netdev
[Top] [All Lists]

Re: jumbo on 8169

To: Luca Bortot <l.bortot@xxxxxxx>
Subject: Re: jumbo on 8169
From: Francois Romieu <romieu@xxxxxxxxxxxxx>
Date: Mon, 27 Dec 2004 17:38:02 +0100
Cc: netdev@xxxxxxxxxxx
In-reply-to: <41D01562.4090606@xxxxxxx>
References: <41CFF27A.2070008@xxxxxxx> <20041227123136.GA25187@xxxxxxxxxxxxxxxxxxxxxxxxxx> <41D01562.4090606@xxxxxxx>
Resent-date: Mon, 27 Dec 2004 18:47:00 +0100
Resent-from: romieu@xxxxxxxxxxxxx
Resent-message-id: <200412271747.iBRHl0CM028784@xxxxxxxxxxxxxxxxxxxxxxxxxx>
Resent-to: netdev@xxxxxxxxxxx
Sender: netdev-bounce@xxxxxxxxxxx
User-agent: Mutt/1.4.1i
Luca Bortot <l.bortot@xxxxxxx> :
[...]
> based on a simple tcp test I made (writes zeroes to a socket in 32Kb 
> blocks and prints the write speed), these are the results (win box cpu 
> not reported - always under 10% load):
> 
> BEFORE PATCH (mtu 1500)
> speed ~38 MB/s
> cpu idle 10%
> cpu system 90%
> 
> AFTER PATCH (mtu 7000)
> speed ~45MB/s
> cpu idle 40%
> cpu system 60%

TSO may make a difference for a TCP test. See ethtool help to enable it.
You can experiment with Tx csum/SG as well.

I'll welcome a complete dmesg and lspci -vx as they pretty well describe
the working combinations.

--
Ueimor

<Prev in Thread] Current Thread [Next in Thread>