Scaling Run Sync Up

• Jonas' Proxy
• AthenaPK
• RIOT

Work in Progress

• Jonas' is validating Darwin Power9 builds while we wait for him to get RZAnsel access

Target Problems

Proposal

1. Pick a set of problems for each application.
2. Start with minimum resources on each architecture (Broadwell, V100s) to run that problem.
3. Strong scale.

General

Recommend: Set PARTHENON_DISABLE_HDF5=ON.

AthenaPK

We will reproduce an existing problem Galen has already run.

Previous discussion: https://github.com/lanl/parthenon/issues/301

RIOT

Problem: Multi-material mixing problem. Kelvin-Helmholtz

American Flag simulation.

Variables:

• AMR vs. uni-grid (same level of max refinement)
• 2 vs. 4 cell ghost zones

We'd like to do a comparison of two different ghost zone sizes 2 vs. 4 for integration (cuts communications with bigger ghost zones). We need to think about this affects problem scaling since ghost zones can have an outsize effect on memory usage.

Proxy App

Problem:

• ???

Target:

• We need to scale the computational density of FillDerived. Probably just a factor multiple on the number of iterations in FillDerived. Will try to compare to RIOT.

Variables:

• Load Imbalance - target 3 different scenarios along axis from very homogeneous to a handful of extreme outliers