Continuous Integration in Devito

We use GitHub Actions for Continuous Integration.

Some of the workflows, in particular CI-core, which executes all of the core Devito tests, run on VMs that GitHub Actions provides for free to open source repositories.

Some other workflows run in the devito-cluster, which comprises nodes owned by Devito Codes Ltd, some of which donated by various companies.

The devito-cluster workflow matrix

node	CI-gpu	CI-mpi	asv	examples-MPI	docker-publish GPU
kimogila	x[OMP]		x		x
sarlacc	x[ACC]
nexu				x
bantha
rancor
macdevito

Nodes specification

macdevito

• MacBook (macOS Catalina)

kimogila [stable]

• Architecture: x86_64
• CPU(s): 8
• Thread(s) per core: 2
• Core(s) per socket: 4
• Socket(s): 1
• NUMA node(s): 1
• Model name: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz
• Total online memory: 31G
• GPU(s): NVidia RTX 3070

sarlacc [stable]

https://www.dell.com/support/home/en-uk/product-support/servicetag/0-R2FySTAydTZQU04ra0p5SmgwcE9MZz090/overview

• Architecture: x86_64
• CPU(s): 12
• Thread(s) per core: 2
• Core(s) per socket: 6
• Socket(s): 2
• Model name: Intel(R) Xeon(R) CPU E5-2630 0 @ 2.30GHz
• Total online memory: 48G
• GPU(s): NVidia RTX 3070

nexu [stable]

https://www.dell.com/support/home/en-uk/product-support/servicetag/0-cDlvYWZiZTZJd3Z1K05Wdys0dGpRQT090/overview

• Architecture: x86_64
• CPU(s): 12
• Thread(s) per core: 2
• Core(s) per socket: 6
• Socket(s): 1
• NUMA node(s): 1
• Model name: Intel(R) Xeon(R) CPU E5-2640 0 @ 2.50GHz
• Total online memory: 64G
• GPU(s): -

bantha [stable]

• Architecture: x86_64
• CPU(s): 12
• Thread(s) per core: 2
• Core(s) per socket: 6
• Socket(s): 1
• Model name: AMD Ryzen 5 2600
• Total online memory: 32GB
• GPU(s): NVidia RTX 3090

rancor [stable]

• Architecture: x86_64
• CPU(s): 16
• Thread(s) per core: 2
• Core(s) per socket: 8
• Socket(s): 1
• Model name: Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz
• Total online memory: 48G
• GPU(s): GeForce GTX 1660 Ti, GeForce RTX 2060

TODO

• Move examples-mpi from kimogila to nexu
  • Add test with larger MPI ranks (up to mpirun -n 8 ...)
• Set up timeout to kill builds after a short period of silence (What should the metric be?)
• Restrict builds on self-hosted runners to PRs (not all branches)? (TBD)
• Script to monitor host and device memory consumption and report it in the build output? (TBD, 3rd party package?)
  • htop (only for CPUs)
  • gpustat (only nvidia GPUs), a nice wrapper for nvidia-smi
  • radeontop (only for amd GPUs)
  • maybe we should develop our own layer on top of these, in lack of an alternative...
• Steal useful ideas for CI from other open source projects?
• Remove the now obsolete DEVITO_BACKEND env var from the workflow files
• Review and clean up various workflows. This includes updating out of date/obsolete actions.
• Move Docker GPU workflow to another machine
• Migrate CI-mpi to our own runners? (TBD)
• More GPU testing?
• Parallelize GPU tests (pytest-n <num_of_phys_vores> ...) 'cause adjoint tests are quite expensive
• Clean up install instructions openacc/openmp
• Setup organization-level self-hosted runners so that we can use the devito-cluster for both CI and TheMatrix. See here
• Action in the private_runners repo that "locks" a specific node for a given number of minutes standing on a wait(nminutes)
• Gerards list
  • MI50s setup - have cards, seeking server to host them
  • A100s setup - waiting on delivery
  • install cluster management and monitoring software on the devito-cluster
  • setup GitHub authentication
  • give George access to bantha
• Move jupyter workflow to the devito-cluster (think about all those non deterministic bugs we saw in the past, often due to an obscure interplay of caching, dask, mac OS, sympy...)