BLonD code development meeting

Name: BLonD code development meeting
Start: 2020-08-28T10:30:00+02:00
End: 2020-08-28T11:30:00+02:00
Location: CERN

Friday 28 Aug 2020, 10:30 → 11:30 Europe/Zurich

CERN

Video conferencing only

Hide

operations profiled: computation, communication, and serial
data parallel version: all workers do same serial and computation parts
- overhead of serial part done by all workers
- drawback of waiting for synchronisation
improved version: task parallelism
- different workers can do different serial parts
- GPU serial operations are faster and can share results with CPU
- minimise the sync time
load balancing across CPU and GPU
LHC results: up to 5x speed up with 2 GPUs (no parallelism) compared to task-parallel CPU version
- overhead in communication when using GPU+CPU mixed mode (but this is due to initial turns being more comp. heavy with load balancing)
- similar results for PS
- LHC case with 2 nodes runs faster than with 4 or 1 (not the case in PS)

Kostis: Approximate computing

good results for PS, SPS, and LHC in accuracy
PSB (Simon): should test unstable beam case. Kostis: could switch on/off the approximation, but this could impact results
a tool to be used with care!! cannot be applied in all cases.
- e.g. for coarse scans or long simulations
could be combined with e.g. float32 approximation (never done before)

Alex: RF parameters pseudocode -- what would be the best way to divide RFStation: by harmonic, or system?

Option 1: input every system, combine harmonics, then add all systems or harmonics to RFStation
- System = feedback+cavity+amplifier, feedback an option for tracking
Finemet: can combine different harmonics?
- combine vector sum is a special task, not really tracking
  - Toy model for next meeting concerning Finemet

There are minutes attached to this event. Show them.