > It was never intended to be. It's primary advantage is that it can > be used on multiple nodes in a cluster at the same time. If your > application (exporting it) takes that advantage away, you've lost > the main (possibly only) advantage in using it. > > Steve Whitehouse actually touched on this when he asked if you had > it set up active/passive, but I didn't see where you answered that > question (I may have missed it, though). > Yes, I was just trying to get my head around the configuration in this case to make sure that it was ok and so far I've not spotted anything that is obviously likely to make the performance via NFS slow. However this is more of an NFS than a GFS2 question I think - GFS2 seems to be providing reasonable performance and it is the addition of NFS that is causing the performance issue. I didn't see an answer to the question either, but I'm assuming that means that it is not a problem. Also, as regards GFS2 being high performance or not, that largely depends on what the workload is, and for the test which was outlined there is no reason why GFS2 should not provide pretty good performance and that seems to be bourne out by the reported results. > Now, if you still want to use GFS2, there are some things that I can > help you with from the GFS side that can speed it up considerably > (depending on the kind of access going on with it), which may actually > turn out to be the cause of the problem that you're experiencing > (interaction between NFS and GFS2), but I would first recommend that > you use a different filesystem all together if your only purpose for > it is NFS exporting. > I don't agree. Within the restrictions that we've noted above there is nothing wrong with exporting GFS2 via NFS, Steve. > I'm sorry I don't have much advice to give for the NFS specific > stuff, though. > > tg. > > On Tue, Sep 24, 2013 at 02:08:01PM +0200, Olivier Desport wrote: > > Le 24/09/2013 14:03, Raymond Burkholder a e'crit : > > > > One possibility is to try changing rsize and wsize at the client end. Is > > the 10G network used for NFS the same network as the one used for the > > iSCSI? > > I've tried several values of rsize and wsize and it doesn't > > change anything. iSCSI and NFS are not on the same VLAN. I'm > > testing with servers which are on the same chassis (NFS/GFS > > servers are on blades and client is a KVM guest with virtio > > network card). iSCSI and NFS use separates network in the > > chassis but each blade has to share iSCSI (50% to 100% > > of 10 Gbits/s) and LAN (50% to 100%). > > > > iperf between server and client : 3.4 Gbits/s (435 Mbytes/s). > > > > You could also try using Jumbo Frames on the network. > > > > I've tried but the NFS speed is the same. > > > > -- > > Linux-cluster mailing list > > Linux-cluster@xxxxxxxxxx > > https://www.redhat.com/mailman/listinfo/linux-cluster > -- Linux-cluster mailing list Linux-cluster@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/linux-cluster
