[HCoop-Discuss] local advocacy
David Crawshaw
crawshaw at hcoop.net
Thu May 1 19:06:12 EDT 2008
Adam Megacz <megacz at hcoop.net>:
> Are you still using BigPond? Pings from mire to them are only 250ms,
> which must make SSH really annoying, but isn't "high latency" from
> AFS's perspective (indeed, back in the 80's when AFS was developed
> that was probably considered a "normal" ping time).
>
> If you are experiencing packet loss, that may affect AFS more severely
> than TCP-based protocols. But you should switch providers if you're
> getting packet loss on a regular basis -- that is not normal.
No serious packet loss here. And while your theory about 250ms not
being slow for AFS is a good one, evidence seems to suggest that Rx
has trouble adapting to the link. Testing different file sizes:
75KB: scp is 4.7 times faster
161KB: scp is 2.5 times faster
1000KB: scp is 1.6 times faster
(This is with scp establishing a new connection each time, which has
some significant overhead, so the small file results are if anything,
understated.)
Now it might not be latency, it might be some crazy burst-iness of the
link, or this might just be noise because I didn't run the tests
enough times, but something seems amiss.
Anyway, I don't have the time to dedicate to an analysis right now.
d.
$ time cp hcoop-a4-draft.pdf /afs/hcoop.net/user/c/cr/crawshaw/public_html/
real 0m41.258s
user 0m0.001s
sys 0m0.011s
$ time scp hcoop-a4-draft.pdf mire:public_html
hcoop-a4-draft.pdf 100% 161KB 32.2KB/s 00:05
real 0m16.104s
user 0m0.074s
sys 0m0.040s
$ time cp smile-lecture1.pdf /afs/hcoop.net/user/c/cr/crawshaw/
real 3m59.237s
user 0m0.001s
sys 0m0.009s
$ time scp smile-lecture1.pdf mire:~
smile-lecture1.pdf 100% 1000KB 8.1KB/s 02:03
real 2m29.108s
user 0m0.121s
sys 0m0.068s
time cp rx-spec.pdf /afs/hcoop.net/user/c/cr/crawshaw/
cp: /afs/hcoop.net/user/c/cr/crawshaw/rx-spec.pdf: Operation timed out
real 0m57.294s
user 0m0.001s
sys 0m0.003s
$ time scp rx-spec.pdf mire:~
rx-spec.pdf 100% 75KB 75.0KB/s 00:01
real 0m12.135s
user 0m0.069s
sys 0m0.039s
More information about the HCoop-Discuss
mailing list