Version 5 (modified by mcastril, 9 months ago) (diff)

HPC-10_mcastril_HPDAonline DiagGPU

Last edition: 06/11/20 09:35:58 by mcastril

The PI is responsible to closely follow the progress of the action, and especially to contact NEMO project manager if the delay on preview (or review) are longer than the 2 weeks expected.

  1. Summary
  2. Preview
  3. Tests
  4. Review


Action HPC-10_mcastril_HPDAonline DiagGPU
PI(S) Miguel Castrillo
Digest High Performance GPU Diagnostics Online - 2nd Phase. After having successfully ported the dia_hsb diagnostic into a toy model, achieving 50x speedup, this task will focus on implementing the rest of the diagnostics and improving the data transfer between CPU and GPU.
Dependencies HPC-04_MCastrillo_HPDAonlineDiagGPU (completed)
Branch source:/NEMO/branches/{YEAR}/dev_r{REV}_{ACTION_NAME}
Previewer(s) Italo Epicoco
Reviewer(s) Italo Epicoco
Ticket #2368


High performance data analytics solutions aiming at tackling the online diagnostics of the NEMO model will be explored as complementary components in the model diagnostics software eco-system. Online techniques leveraging fast (low latency and real-time) data analytics approaches (e.g. on fat nodes) will be evaluated in real cluster environments. In particular, an interface of NEMO to the High Performance Data Analitics (HPDA) framework will be designed and implemented for online diagnostics.

The rationale of this activity is to improve the NEMO computational performance by executing the computations for diagnostics on GPU.


As first step, the portability of NEMO diagnostic calculations to GPUs has been analyzed, exploring how to adapt these regions from the current MPI implementation to the CUDA paradigm. A toy model has been created to perform preliminary tests, that were done using the dia_hsb diagnostic. The code itself is executed 50x faster than in a single CPU but the data transfer to and from GPU is the main bottleneck.

The plan for the next months is to work on removing the CPU/GPU communications from the critical path, by using an asynchronous method or parallelizing them with a computation phase. At the same point the efficiency of the overall solution will be improved, by mitigating the impact of the offloaded data.

The last step will be to extend this approach to the rest of the diagnostics.

Documentation updates