On May 28, 2014 8:27 PM, Ben Scott <[email protected]> wrote:
>
> On Wed, May 28, 2014 at 3:08 PM, Matthew W. Ross
> <[email protected]> wrote:
> >> Latency is the real killer.
> >
> > How do you reduce latency?
>
> Make the cable shorter.
>
> Ha ha, only serious.
>
My favorite kind of humor; it's funny because it's true. Humor noted and appreciated. Continue...
> For example: If you've got six servers, each with a gigabit link,
> plugged into your 2510, and then have a single gigabit link connecting
> your 2510 to your 5308xl, the traffic from those servers could
> overwhelm the uplink. The buffer on the 2510 will fill and then start
> dropping frames. The servers go into retransmit and make the problem
> worse. Everything goes into exponential backoff until aggregate
> transmit rate drops below the uplink speed. Things start working
> again, so transmit rate increases. Lather rinse repeat.
Is there a term for this kind of buffer overflow on a switch? I want to know so I know what to look for if/when this problem comes up again.
> Did you check port statistics on switches and servers? Check the
> logs on the switches? Are all the fault finders enabled on the
> switches?
Logs were not helpful, and counters did not show a lot of errors. I was not looking at any "fault finders," but I will be now.
> > On a similar note, We used to be running HP 2910s as our iSCSI traffic
> > connection switches between our 3 VMware hosts and our EqualLogic, but we
> > were again having latency issues.
>
> Define "latency issues". Please note that "latency" does not mean
> "network trouble". :)
Definition in this case: VMware logs complaining of latency issues while communicating with the EqualLogic SAN. I don't remember the exact error, but it was up to about .5 seconds in latency. Thus, I figured there was a network latency problem.
> > As recommended by our vendor, we upgraded
> > to Dell PowerConnect 6224s. On paper, these two switches are quite similar,
> > but the Dell switches handle the latency where the HPs were not.
>
> Uh.... "handle the latency"? :) Latency isn't something switches
> handle. It's something they cause, but usually not significantly.
>
> The specs for those switches do look pretty similar, so I too would
> expect them to handle similar loads.
I expected the HPs to work fine as well. I do not have a valid explanation on why the did not suffice while the Dells did.
> When you get into performance tuning of networks, configuration can
> matter a lot. Stuff that was wrong before but didn't cause a problem
> because it was small becomes a problem because now it's big. And
> different vendors each have different design assumptions, defaults,
> and quirks. As a result, you can get situations where vendor X works
> with vendor Y "out of the box", but fails with vendor Z. So you can
> replace all your X with Z, or you can adjust configuration. And then
> what do you do when you discover Z doesn't work with W?
Makes perfect sense. Time for me to study. Thanks Ben.
> -- Ben
>
>
--Matt
