Yet another bandwidth question

Hi all,

I did 5 (extended) tests with the bandwidthTest example and plotted the median.

The workstation is a Core2Duo E6600@2.4GHz, 2x 1GB Corsair Value Select@533 MHz, ASUS P5B Deluxe (P965,ICH8R), 2x ASUS ENGTX280 and runs OpenSuse 10.2 (Kernel 2.6.26.5) with driver version 177.67. (Yes, the FSB is really slow. The 2nd card is only connected with 4 PCIe lanes.)

[attachment=7537:attachment]

[attachment=7538:attachment]

Two questions came up when looking at the results:

  • What is causing the step at 1MB for pageable memory? (Hardware or software? Does the P965 open more PCIe lanes for larger data packages?)
  • Why is the device–>host bandwidth higher than host–> device for the 2nd device and the other way round for #1?

I would really appreciate if somebody would have an explanation.

Thanks,
Tobi
htod.png
dtoh.png

As of now, any input will be appreciated :-)

tohahn: These plots are really painful to read. Try to recreate them in a more clear way.

Pageable memory performance isn’t really meaningful, as it’s dependent on OS, kernel, where memory is laid out, whether or not stuff is paged, etc. etc. etc.

The other one is basically that your BIOS is doing something strange. Lots of BIOSes do strange things. Try updating it, it might help.

@tmurray: As I have 2 GB of GPU memory and only 2GB of RAM, I can’t allocate everything as pinned memory. Thus, pageable memory performance is also interesting to me. Gaining a 50% higher bandwidth when transferring 16 MB compared to 1 MB is quite something.

Thanks for the BIOS hint, I’ll have a try on it.

@Mafos: A linear x-axis wouldn’t show anything and not labeling the axis with 2^x would generate huge numbers. So, I don’t know how to make it easier for you.

What you can basically read from them is a maximum pinned memory transfer speed of

                |   #1    |  #2

-----------------------------------

host-to-device  |  2500   |  650

device-to-host  |  1900   |  800

and I think, the steps at 1 MB (=2^10 kB) are clearly to be seen in the first plot.

Well, they’re fine, but I would differ the 4 almost-the-same sybmols and use higher resolution to make it a bit more clear It’s just a suggestion though.

To put an end to this issue: The BIOS was updated, new test were conducted, but the plots look the same. I got a 1% higher bandwidth for device 0 but that’s all.
Guess, I’ll have to accept the results as they are and don’t aks for an explanation.