[ODK] Meeting 2018 10 03

Minutes for the October 3, 2018 meeting

1. Previous work
   • Alexis: work on existing structure to compress data, used to send GMP integers via MPI. Even send matrices (and Integer matrices). The existing code seemed messy. This code has been cleaned and optimized. One issue remains: size of GMP elements: 32- vs. 64-bit. Should be developed for portability purposes. Furthermore, a specific test was written for serialization. This test does not uses MPI. Works on local workstation. Need to perform benchmarks to see if MPI communication is improved. Was designed to be modular.
   • Zhu: mpi-comm-interface: cleanup and testing in progress. Modular<int64> seems to be not working. Probably because 64 bit is not implemented.
   • ETA: Monday, October 8 for the serialization.
2. Next ideas
   • Refactor: solve, rational_solver.
   • Coverage: improve it, get a strong test suite.
   • Perfs in MPI: check if anything breaks.
3. ODK project review
   • From Feb. 2017 to Aug. 2018 which includes all deliverables done during this time period. Need to give status update on Work Package 5, the one discussed here. 3 reviewers, only one who will know exactly what HPC is. Last time, this reviewer was not fully convinced by what was presented (B. Hart's work on FLINT improvements on vectorisation). Especially, he was expecting some elements about clusters, GFlops and so on... And this time, they expect some results and will be especially focussed on that. Need to have some actual computation time/GFlops of a code running on a fairly large cluster (20 nodes, 20 cores/node).
   • Your mission, should you choose to accept it, will be to take Zhu's code on rational solver and make it run on a cluster. Run it on around 200 cores, and the goal is to create a fairly large instance, a few hours to a day of computation time. For instance a 10k x 10k system. Clusters available:
     • Luke
     • Dahu (roughly 20 nodes, 20 cores/node)
   • Goals:
     • It works!
     • It scales!
     • It has good performances! (GFlops/NBCores ~= theoretical GFlops per core)
   • Deadline:
     • Soft: October 26th, 2018.
     • Hard: October 28th, 2018.
   • Zhu already worked on Luke, using 2 cores. He will share his experience with Alexis.
   • Technical details: full MPI for starters. If it does not work, MPI for the nodes and OpenMP for the cores. Bottleneck: CRT reconstruction.
   • Estimation of computation time for 10k x 10k system: 39 min per core without reconstruction, and without comms, for a bitsize of 10 on 20 cores.
4. Other
   • Weekly meeting on the 31st of October is cancelled.
5. tl;dr
   • Topmost priority: Benchmark rational solver on a cluster