Subject: Re: [vserver] NFS shares or iSCSI
From: "Michael S. Zick" <mszick@morethan.org>
Date: Mon, 12 Dec 2011 08:22:07 -0600

On Mon December 12 2011, Eugen Leitl wrote:
> On Mon, Dec 12, 2011 at 02:23:39PM +0100, Herbert Poetzl wrote:
> 
> > so, what performance do you get over this setup
> > ignoring any filesystem overhead? 50MB/s 100MB/s ?
> 
> I will have to nuke the system in order to make it
> production, but it should be able to saturate GBit/s.
> I don't know how many IOPS it would have, but it
> is likely to have at least an order of magnitude 
> better performance than the local RAID 1 pair
> of spindles.
> 
> > how many packets (I/O operations) can you transmit
> > in a second? 10k? 20k? ...
> 
> I'm expecting about 1 k IOPS, but I haven't measured
> that yet.
>  
> > I guess a single SSD, attached via SATA 6GB locally
> > will seriously outperform whatever solution you have
> > in mind, but of course, you cannot share that between
> > different hosts easily
> 
> I don't have a spare slot in the server (it's an old
> SunFire X2100 M2) and I wouldn't be able to afford 2 TByte
> of SSD in any case.
>

Why not use a small SATA-3 SSD on each of the host machines
as a cache in front of your network file or block system?

Something like:
http://www.kingston.com/ssd/kc100.asp

Which would (in the 120Gbyte size) run you about $2/Gbyte and
give each client machine a local 90K IOPS cache in front of
your old server network connected at only 1Gbs.
Caching locally about 5% of the network data would greatly
reduce your network load.

Unless you can upgrade your networking to 10Gbs fiber, I think
this would give you about the next best performance.

Mike 
> I think I'll go with an NFS. I have some 3 consumer NAS
> which export NFS as well, and I haven't done iSCSI yet
> so NFS is definitely simpler to deploy.
>