Choose timezone
Your profile timezone:
(Bi-)Weekly meeting green compute team
→
Europe/London
, , , ,
Description
Live notes: see Teams channel / on demand for non-UofM.
Zoom link:
https://cern.zoom.us/j/69108649411?pwd=BhqU0RERtnPf2gtK872m4gSM6izuZx.1
In this meeting, we will discuss status of the joint work with Glasgow, and connections to the HEPScore/HEPBenchmarks work.
Participants: CD (UofM), Sakshi Kumar (GSoC/UofM), Robin (from HEPScore CERN - GPU benchmarking and efficiency studies), Rosie (new summer student at UofM), Sam Skipsey (Glasgow), Alessandra Forti (UofM), Albert (Bruno) Borbely (from Glasgow), Emanuele Simili (Glasgow)
Sakshi, Emanuele, Michael - exporter for power accounting
Work done so far:
_exporter_
- Unified node_get [condorInfo, IPMI, PSInfo] to have only one script to run to get all the information.
- Streamlined some of the code
So far, no access to running jobs from HTCondor, so mocking some outputs.
Emanuele will test this on his machine, and provide more input for Sakshi.
_plotting_
Not a lot of data, very spiky - need to backfill the database, but Emanuele thinks it’s a good start.
Added also some VO stats (e.g. outliers in variance, could have done more ML but small dataset).
Question to Sam, Emanuele and Alessandra: what would be useful to look at in terms of outliers?
Alessandra: Could be compared in terms of accounting system from WLCG, since this is calculated from the batch system. There may be some differences across sites as well. The sum of these should give what we then give to WLCG as our official CPU numbers.
Bruno: downwards fluctuations would be interesting as this would mark a site not working. If I see that there are no jobs then things may be broken. It would also be interesting to check when the VO is pushing more jobs (e.g. reconstruction campaigns).
Next steps:
- Doing scripts to scrape laptop data, plotting that
- Generate other datasets and do ML on it, and be ready when we scale it up
Rosie Schiffmann - user-driven green compute
So far, work on collecting links for green software tools for HEP Software Foundation website. Prototype webpage available.
Target audience: users who want to measure the energy consumption of their own software package or ML. Sort of a living review. Note that ‘sustainability’ in time not clear - who maintains a living review?
Links to Green MetaData (work with Sakshi and Michael): how did you calculate the carbon cost of your paper? Metadata about the various pieces, and one can go to Rosie’s page.
Next: use some of these tools trying to evaluate the energy consumption of ML using online calculators and embedded package. Also tidying up Luis Villar’s code and plots on Noether.
What to do next in collaboration with Sam (following Julia studies)?
CodeCarbon with non-Python isn’t great (lots of scripts) → read from RAPL (in Julia) and then estimate RAM on top of that. Essentially a reimplementation of CodeCarbon in Julia, do the power measurement in a separate thread.
- Sam is happy to share the GitHub and talking through the repository with Rosie. Meeting to be organised on mattermost.
- This will become another example code for others wanting to use Julia
AOB
Emanuele + Michael chat about Prometheus —> will follow up offline (also need to discuss filtering).
Caterina’s AOB:
- WLCG Environmental Sustainability Forum kicking off after the summer, see the WLCG December sustainability workshop summary document for context: https://indico.cern.ch/event/1450885/attachments/3003697/5293988/WLCG%20Sustainability%20Summary.pdf
- WLCG heterogeneous architectures workshop in the works as well
Robin and Bruno are both working on heterogeneous workloads and would like to sync up:
- Pepper, MadGraph MC generator
- MadGraph
- Celeritas and AdePT, looking at how power/energy are being used
- CD mention also GAPS: https://arxiv.org/abs/2403.08692v1
There are minutes attached to this event.
Show them.
