Agenda
https://indico.cern.ch/event/735132/
Profilers
- Andrea - main goal is to correlate with a line of code. There is an overlap with
deployment and infrastructure monitoring. How well does an
application work on different architectures? Performance and Cost
Model Group (in WLCG) looking at this.
- Hooks to understand where certain parts of an application start/end
are useful. E.g., identify the start of the event loop.
- Would like to be able to profile (if needed) on the grid. Debug
end to end performance, not just on development machine.
- Is the idea of a warehouse useful? Could this be a service from CERN
IT?
- It’s useful. Needs to be a concrete proposal.
- How do you want to analyse the data? This informs what we want to
store. Maybe not so much need for long term archiving of the detailed data.
- Need a lot of metadata to link it back to the right code
versionin git.
- Software in production becomes of historical interest; at least
some summary data is important.
- The bleeding edge has a useful lifetime of only a few days/weeks
during the development cycles. A lot of churn on HEAD and code
often broken or unstable.
- Some vision of how to come up with common formats, but needs some
expert input to turn this into a DB schema that would scale.
- Is it really worthwhile to compare between different tools - they measure
different things (explains the zoo of the tools - they do
different things).
- There have been some useful examples of cross profiling in ALICE.
- Also now much easier to analyse large profile dumps.
- Keep more ‘raw’ data and analyse on the fly.
- Is it possible extract data from VTune? Probably not - proprietary.
- Comparison between different developers’ work could also be useful as
a collaborative tool.
- e.g., ATLAS SPOT monitoring is just a ‘red flag’ style - something more generic would help and share efforts between
experiments.
- Progress by gathering people who could work on a prototype of the
web based presentation. Concentrate on one or two profilers first.
- Bootstrap guide for different tools would be useful.
- There is some training that people have developed and we should
inventory that and point to it.
- Getting users started is a very good thing to contribute to.
- Peter - as a developer got some very incomprehensible information
back to try and understand a difficult problem. Developers are
very busy and under a lot of pressure, not profiling experts.
- Need to optimise the efficiency of our coders - what we want to
develop has to help that.
- CMS igprof experience was that providing a service to
developers did help a lot.
Static Analysers
Presentation held back until the next meeting.