Benedikt Hegner showed 6 slides.
Frank Wuerthwein: Go where the money is. KNL will be relevant in the US. For CMS Reconstruction will be the most cpu intensive data processing activity, and of that Tracking is the biggest part
Marco Cattaneo: C++14 is coming, we all need to become familiar and develop expertise, need for common support
Michel Jouvin : HSF TN’s were started in order to address this, so that expertise can be shared
Mattias Wadenstein : easier to deal with small projects; validation - use state of the art compilers, haven’t done anything yet on behalf of the entire community
Fons Rademakers :
- Need to keep tracking new (software) technologies such as Apple/Swift
- Fast memories will have an impact - build a demonstrator to study them
- ROOT format is used for storage model but this may be outdated - need to look at new technologies such as Kudu, C++ Object Storage
- Interesting results coming out of taking a fresh look - Intel competition on Xenon Phi - student managed to increase the speed of neurobiology code by a factor of x320
- Make HSF a legal entity (cf Apache Foundation), handle Copyright in a better way, helps to get funding from industry, profit from work done in Machine Learning
Benedikt Hegner :
- several comments to make..
- legal entity : this is a chicken/egg situation, HSF started small by a few volunteers, need to reach a critical mass and demonstrate feasibility of initiative, now need more participation from the community
- Concurrency Forum fosters 1st round of demonstrators, now need to start a second round e.g. fast memories
- Different languages have been looked at e.g. Go, but no real outcome, take a look at Julia, ….
Ian Bird :
- profiling tools were mentioned, it would be good to have measures of what can be achieved.
- savings of 10% here and there are crucial since they have a big impact on costs. If they exist they should be used
- I’d like to see some metrics of what can be done
Danilo Piparo : we need to build a matrix of costs, for example see what we are prepared to give up; one example is precision, floating point give up IEEE performance
Liz Sexton-Kennedy : we have one person to do it for the experiment, this is a fundable model, have developer and user work closely together to achieve results
Latchezar Betev : for simulation code is focused in one place, therefore optimisation can be done for whole community, surprised more emphasis has not been given to this in the discussion
Marco Cattaneo :
- careers are important, need to raise the profile of developers
- not a part-time job for a student, engineers are important
- middleware is not a job for a physicist, get this on the radar of management
Stefan Roiser : try to involve SE institutes - those close to our physics depts. and collaborate with computer scientists; in LHCb we need to change our framework and in particular the data model in order to exploit vectorisation for example, currently we are not using SIMD
Ian Bird: there are ways of giving recognition (see mail thread that was circulating earlier this year; idea is to look at code that is accessed/used in git repositories such that articles can reference what has been used in a piece of work, this allows developers to build their CVs
- I agree its important to work with computer scientists, however the biggest role is still played by physicists;
- have to understand the impact of any software change on physics results
Benedikt Hegner : yes we need physicists for validating impact on physics results, but this shouldn’t prevent us for being outward-looking and learning from other fields, there is a place for non-physicists and this will help us to avoid re-inventing the wheel
Miron Livny : From experience, solutions owned by experiment(s) don’t become common, only solutions that are owned outside the experiments have a good chance to become common solutions.
Liz Sexton-Kennedy : Disagrees. Frontier and xrootd were solutions that grew out of a single experiment (CDF, BaBar). The solution has to be proven to work at scale for somebody before it is accepted by another. Common solutions need to be open source because it increases the trust between developers from different organizations.
- Here in Netherlands we benefit from participation of non-physicists
- However experience shows it is hard to rethink an algorithm if you are not a physicist
- Try community building i.e. one good place where people can talk to each other
Oxana : we need to lower the threashold for invoving non-physicists who would like to help
Frank Wuerthwein : If I look in CMS, the biggest contribution to tracking came from 5 people, 3 of whom have permanent jobs; I’d suggest that physicists have a harder time becoming permanent
Benedikt Hegner :
- HSF is everybody, you are all invited to contribute, need to assemble people who are keen to do technology tracking
- In reply to question from John, the HSF Workshop will take place at LAL, Orsay, Paris and the current favoured time is first week of May, time hopefully to be fixed early next week
note taker : John Harvey