linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* Why throughput increases as MTU size is increased
@ 2003-04-30 14:10 Shesha
  2003-04-30 15:30 ` Stephen Satchell
  0 siblings, 1 reply; 2+ messages in thread
From: Shesha @ 2003-04-30 14:10 UTC (permalink / raw)
  To: linux-kernel, kernelnewbies

Hi All,
 I have a question which I am not able to answer myself.I request you all to
give
me some input.
 When I measure the performance of iSCSI on XScale with MTU size = 1500 bytes,
a throughput of 32 Mbps was observed. As the MTU size was increased, the
throughput also increased.
1500 -> 32 Mbps
3000 -> 56
4500 -> 80
6000 -> 100
7500 -> 108
9000 -> 108

Actually the throughput saturates. I thought, the per-packet overhead
decreases as the MTU
size increases. This contributes to the performance improvement. And the
saturation is achieved because, the iSCSI max PDU size is 8k. Even if
we increase the MTU size beyond 8k, we will not see any change because,
iSCSI
devivers a max of 8K PDU to TCP. Therefore a saturation in throughput is
observed.

But  the question is, Am I thinking correctly?
secondly, if yes,does the per-packet over head decrease the performance so
much. we are observing, somewhere like 4 times improvement in throughput. Can
there be any other reason for this observation. 

Thanking you 
Shesha





^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: Why throughput increases as MTU size is increased
  2003-04-30 14:10 Why throughput increases as MTU size is increased Shesha
@ 2003-04-30 15:30 ` Stephen Satchell
  0 siblings, 0 replies; 2+ messages in thread
From: Stephen Satchell @ 2003-04-30 15:30 UTC (permalink / raw)
  To: Shesha, linux-kernel, kernelnewbies

At 07:10 AM 4/30/03 -0700, Shesha@asu.edu wrote:
>  When I measure the performance of iSCSI on XScale with MTU size = 1500 
> bytes,
>a throughput of 32 Mbps was observed. As the MTU size was increased, the
>throughput also increased.
>1500 -> 32 Mbps
>3000 -> 56
>4500 -> 80
>6000 -> 100
>7500 -> 108
>9000 -> 108

The non-linear increase as you change MTU is the dead give-away:  the 
slow-start algorithm is working just fine.  The reason you didn't get 
double the throughput as you doubled the MTU is that the slow-start 
algorithm ramps based on packet count, so as you increase the MTU the 
packet rate slope remains reasonably constant.  If you have the source 
code, you can turn off slow-start using setsocketopt(), and see what happens.

Have you considered trying a longer file?


--
X -> unknown; Spurt -> drip of water under pressure
Expert -> X-Spurt -> Unknown drip under pressure.


^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2003-04-30 15:18 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2003-04-30 14:10 Why throughput increases as MTU size is increased Shesha
2003-04-30 15:30 ` Stephen Satchell

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).