[p2pBandwidthLatencyTest] unknown test items.

Accelerated Computing CUDA CUDA Programming and Performance
1

In p2p test, we have p2ptest program and [p2pBandwidthLatencyTest] program.

I don’t understand the test item listed in this test:

  1. Unidirectional P2P=Disabled Bandwidth Matrix (GB/s )
    therefore, this is a bidirectional test?

  2. Unidirectional P2P=Enabled Bandwidth Matrix (GB/s )
    therefore, this is a unidirectional test? who is the source, who is the target?

  3. Bidirectional P2P=Disabled Bandwidth Matrix (GB/s )
    therefore, this is a unidirectional test? what’s the difference between this one and “Unidirectional P2P=Enabled Bandwidth Matrix”?

  4. Bidirectional P2P=Enabled Bandwidth Matrix (GB/s )
    when setting bidirectional P2P=enabled, what’s the setting for unidirectional p2p?

Sorry, I am a newbie, thank you for help.

2

Its a unidirectional test. That means data transfers are only flowing in one direction at a given time (from one device to the other). For this test, P2P is disabled.

It’s a unidirectional test. P2P is enabled. For a given point in the matrix, the GPU listed on one of the axes is the source, the GPU listed on the other axis for that point is the target.

It’s a bidirectional test. That means that transfers are initiated in both directions, from one GPU to the other and vice versa, with the transfers intended to run “at the same time”. For this test, P2P is disabled.

3

Thank you.

I have another question.

When we executed the p2ptest program, what’s the test condition compare to [p2pBandwidthLatencyTest] program?

In my p2ptest test result, I found the result didn’t match to [p2pBandwidthLatencyTest].

4

I don’t know what p2ptest is. It doesn’t appear to be a CUDA sample code like p2pBandwidthLatencyTest

5

I got this p2ptest program from document -

or should I post to another forum board?

6

What’s the title of the document? Who is the author? Where did you get it?

8

You seem to imply that this is a document that is internal to your place of employment or one that you received under NDA (non-disclosure agreement). If so, there is no point in bringing up the document or any of its content in a public forum. In fact, it should strictly be avoided.

You will need to work out any of the differences between multiple test programs on your own. Some ideas: If one program is focused on latency, and another on bandwidth, that might explain some of the differences. Or they are using different timing methodologies. Or employ different communication modes.

9

Sorry, I am a newbie to this forum, in order to avoid any inappropriate message, please delete this thread directly.

Thanks.