[CP2K-user] Benchmark results with specified iterations

Travis polla... at gmail.com
Sun Feb 2 17:54:09 UTC 2020


You can delay collection of statistics during an MD run for a specified 
amount of time,


This includes time per MD step, not just energies and so on.


On Sunday, February 2, 2020 at 11:35:36 AM UTC-5, vitesse lin wrote:
> Hi developers,
> CP2K may need extra cost in the first launch kernel when using GPU. It 
> needs to load related code into GPU memory. The cost should depends on the 
> size of kernel object code.
> This behavior will not need to do again in the process life cycle if GPU 
> memory is enough.
> Is there any way to by pass the first launch time when we are doing the 
> benchmark testing? In this way, we can observe the real cost of total 
> kernel cost.
> For an example, running a benchmark test with specified iterations in one 
> CP2K process lifecycle and having a result of average without the first 
> time.
> Thanks,
> Vitesse.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20200202/da25f377/attachment.htm>

More information about the CP2K-user mailing list