March 2021 - Ecp-aurora-sdk-discuss - mailman3.cels.anl.gov

GPU Debug Support with oneapi
by Muralidharan, Servesh 15 Mar '21

15 Mar '21

Hello folks, The queue “iris_debug” has the necessary updates to the kernel config to enable GPU debugging. This is currently in beta testing phase so feel free to test it and give us an update should you run into any particular issue. After a week or two of testing, we will migrate the rest of the nodes in “iris” queue to have this support as well. Loading the latest oneapi module should give you `gdb-oneapi` in your environment that can be used for this purpose. On behalf of early hardware team. -- Servesh

1 0

ATTN: ECP Users with access to Arcticus Queue
by Muralidharan, Servesh 15 Mar '21

15 Mar '21

Hello all, We have added several new Arcticus nodes, each with a single ATS GPU. These SKUs are still early A1 silicon samples but are running at a much higher frequency in comparison to the older cards. These nodes are particularly enticing as they are higher performance SKUs and with many more of them can greatly aid in developing/testing your code as we move towards Aurora. For those who don’t have access, but would like to request access, please follow the steps here: https://confluence.exascaleproject.org/display/HIAI01/Aurora+Early+Access+-… These are not production SKUs and still have the some hardware issues as with the early silicon on the previous cards, so expect them to behave as such. However, there have been numerous updates to the firmware, drivers, OS image, etc., rolled into this migration. That should help in performance/stability improvements. This migration also introduces validation checks to ensure the GPU is behaving normally on an user job termination and if there is an unexpected behavior the nodes are marked down in the queue. An admin intervenes periodically to ensure these nodes are good before placing them back for user access. Should you have a workshop or other events planned, please send a reservation request ahead of time so admins are aware of the priority. We will add the details to the config changes of the node itself in the coming days to this page, https://wiki.jlse.anl.gov/display/inteldga/Arcticus Please update your environment to use the latest oneapi module, `2020.12.15.005`. Anything older might not be fully supported with this update. On behalf of the early hardware team. -- Servesh

1 0

Publishing NDA work
by Knight, Christopher J. 01 Mar '21

01 Mar '21

Hi all, With the ECP Annual Meeting around the corner, we wanted to send a quick reminder that presentations involving NDA material need to acquire the proper approval. A summary of the process is here on JLSE’s confluence page: https://wiki.jlse.anl.gov/display/inteldga/Publishing+NDA+Work . Some additional comments are below. Please review the process and check if your work involving Intel hardware & software needs to be reviewed. As the ECP deadline for uploading presentation slides is April 5th, folks should plan on submitting their work for review no later than March 22nd. Earlier submission for review is encouraged. chris ===== * Approval is not needed for papers that use production (not beta) developer tools on publicly released hardware * For papers that mention the use of the non-public ATS devices (or other future unreleased hardware) * Please submit specific claims including full prose and data planned to be disclosed. Worst case, two weeks is needed to respond so the sooner the better for submissions. * Mentions of functionality progress (e.g. % of application modules completed) are likely to be approved. * Given the low performance of the current ATS sample hardware, we believe it’s very unlikely that performance data would be approved for publication. * Please include all specific claims including full prose and data planned to be disclosed. * Example: “We now have 75% of our application running with no high or critical bugs on ATS samples in preparation for Aurora."

1 0