[Gluster-users] Switch recommendations
d.a.bretherton at reading.ac.uk
Wed Feb 1 15:23:03 UTC 2012
Thanks for this detailed posting. Answers are in line.
On 01/31/2012 12:51 PM, John Lauro wrote:
> The latency shouldn't be that bad to really be noticeable unless the links
> are really saturated. One thing I don't like about the Dell 5424 switches
> (but otherwise great switches, and you can design around the flaw if you
> are aware of it) is the buffers are a little small and they will send
> pause packets fairly easily. Make sure you don't have anything that does
> a moderate amount of traffic on 100mbps port, as it can easily backup all
> the other ports, especially if it's on a different switch and the switch
> to switch port is paused...
I didn't realise 100mbps connections could cause a problem. Fortunately
there aren't any, I just checked; no orange lights anywhere. I also
checked the port statistics and found no pause frame errors on any of
> Run some ping tests between servers on the same switch, and servers that
> are going over your switch to switch link. It could be you just have a
> bad cable somewhere...
> Generally I use the following for ping tests:
> ping serverip -s 16000 -c 1000 -q -A
> PING xxx.xxx.xxx.xxx (xxx.xxx.xxx.xxx) 8000(8028) bytes of data.
> --- xxx.xxx.xxx.xxx ping statistics ---
> 1000 packets transmitted, 1000 received, 0% packet loss, time 1035ms
> rtt min/avg/max/mdev = 0.631/0.834/5.110/0.235 ms, ipg/ewma 1.036/0.802 ms
I ran the test on four clients, 3 times each, for all the servers. Most
results were similar to your example above, with some tests taking 2 or
3 times longer. For three of the clients, ping tests on servers on
different switches took roughly twice as long on average, which suggests
that connecting everything to the same switch or using stackable
switches would be a good idea as I suspected. However the test results
for the fourth client showed no pattern at all, so I probably need to do
some more tests on more of the clients to be confident enough to draw
any firm conclusions. Thanks for pointing me in the right direction.
More information about the Gluster-users