GPU PDHG part on separate instance

回答済み

Per Jernström

2025年10月24日 13:39

• 
• 

Hi,

Been testing beta 13 version with GPU support. In many cases, especially for large problem instances, the GPU part is relatively fast, but the following crossover and everything after that can take a long time. 
Hence a feature request, maybe initially for instant cloud. It would be nice to have two instances starting at the same time, one GPU and one without. Once the GPU part is complete the problem would continue on the non-GPU machine, saving valuable GPU-time.

I tried to run multiple problems on the same machine, staging them in time (next one starting as soon as GPU phase on the previous is complete). It works, but memory management was hard. GPU phase ending seems to create a spike in host machine RAM usage. Moving the problem to a separate machine would therefore be more stable.

関連

• PDHG
• GPU

0

• 

  David Torres Sanchez

  • Gurobi Staff

  2025年10月27日 10:22 編集済み

  Hi Per,
  
  Thanks for getting in touch and for testing the beta!
  
  This seems like quite a specific application. There would be a significant overhead between the machines (GPU → host → other machine). Have you tried benchmarking against defaults or barrier? 

  What you can also try is solve the problem with Crossover=0 Method=6 PDHGGPU=1 Threads=1 and save the solution to an .attr file. Once this is done you can kick off the next GPU job (with the same parameters), and a new CPU job that runs crossover on the solution (you can do this by reading the .attr file and Method=6 PDHGIterLimit=1 Threads=1 this will skip straight to crossover after presolve and use a single thread. You will ~~not~~ only be able to queue n-1 CPU-only jobs as well as a single GPU job.

  But my guess would be that this would slower than solving in parallel or even serial with defaults.

  Cheers,
  David

  0

サインインしてコメントを残してください。