[Gluster-users] Switch recommendations

[John Lauro]

> Incidentally, using a LAG is not going to improve your latency, and
> depending on workload it might not improve your peak throughput either.
> The
> algorithm which distributes the frames between links is usually a hash
of
> src/dest MAC and/or IP addresses, so if all the traffic is between one
pair
> of hosts, it will all go down one link.  (Anything else might result in
> packet reordering)

True, if all the traffic is between exactly two nodes.  However, assuming
multiple servers and clients it can improve latency.

It depends on the cause of the latency problems.  If a link is overloaded,
or even if that link itself isn't overloaded but a port behind it is (ie:
100mbps port), then traffic taking the other link will not be paused and
so the latency will improve.


Some network monitoring on the switch ports to get utilization graphs of
the ports would be good.  Especially statistics on received and
transmitted pause frames and they are a real latency killer.