-
@Conengmo what do you think? Take a look at https://closember.org/ and let me know.
I believe there are some stale PRs, already resolved or no longer relevant issues that could use a clean-up!
-
This notebook assumes that we are using an older style psf (for example, `psf.fits` which is included in this repo) and does not fully support using a newer psf (`/global/cscratch1/sd/stephey/desitest…
-
Understand how we can make sure cpu and gpu work are happening asynchronously.
For example, can we do the I/O on the cpu side while simultaneously doing extractions on the GPU side?
Is mpi a go…
-
During the hackathon we saw speedups from using NVIDIA's MPS framework. We also saw some overhead we didn't fully understand.
Explore
* The best configuration of MPI + MPS ranks per frame (and p…
-
Once the cpu matrix-based version of gpu_specter is ready (Issues #4, #5, #6, and #7 have been addressed), it should be fully benchmarked against the current master version of desispec and specter.
…
-
When the gpu version of gpu_specter is fully ready, profile the code to understand where our major gpu bottlenecks are.
At the moment our best tools are nsight systems and nvvp (for multiple mpi r…
-
Once we have settled on a correct version of the cpu gpu_specter (i.e. after Issues #4, #5, and #6 are complete), we will need some unit tests to make sure that we maintain this correctness.
-
It has happened several times already (gh-7606, gh-8870) that we had to fix bugs in numpy itself due to a missing call to ``customize`` after creating a new compiler instance. External users are proba…
-
Our version of gpu_specter (even in it's cpu form) should have the same connectivity as the current cpu versions of desisim and specter and be fully interchangeable with the cpu version.
This mean…
-
I have a script which uses testing.postgres to create a temporary database to test against.
My script has a SIGINT handler which handles teardown logic, which I want to run before the database is ter…