how to improve the throughput of linux network

how to improve the throughput of linux network

[zhengchuanbo]

we use linux as our router. i just tested the performance of the router with smartbits, and i found that the throughput of 64byte frame is only 25%, about 35kpps. 
someone mentioned that the throughput of 64byte frame could reach 70kpps.so i wish i could improve the performance of our router,but i don't know how to do that.
i looked for some solution,and found some article mentioned the NAPI. it changed the driver to polling mode,so that the interrupt is no too much. but i could not find  drivers for our router.(eepro100 card). has the polling mode driver been used in linux?
i think there should be some other methods to improve the performance.but what is the most efficient one?
thanks for help. please cc.

chuanbo zheng

Re: how to improve the throughput of linux network

[Jason Lunz]

zhengcb@netpower.com.cn said:
> we use linux as our router. i just tested the performance of the
> router with smartbits, and i found that the throughput of 64byte frame
> is only 25%, about 35kpps. 
>
> someone mentioned that the throughput of 64byte frame could reach
> 70kpps.so i wish i could improve the performance of our router,but i
> don't know how to do that.
>
> i looked for some solution, and found some article mentioned the NAPI.
> it changed the driver to polling mode,so that the interrupt is no too
> much. but i could not find  drivers for our router.(eepro100 card).
> has the polling mode driver been used in linux?

I made NAPI core and driver patches for linux 2.4; they're at
http://gtf.org/lunz/linux/net/. There's an eepro100 driver there.

Please send me (and the linux netdev list) the results of any testing
you do.

Jason

Re: how to improve the throughput of linux network

[Jason Lunz]

On Tue, Jul 16, 2002 at  6:59PM +0200, Filip Sneppe (Yucom) wrote:
> Hey, that's funny. I downloaded NAPI patches from:
> ftp://robur.slu.se/pub/Linux/net-development/NAPI/

Yes, that's where the NAPI stuff originates. The core patch from that
site is what made it into 2.5, and I backported it to 2.4 from there.
There are no substantial differences between the three.

> How different are those patches from yours ? Did you just make
> sure they all applied cleanly ?

While the core is mostly a matter of changing patch offsets, the driver
patches are still a moving target. There are two different flavors of
the e1000 NAPI patch, for example. The most recent tulip patch from the
above FTP site is missing tulip_misc.c, so it must be replaced with a
version from another patch.  That file really is only code to report
stats in /proc, though. There's yet another napified tulip driver that
jamal posted a link to; I have yet to investigate that one.

But to answer your question, yes. I just found the most recent versions
of various napi efforts and made them apply to a recent 2.4 kernel so
people can easily try out the new net core.

Jason

RE:Re: Re: how to improve the throughput of linux network

[Hell.Surfers]

[-- Attachment #1: Type: text/plain, Size: 408 bytes --]

well only if it was used little amounts, like once every hour, it would dynamically unload in between, and dont modules, recompile on each use? I was told they were, although im more of a sound driver person.

- "Yes. Yes. OKAY.", Installing Microsoft software has always felt like an argument with your Mum (alledgedly).

On Mon, 15 Jul 2002 22:36:52 -0400 (EDT) Mark Hahn <hahn@physics.mcmaster.ca> wrote:

[-- Attachment #2: Type: message/rfc822, Size: 1103 bytes --]

From: Mark Hahn <hahn@physics.mcmaster.ca>
To: <Hell.Surfers@cwctv.net>
Subject: RE:Re: Re: how to improve the throughput of linux network
Date: Mon, 15 Jul 2002 22:36:52 -0400 (EDT)
Message-ID: <Pine.LNX.4.33.0207152236400.25831-100000@coffee.psychology.mcmaster.ca>

> Using it as a module would only slow you down if netfilter is required, because itwould load and unload, contstantly, causing you to remember what the 486 was like.

why would it get unloaded repeatedly?

Re: how to improve the throughput of linux network

[Thunder from the hill]

Hi,

On Tue, 16 Jul 2002 Hell.Surfers@cwctv.net wrote:
> well only if it was used little amounts, like once every hour, it would
> dynamically unload in between,

That's ok then. It shouldn't produce significant overhead. But on the 
routers that I run I have either no netfilters at all, or they keep 
running, so even if they were a module, they'd never have any time to 
unload.

