search

ComputationalRadiationPhysics / jungfrau-photoncounter

Conversion of Jungfrau pixel detector data to photon count rate
GNU General Public License v3.0
2 stars 2 forks source link

Upload/Download/Kernel Parallelism #41

Closed kloppstock closed 6 years ago

kloppstock commented 6 years ago

Check to which extent upload/download/kernels are running parallel (as discussed in the meeting on 25. May).

lopez-c commented 6 years ago

Hi, This is an example of what I tried to describe during the meeting. The GPU application has been modified not to transfer pedestal data from one device/stream to another (since we assume that one detector is always connected to the same node)

As you can see, upload/kernel/download parallelism works fine for the first execution of both streams, but the second upload of data of stream 14 starts right after download of stream 15 and not after the download of stream 14. This is program in the GPU application this way, but it would be great to remove this dependency, since the overall throughput of the application could be increased significantly

test_stream

Cheers, Carlos

kloppstock commented 6 years ago

Hello Carlos,

I have looked into this a little bit. It is not possible to say exactly where this came from. Depending on which version you used, it is possible that some images are being stored in the main routine. For this we synchronized the streams and blocked further uploads. This was just for debugging and has since been removed. If you could send us your modified version, we could look further into it.

Cheers, Jonas

lopez-c commented 6 years ago

Hi Jonas, Yes, we have identified the point that delayed the execution next execution of the stream, and we can now overlap up and down transfers.

Cheers, Carlos