RE: jumbo frames

["Tony Hain" <alh-ietf () tndh net>]

Roeland you are talking about jumbo frames from the end system lan, while
John is talking about only using the jumbo frames between the routers. My
point was that in John's environment the packets will all be 1500 since the
packets are restricted to that size just to get to the router with the GE
interface. I understand that there are perf gains as long as the entire path
supports the larger packets, but I don't understand the claim that having a
bigger pipe in the middle helps.

Tony

-----Original Message-----
From: owner-nanog () merit edu [mailto:owner-nanog () merit edu]On Behalf Of
Roeland Meyer
Sent: Thursday, April 26, 2001 12:13 PM
To: 'alh-ietf () tndh net'; John Fraizer; Paul Lantinga
Cc: nanog () merit edu
Subject: RE: jumbo frames


> From: Tony Hain [mailto:alh-ietf () tndh net]
> Sent: Thursday, April 26, 2001 11:47 AM
>
> April 26, 2001 9:29 AM John Fraizer wrote:
> > We only have jumbo frames enabled on router<->router links.
>  The GigE
> > ports facing the aggregation switches runs standard 1500 MTU.
>
> Hence my original question. Packets across the GE will be
> 1500 unless you
> are packing them.
>
> April 25, 2001 8:10 PM John Fraizer wrote:
> > Partially because I can.  Partially because there seems to be a
> > performance increase when you start stuffing the pipe.
>
> Assuming you are just passing the packets as received from
> the aggregation
> switch, this would only happen if your router hardware was better at
> managing jumbo buffer allocations than 1500B ones. Clearly it
> will waste
> large chunks of memory, so do you have measurements to show the actual
> performance increase?

This depends on the type of traffic. We use it to enhance performance of the
data tier. We've jiggered the TCP/IP stacks for ~4500 byte packets and have
knee-capped the slow-start algorithm (which doesn't make sense in a pure
switched environment anyway). What we then wind up with, is a dedicated data
LAN between our application servers and the Oracle database servers. It
comes out to about an order of magnitude increase in performance and SQL
query responsiveness. At first we went to jumbo packets. We saw such a
radical improvement that we started investigating and found the slow-start
issue. Jumbo packets are one way around the slow-start problem if you can't
jigger the stack itself. Most of the queries are reasonably short so we saw
some serious improvements by killing the slow-start. If you can modify your
IP stack then it still pays to use jumbo packets because you reduce the
overhead on the wire.

We got sufficient performance improvement that we were able to defer GigE
implementation, at some sites. Those sites are using switched FDX 100baseTX.