Choose timezone
Your profile timezone:
Powered by Indico
v3.3.5-pre
Graeme Stewart is inviting you to a scheduled Zoom meeting.
Topic: EP R&D Software Meeting
Time: This is a recurring meeting Meet anytime
Join Zoom Meeting
https://cern.zoom.us/j/91901221096?pwd=SERFUkkzMDRwdkNvdXp6dWthK3Q0UT09
Meeting ID: 919 0122 1096
Passcode: rd7wp
Find your local number: https://cern.zoom.us/u/ae7prgeVJ
Join by SIP
91901221096@188.184.110.70
91901221096@188.184.85.92
Join by H.323
188.184.110.70
188.184.85.92
Meeting ID: 919 0122 1096
Passcode: 844111
It has taken a long time to get these machines into service and all of the delays are not very well understood. Is the IT installation service overloaded?
Quantum computing lectures are happening on Friday mornings - still unclear how HEP will use this at the moment, but an interesting subject. (First lecture was very informative, but rather long.)
Acts data model is for efficient transient processing.
As Acts is not a full reco framework, how do you solve issues to do with plugins? A. Acts doesn’t use dynamically loadable plugins. Define what you need at compile time instead.
API can change across major versions, but within a major version it won’t change. Semantic versioning.
If an experiment needs something special to act on their source data, then the source_link_t can hold this and it will be available to the correct code, but at compile time. Did not want plugable system with overheads of virtual interfaces that affect performance.
Making the GPU code port generic, for multiple experiments, is hard.
Pere - for the headers and public API, do you separate the public headers from the implementation? A. Yes (also in differnent namespaces). Internal headers are available (needed by templates), and can be used by mischiveous clients, but then people have to use the clearly marked internal namespace so they know they are doing wrong…
Pere - can the installation be shared? A. Yes, e.g., ATLAS builds Acts as an external, but then can be configured via headers. There are difficulties, e.g., needed to override the logger setup.
Erica - why were triangles chosen on the GPU rather than geometry surfaces? A. Each surface needs its own implementaion, which makes the code branchy. However, triangles are all the same, so the code becomes simpler and more prallelisable, even if some geomentric fidelity is lost. The current implementation is slow as it has no heirarchy. BVH!
Erica - validation, how do you match simulated particles to tracks. A. this is there, but not quite end to end yet. There is some R&D to have this and to use, e.g., autodiff to implement easier checks. Validation can always be improved.
Graeme - what’s the most important thing that EP R&D could contribute? A. Looking at event data models that could work on GPUs as well. And full chain validation. Some smaller R&Ds in the pipeline too, like a graph network navigator.