Intel e1000 gigabit poor performance

Hi all,

Is it even possible to achieve gigabit speeds using intel's 82540EM
chipsets on rh9?

I installed the latest e1000 driver from Intel, tried some of the tips
listed here, etc., but the fastest I get is ~4.0M. Also, the card "bursts"
data, i.e. 0 to 4m,.

If anyone knows anyone who got it to work, or maybe a better
troubleshooting guide than intel's own website please share!

R.

Spongebob <(E-Mail Removed)> wrote:
> Is it even possible to achieve gigabit speeds using intel's 82540EM
> chipsets on rh9?

> I installed the latest e1000 driver from Intel, tried some of the
> tips listed here, etc., but the fastest I get is ~4.0M. Also, the
> card "bursts" data, i.e. 0 to 4m,.

I've no idea if my troubleshooting is any better than Intel's, but can
you also tell us what the CPU util was, what window sizes you were
using, were you doing "plain" netowrk transfers or were there
filesystems invovled, that sort of stuff?

rick jones
http://www.netperf.org/
--
oxymoron n, Hummer H2 with California Save Our Coasts and Oceans plates
these opinions are mine, all mine; HP might not want them anyway...
feel free to post, OR email to raj in cup.hp.com but NOT BOTH...

On Mon, 27 Oct 2003 20:12:15 +0000, Rick Jones wrote:

> Spongebob <(E-Mail Removed)> wrote:
>> Is it even possible to achieve gigabit speeds using intel's 82540EM
>> chipsets on rh9?

> I've no idea if my troubleshooting is any better than Intel's, but can
> you also tell us what the CPU util was, what window sizes you were
> using, were you doing "plain" netowrk transfers or were there
> filesystems invovled, that sort of stuff?
>
> rick jones
> http://www.netperf.org/


I'm not sure what you are asking. CPU utilization is negligible, but "window
sizes"?

I got the procfgd/procfg prgrams from intel to work, and this is what I
get...

spongebob:~> procfg -U ral -P sausage666 adapters

UNIQUE NAME SYSTEM NAME LINK SPEED DUPLEX TEAM TYPE
------------- ------------- ---- ----- ------ -------- ---------
01:00.0 eth0 UP 1000 FULL Intel1000



I am running rh9 on a 1GHz intel.
I am moving/reading/copying files from a NAS NFS server, also on the
gigabit network.
I have the same problem when connecting (through the network) to another
computer (dual 2.5GHz) with the same intel chipset for a NIC.


I guess what I would really like to know, is if anyone have got this card
to reach 1Gigabit speeds. If so I can go through the trouble of fixing
them.

TIA.

R.

Spongebob wrote:
> Hi all,
>
> Is it even possible to achieve gigabit speeds using intel's 82540EM
> chipsets on rh9?
>
> I installed the latest e1000 driver from Intel, tried some of the tips
> listed here, etc., but the fastest I get is ~4.0M. Also, the card "bursts"
> data, i.e. 0 to 4m,.
>
> If anyone knows anyone who got it to work, or maybe a better
> troubleshooting guide than intel's own website please share!
>
> R.
>
>

I recently put a Intel box on our data centre with an Intel gigabit
ethernet card.

Before putting the box in the data centre, I tested at the office first
( the office network does not support gigabit yet ), put options in the
/etc/modules.conf to force it to 100Mb full-duplex, but the performance
was very poor. It turns out /sbin/ifconfig was showing lots of "errors"
on the interface on TX, even while the interface was add 100Mb full-duplex.

Had to remove all the options I put in /etc/modules.conf.
After removing all the forced options I put, the interface was still at
100Mb full-duplex, but no more TX erorrs!

I tried both the driver that came with the latest 2.4.x kernel, and the
one from Intel, which I think IS the one built into the latest 2.4.x
kernel anyway ( same module options ), but it was not there in the
earlier 2.4.x kernels.

After a few more rounds of testing, put it on the data centre. Only
after a few days did I realise it was not at running at full-duplex at
the data centre. Damn! Could not reboot it now since it is live.

Spongebob <(E-Mail Removed)> wrote:
> I'm not sure what you are asking. CPU utilization is negligible, but "window
> sizes"?

Windowing - part of the basic TCP flow control mechanism, intended to
prevent a sending TCP from overwhelming a receiving TCP. One of the
(many) possible limits to the performance of a TCP connection is the
window size divided by the round-trip time.

If you are running netperf, you might try different values for -s and -S.

> I am running rh9 on a 1GHz intel.
> I am moving/reading/copying files from a NAS NFS server, also on the
> gigabit network.

NFS isn't "really" a bulk-transfer even with read/write. It is a
request/response protocol and as such can be quite sensitive to
latency. Some driver/NIC combinations have been known to set their
default interrupt throttles to support bulk transfers using TCP, and
leave request/response performance (eg netperf TCP_RR) suffering.

Also, NFS may or may not be running over TCP.

I would eliminate fs performance and NFS and such by running netperf
and see how fast you can go with that.

rick jones
--
firebug n, the idiot who tosses a lit cigarette out his car window
these opinions are mine, all mine; HP might not want them anyway...
feel free to post, OR email to raj in cup.hp.com but NOT BOTH...

In article <(E-Mail Removed)>, Spongebob wrote:
>
> I am running rh9 on a 1GHz intel.
> I am moving/reading/copying files from a NAS NFS server, also on the
> gigabit network.
> I have the same problem when connecting (through the network) to another
> computer (dual 2.5GHz) with the same intel chipset for a NIC.

To really test the NIC, you need to eliminate as many other variables
as possible, and that NAS and NFS are two big ones. Look for
programs like ttcp, netperf, netpipe and others that will *just* test
the network connection.

If you eliminate and/or optimize the NIC, then you can start fiddling
with NFS options and the NAS setup.

--
Joshua Baker-LePain
Department of Biomedical Engineering
Duke University

> I'm not sure what you are asking. CPU utilization is negligible, but
"window
> sizes"?

Ethernet was originally intended to run at 10 megabits/second.
Increasing the speed by a factor of 100 without increasing the maximum
frame (window) size was a very "kludgy" way to go about things, and is far
less than optimal.

With a maximum frame size of 1500 bytes (12 kilobits), at full gigabit
speed, you're talking about transmitting/receiving nearly 100,000 packets
per second, and there's a LOT more overhead in processing 100,000 small
packets vs. 5,000 large packets.

If you have a switch that supports "jumbo frames", you can increase the
maximum frame size to 64K, which can reduce the number of packets by up to
a factor of 40, which makes a big difference.

If you don't understand what I've just said, there are plenty of
tutorials and primers on ethernet that will help you out. On to your next
comments:

> I am running rh9 on a 1GHz intel.
> I am moving/reading/copying files from a NAS NFS server, also on the
> gigabit network.

Then you're not testing the speed of the network card, you're testing
the speed of hard drives, CPU's, etc.. Think about it: If you were to
transfer a full gigabit/second on an NFS transaction, the disks at *both
ends* would have to be sustaining more than 100 megabytes/second. Can
your disks really read and write 100 MB/second? Can your NAS really
*process* 100 megabytes of NFS information per second?

> I guess what I would really like to know, is if anyone have got this
card
> to reach 1Gigabit speeds. If so I can go through the trouble of fixing
> them.

You can certainly get the card to reach gigabit speed. Getting NFS
transfers to work at gigabit speeds, however, is something you probably
shouldn't hold your breath for.

steve