TCP Offload Engine

[Rootax]

Hi there,

The Fit-pc2 is a wonderful little box, but i think the network part of it can be better.

Indeed, with the 1.1ghz version, the cpu is at 100% and you can't go higher than 50-55mo/sec (approx, between 2 windows 7), even with jumboframes&co enabled.

So, in a kind of Fit-PC "server edition", a "real" GB network card with a good TOE Engine would be usefull, i think.

(And sorry for my bad english, i'm french)

[teetee1]

I am curious about the way you used for testing:

hardware: what type of switch and cable did you use? Which NIC and how was it configured on the other end?
software: which protocol did you run for the measurement and what software did you use? What's the CPU utilization on fit-pc2 during the file transfer process in the test?

When you said 50-55 mo/sec, did you mean 50-55 MB/sec? 50-55MB/sec reading sounds like the bottleneck speed of a hard drive if the data was copied from fit-pc2. It will be a very good measurement (or a fast hard drive) if the 50-55MB/s reading was from measuring data writing into fit-pc2.

[Rootax]

Yep i mean MB (i always talk in octet, that's why i used Mo), so it's the max speed of my WD Scorpio Blue 160gb.

I'm not too concerned about the speed, since the HD is the bottleneck, but it's the fact that the cpu is at 100% during the file transfer. My setup is :

- My PC (Q6600, 4gb of RAM, nVidia MB with nVidia network controller,fast raid5 array with an Areca controler, blabla)
- A Netgear GS108 8ports Gb switch, cat5e cables
- My Fit PC2, 1.1ghz version.
- Windows 7 v7600 on both.

From my pc, i map a network drive to the Fit-PC2, then copy-past the files from it (the data are on the fit-pc2, i copy past them to my main pc). So i guess the protocol used here is SMB v2.

While the copy is on, the "system.exe" process on the fit-pc2 is eating all the cpu, that's what is strange to me.

Sorry again for my english.

[teetee1]

I would use something that won't make cpu work 100% for file transfer(ex. ftp) and create a ram disk (~600mb maybe?) on fit-pc2 and read out the file from the other machine to estimate the NIC speed. That way the cpu and the hard drive won't be the bottleneck and therefore it's more likely to measure the true network transfer speed.

ex.
fit-pc2(ftp server) - gigabit switch - ftp client

[madgrt]

I would use something that won't make cpu work 100% for file transfer(ex. ftp) and create a ram disk (~600mb maybe?) on fit-pc2 and read out the file from the other machine to estimate the NIC speed. That way the cpu and the hard drive won't be the bottleneck and therefore it's more likely to measure the true network transfer speed.

ex.
fit-pc2(ftp server) - gigabit switch - ftp client

Transfering between RAM disks is a good idea for the test, but how about just using a crossover cable and remove the switch altogether. If you make one for the test, (instead of buying one), don't make it shorter than 1 meter.
FTP would be easy to see a result at the end of the transfer with "hash" printing turned on.

[Rootax]

As i said, the speed is ok, since it is as fast as the fit-pc2 hard drive. I'm more concerned about the 100% cpu usage. Using ftp is maybe better on that point, but in a everyday usage, a copy/past from a network drive is much more simple...

[madgrt]

As i said, the speed is ok, since it is as fast as the fit-pc2 hard drive. I'm more concerned about the 100% cpu usage.

What cpu usage do you get on the quad box when you do the transfers? Is it proportional cpu usage?
The other thing is that the Window 7 + the drivers you are using are still technically in beta, even though 7600 must nearly be RTM. It's not the greatest test base, don't you have something else you can boot into?

[Rootax]

I had the fit-pc2 under vista sp2 before (my quad box too), and it was the same thing.

I'm not at home now, but i think the cpu occupation on the quad was ridiculous (i mean very low by that), not proportional at all.

i'll test with the ftp protocol and see the results.