• gh repo create --source=. combines three steps: creates the remote repo, adds it as origin, and pushes - much faster than doing each manually
• Nested git repos are a common gotcha - git treats them as submodules by default, which can cause confusion. Flattening (removing inner .git) is simplest when the nested repo doesn’t need independent versioning

Gavel vs FGD: These papers solve complementary problems:

• FGD (the paper we just read): Handles partial GPU allocation (0.3 GPU, 0.7 GPU) and the fragmentation it causes
• Gavel: Handles heterogeneous GPU types (V100 vs K80) and how to fairly/efficiently allocate across them

In a real cluster with both problems, you’d want elements of both solutions.

Effective throughput is the key abstraction. It collapses the complexity of heterogeneous allocations into a single number per job. This lets Gavel reuse the structure of existing policies (fairness = max-min, makespan = min-max duration) while making them heterogeneity-aware. The insight is that most policies are already functions of throughput—Gavel just generalizes what “throughput” means.