Ppl-accel mailing list

["Choi, Jaemin" <jchoi157 AT illinois.edu>]

1. GPU PEs (drones)

- Tested concept with 2 benchmarks, busywait and stencil2d, and in both cases it seems to perform similarly to using LIFO for the callback messages

- No performance improvement seen with stencil2d, since GPU work finishes much earlier than CPU work per iteration

- See screenshots on TMS

- Need to test the case when the same task runs slower on the GPU (something other than stencil2d)

2. Tracing API

- Exchanged emails with NVIDIA developers

- Was recommended using CUPTI Activity API, with Callback API as needed

- Tested with activity_trace_async sample application to see if measured timings were accurate, and they seem to be so

- Need to integrate into HAPI

3. Preparing for 6.9.0 (Remaining Redmine Issues)

- Update documentation: maybe section title should be changed to 'Hybrid API' instead of 'GPU Manager Library', as Hybrid API is all that is exposed to the user

- Add CUDA as an option in the build script

- Delete GPU dummy mempool: should be easy to do

4. Others

- Talked to a researcher from AMD (Dr. Bradford Beckmann) during my poster presentation at SC, and he told me that their HIP library is able to interrupt the CPU when GPU work is complete. This could be an easy alternative to the solution in the poster, but possibly only with AMD GPUs.

- AMD seems to be doing some interesting work on GPGPU, check out their website if anyone is interested: https://gpuopen.com/professional-compute/

Jaemin Choi

Ph.D. Student in Computer Science

Research Assistant, Parallel Programming Laboratory

University of Illinois Urbana-Champaign