Re: Jumbo Frames?
On Mon, Jan 30, 2006 at 01:58:56PM -0800, mike wrote:
> Thanks for the reply.
>
> I've tried that and now this is giving me confusing results:
>
> A has jumbo frames (mtu 9000)
> B has jumbo frames (mtu 9000)
> C has normal (mtu 1500)
>
> B -> A = 3.4MB/sec
> C -> A = around 50MB/sec
>
> I'm just scp'ing a 170 meg video file back and forth - it's over 10
> times slower... does anything stand out to you instantly as to what
> the problem would be?
C -> A will end up at 1500 since C can only do 1500, obviously. B -> A,
your switch is choking, is my *guess*.
I suspect your switch it not working as well as you think it is,
unfortunately. Is it firmware upgradeable? If so, check for updates.
Also check the Dell forums. I have a unmanaged Dell switch, it exhibits
the same behavior. Unfortunately, not upgradeable in my case ~:^(
You should get better than 50MB/s however. What are the NICs? Using a
hacked version of netcat, to eliminate disk I/O from the picture, I see
steady 90ish MB/s. Between two 64/66 Intel NICs across my Dell switch.
Between any other NIC hardware combos, I see 70-80 MB/s max. Actually,
the only other hardware I have that isn't 32bit is a Broadcom 57XX, also
64/66.
On a dual core, or even a single core, Opteron, the encryption overhead
would be minimal. The heavy encryption calculations comes at key
generation time, ie., when the connection is setup. Once it's going,
I'm going to estimate that the overhead from encryption is less than 1%
of CPU. The default encryption algorythms are pretty fast and efficient.
Try netcat and just time it, for cleaner results. After the first time
you run it, your caches should be warmed up, so you should get pretty
clean results of just the network xfer speed.
a
Reply to: