coveralls
commented
6 months ago Pull Request Test Coverage Report for Build 9520264580
Details
- 25 of 26 (96.15%) changed or added relevant lines in 3 files are covered.
- No unchanged relevant lines lost coverage.
- Overall coverage increased (+0.4%) to 97.838%
| Changes Missing Coverage | Covered Lines | Changed/Added Lines | % | ||
|---|---|---|---|---|---|
| nautilus/pool.py | 17 | 18 | 94.44% | ||
| <!-- | Total: | 25 | 26 | 96.15% | --> |
| Totals | |
|---|---|
| Change from base Build 9114932167: | 0.4% |
| Covered Lines: | 1222 |
| Relevant Lines: | 1249 |
johannesulf
marberi
Basic support for paralleling Nautilus on a Dask cluster. Out high-throughput data center is not optimized or well configured for running MPI jobs. We were discussing running cosmological inference using a code, which apparently relies on Nautilus. Wanting to see if we could avoid MPI, I tested parallelising the basic example in the Nautilus documentation.
After creating a Dask cluster: from dask.distributed import LocalCluster cluster = LocalCluster() client = cluster.get_client()
the interface: sampler = Sampler(prior, likelihood, pass_dict=False, pool=client)
remains unchanged. While the Dask client supports map, the syntax the Dask versions is asynchronous. To avoid this detail being exposed too many locations, the code use wrapper to mimic the same behaviour as the multiprocessing pools.
Testing, all the stages could be parallelized over many machines (not using the local cluster above, but through the dask-jobqueue package).