search

stanford-futuredata / gavel

Code for "Heterogenity-Aware Cluster Scheduling Policies for Deep Learning Workloads", which appeared at OSDI 2020
MIT License
125 stars 31 forks source link

faster problem compilation #237

Closed akshayka closed 3 years ago

akshayka commented 3 years ago

Modifies the construction of the CVXPY problem (get_allocation) in MaxMinFairnessPolicyWithPacking so that it compiles much faster.

I reran some of the experiments in scheduler/notebooks/figures/improved_scalability/max_min_fairness.ipynb. For the large problem (1024 jobs, with packing), the combined time spent in CVXPY + MOSEK is less than Gandiva.

With these changes (enabling multithreaded compilation in CVXPY, see https://github.com/cvxgrp/cvxpy/releases/tag/v1.1.11) image

Compared to without these changes (@ master): image

deepakn94 commented 3 years ago

This looks great! Are the allocations exactly the same as before?

akshayka commented 3 years ago

Yup!

Or, I should say, as long as CVXPY >= 1.1.11 is used. I found a bug in CVXPY when making this change, and I released a fix as part of 1.1.11 (see the release notes linked to above).

deepakn94 commented 3 years ago

Great! Thanks for the patch!