Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 20 Dec 2006 01:26:34 +0100
From:      "Raphael H. Becker" <rabe@p-i-n.com>
To:        freebsd-hackers@freebsd.org
Subject:   Re: FreeBSD-6.1/amd64 bge(4) driver performance problems
Message-ID:  <20061220002634.GB75640@p-i-n.com>
In-Reply-To: <20061128204732.3e8790fd.vlady@gbservices.biz>
References:  <20061128204732.3e8790fd.vlady@gbservices.biz>

next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, Nov 28, 2006 at 08:47:32PM +0200, Vladimir Terziev wrote:
> 	It seems the bge(4) driver has severe performance problems (may be especially in my configuration).
> 	I tried test scp(1) to a remote machine, using one of the BCM5721 NICs. The average speed which has been reached was 200kBps.

I had this problem just before 5.3-RELEASE and still have this with some
combinations of interfaces, switches, ...

http://lists.freebsd.org/pipermail/freebsd-net/2004-September/005017.html

In short: 
* Testing with FTP'ing a large file between hosts.
* 4 (or more) nodes on the net:
  - 2x fxp  10/100
  - 2x bge  10/100/1000
* dumb GBit switch

case 1: ftp from fxp/100 to bge/1000 --> fine (11MBytes/sec)
case 2: ftp from fxp/100 to fxp/100 --> fine (11MBytes/sec)
case 3: ftp from bge/1000 to bge/1000 --> fine (3 parallel transfers on
a cached file, download to /dev/null): around 90MBytes/sec and more 

case 4: ftp from bge/1000 to fxp/100 --> poor: <1MB/sec at the
beginning, falling back to some 100k/sec

Every combination of sender/receiver was tested so there was not a
specific combination of machines which failed.

Summary:
* Every single interface is able to send and to receive with the full bandwidth.
* Every Switch-Port is able to transfer full 1000MBit
* Every cable works perfektly (1000MBit/sec indicates)


The interesiting case is case 3: If the sender is at 1000MBit and the
receiver is on 1000 too, there is no sign of bandwidth-problems, this
only accours on the 1000->100 case.

This was in pre-5-STABLE.



Recently I had this again with bgeX to bgeY on two Dell machines running
6.2-RC1. But in this special case bot machines are connected to two GBit
Switches with 2 interfaces  so both hosts can connect to each other
using switch1 or switch2. --> switch1 (netgear-something) got changed
with a H* p*o-curve  --> solved.

--> Try another switch.


HTH
Raphael




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20061220002634.GB75640>