Hi Joseph,
On Thu, Feb 24, 2011 at 10:39:11AM -0500, Joseph Hardeman wrote:
> Hi guys,
>
> I have been asked if it were possible for Haproxy to receive traffic
> from servers for NFS over TCP, pass that traffic to a storage cluster
> and then the cluster send the data directly to the servers. Sort of the
> same method as LVS-DR would be.
>
> So the flow would go something like this:
>
> Server Pool ? Haproxy ? Storage Cluster VIP's
> ? ?
> ??????????????
>
> I know that Willy has tested Haproxy to 10G through put but they don't
> want Haproxy to become an I/O bottleneck as they scale their
> application. Or if someone has a recommendation to use instead of
> Haproxy or RR DNS for this sort of connections to the Storage Cluster I
> would love to hear it. They enjoy Haproxy on the other applications
> that they are using it for and it is working great.
I'm not fond of using proxies for protocols like NFS, for several reasons :
- proxies hide the source address, and servers often have per-source IP
settings ;
- NFS is extremely sensible to latency. Adding even one millisecond of
transit can become very noticeable in directory listings ;
- NFS uses long-lived connections, that means that you can't reconfigure
your proxy without breaking them ;
- at one point the proxy can still become a bottleneck, especially for
the outgoing traffic. On a local switched network, it can be enough
to plug a second network interface into the server and double its
bandwidth. When all the traffic passes through a proxy this is much
more complex.
If I were you, I'd either use some active/standby mechanism such as
heartbeat, keepalived or anything else to provide HA to the storage
cluster, or if you need some LB, just make use of LVS+DR. In both
cases, the downlink traffic does not experience any risk of bottleneck.
Regards,
Willy