> and dont modules, recompile on each use?

No, because that would be pointless. And it would keep the kernel from 
running properly on systems without a c compiler.

							Regards,
							Thunder
-- 
(Use http://www.ebb.org/ungeek if you can't decode)
------BEGIN GEEK CODE BLOCK------
Version: 3.12
GCS/E/G/S/AT d- s++:-- a? C++$ ULAVHI++++$ P++$ L++++(+++++)$ E W-$
N--- o?  K? w-- O- M V$ PS+ PE- Y- PGP+ t+ 5+ X+ R- !tv b++ DI? !D G
e++++ h* r--- y- 
------END GEEK CODE BLOCK------

Re: how to improve the throughput of linux network

[Eric W. Biederman]

Thunder from the hill <thunder@ngforever.de> writes:

> Hi,
> 
> On Tue, 16 Jul 2002 Hell.Surfers@cwctv.net wrote:
> > well only if it was used little amounts, like once every hour, it would
> > dynamically unload in between,
> 
> That's ok then. It shouldn't produce significant overhead. But on the 
> routers that I run I have either no netfilters at all, or they keep 
> running, so even if they were a module, they'd never have any time to 
> unload.

Only if you have a cron job running rmmod -a will the module unload.
The kernel never unloads modules without being asked.

Eric

Re: Re: how to improve the throughput of linux network

[zhengchuanbo]

>On Tue, 2002-07-16 at 03:13, zhengchuanbo wrote:
>> 
>> we use linux as our router. i just tested the performance of the router with smartbits, and i found that the throughput of 64byte 
>> i looked for some solution,and found some article mentioned the NAPI. it changed the driver to polling mode,so that the interrupt is no too much. but i could not find  drivers for our router.(eepro100 card). has the polling mode driver been used in linux?

>Try this url:
>
>ftp://robur.slu.se/pub/Linux/net-development/NAPI/
>
>Also check out 2.5 kernels - they alreacdy have NAPI, there's is at
>least some documentation under linux/Documentation/
>
>Could you put some numbers online after your tests ?
>I already have this page, I don't have a Smartbits, though :-):
>
>http://www.filip.sneppe.yucom.be/linux/netfilter/performance/benchmarks.htm
>
>Regards,
>Filip

i got the patch for NAPI,and patched it on linux2.4.18. it worked. the throughput of 128bytes frame improve from 60% to more than 90%. it seems that it has no influnce to frames bigger than 256.

but there is still some problem. when i tested the throught of 64bytes frame,some error occured. in the begining it works well. but after several times of try the linux router can not receive any packets at all.(i found that by run ifconfig when the smartbits is testing). for the other frames it worked very well.

so what's wrong with my test? is there some problem with the patch?

regards,
chuanbo zheng

Re: how to improve the throughput of linux network

[Jason Lunz]

In gmane.linux.kernel, you wrote:
> i got the patch for NAPI,and patched it on linux2.4.18. it worked. the
> throughput of 128bytes frame improve from 60% to more than 90%. it
> seems that it has no influnce to frames bigger than 256.
> 
> but there is still some problem. when i tested the throught of 64bytes
> frame,some error occured. in the begining it works well. but after
> several times of try the linux router can not receive any packets at
> all.(i found that by run ifconfig when the smartbits is testing). for
> the other frames it worked very well.
> 
> so what's wrong with my test? is there some problem with the patch?

Quite possibly. Are you still using the eepro100 NAPI driver? I doubt
that it's gotten wide testing.

The patch at http://gtf.org/lunz/linux/net/ comes from here:
ftp://robur.slu.se/pub/Linux/net-development/NAPI/eepro100/eepro100-napi-020619.tar.gz

The NAPI conversion was done by fxzhang@ict.ac.cn (see the README); you
may want to ask him about problems you're having.

-- 
Jason Lunz			Reflex Security
lunz@reflexsecurity.com		http://www.reflexsecurity.com/