HEP-Workloads mid-cycle Sprint meeting (replacing the regular HEPiX Benchmarking Working Group meeting)
Present Chris Hollowell, Michele Michelotto, Domenico Giordano, Lorenzo Rinaldi, Martina Javurkova, Andrea Sciaba
Documentation can be made available later because the tools should be rather simple and has good help.
HEPSCORE v2 is a new version where we will reorganize stuff in a way
But all the development will continue on version 1
The whole benchmark last for about 6 hours ( 3 CMS benchmarks , 2 Atlas, 1 LHCB) so 18 hours if we run it and take the median
See BMK-41 to see the effect of median on reducing the spread
We may want to run the default one (with the median of 3) or the fast one. Now Domenico will run on the machines where he already has many measurements with HS06, SPEC CPU2017 and others
Problems with DIGI-RECO. It will not fit in a machine with less than 2 GB/core
The script will recognize that situation and switch to RELAXED mode
Manfred propose to run in MIXED mode so that only 25% of the cores run DIGI-RECO with more that 2GB/core, 50% SIM at about 2GB/core and 25% run GEN with much less than 2GB/core. This is an interesting feature that can be addressed at the the level of orchestrator but it should be in the benchmark
The ATLAS jobs must report Wall Clock Time like CMS. Alice also has this problem, We must check LHCB
Manfred on the validation tests.
5 workload are ok
1 workload has minor problem and need improvement
2 have severe issues in stability
Michele Tested recent containers. They are ok
Licenses: we are waiting for input from the guy that went from IT to Technology Transfer. He will be present at the pre-GDB and Domenico will prepare a link a distributed documents where we will write down our request or requirements
Andrea able to build a PATATRACK run ( to take advantage of GPUs)
They are not in CVMFS this leads to an image that is 49 GB
The we go through the agenda of the next pre-GDB to make final adjustements