SV: SV: [Gluster-devel] Re-exporting NFS to vmware

[Fredrik Widlund]

The point was that trying to make NFS a distributed protocol that fits the gluster design does not make any sense at all, so the choices probably are 
1) use a proxy, and standard hardware today can handle up to 10GbE loads which should be enough, and if it's not, just scale out horizontally in additional proxies. The problem today is that the NFS<->gluster chain infers a performance penalty/bottleneck, not that the concept of using a proxy is flawed.
2) use gluster on the clients

I'm not sure which stance you are referring to, I'm just saying that on an fairly updated kernel it works out of the box. The gluster nfs is said to have better performance compared to the kernel nfs, and looking at the IO pattern it makes a _lot_ more sense using the gluster nfs (using the kernel nfs results in very weird pattern, with files being reopened multiple times for a single read, etc), but in our tests both top out on around 500MB/s with a concurrent load anyways. With the gluster nfs we top out maybe 10% higher, but the difference is marginal.

As far as reliability is concerned, in our experience unfs seem to be by far the worst choice.

Kind regards,
Fredrik Widlund

-----Ursprungligt meddelande-----
Från: gluster-devel-bounces+fredrik.widlund=qbrick.com at nongnu.org [mailto:gluster-devel-bounces+fredrik.widlund=qbrick.com at nongnu.org] För Gordan Bobic
Skickat: den 6 januari 2011 18:00
Till: 'gluster-devel at nongnu.org'
Ämne: Re: SV: [Gluster-devel] Re-exporting NFS to vmware

Fredrik Widlund wrote:
> If you're re-exporting a gluster filesystem, the re-exporting node will act as a proxy. As a concept this is fairly natural, and in itself it shouldn't be a problem. 

It's natural, but Sylar was saying it's inefficient - which it is, 
because you end up hammering one node for all the files, including the 
ones that aren't on it.

> And I did say that it is possible to re-export a FUSE filesystem, not impossible. 

IIRC it required fuse patches that weren't in the mainline, and when I 
tried it, it caused reliability problems and some corruption issues that 
went away when I switched to unfsd. Unless I missed something recently, 
I don't think the stance on kernel nfs exporting fuse file systems has 
changed, but I'd love to hear otherwise.

Gordan

_______________________________________________
Gluster-devel mailing list
Gluster-devel at nongnu.org
http://lists.nongnu.org/mailman/listinfo/gluster-devel