Choose timezone
Your profile timezone:
BLonD code development meeting
→
Europe/Zurich
864/1-B04 (CERN)
Tour-de table introduction, welcoming Luis
Markus: OTFB still an issue with the gain margin (follow-up: Helga, Markus)
General practise reminders (Helga):
- How to build docs
- How to check your code modifications
- Hard internal deadline to submit your contribution to the joint publication: 31st Jan. 2019
- Recent merge requests approved, to be merged (Kostis, Helga)
MPI computing (Kostis)
- Distributed BLonD: optimised and parallelised within a single CPU
- MPI for more complex simulations
- Master-worker tasks with hybrid execution model (MPI tasks on the high level with OpenMP threads underneath)
- Operations can be of type (i) parallel, (ii) serial, (iii) communication
- Optimisations done:
- communication minimised: sending bytes, profile as 16-bit int, commands and data sent in one transaction
- minimise serial parts: global profile is a sum of workers’ partial profiles, packed FFTs, remaining serial regions parallelised in C
- minimise synchronisation: more load to tasks that run with the master
- MPI can only be run on HPC batch cluster; resources dedicated to BE department. Get access by submitting a ticket to the service desk. Capacity: 300 CPUs, 20 physical cores each. Short (~1 day queue), long queue (unlimited)
- LHC ramp with 96 bunches: 3 days vs 350 days
- SPS test case with 72 bunches: 20 minutes vs 2 days, 2 FFTs per turn -> more serial region
- PS test case with 21 bunches, multi-turn wake field: 1 hour vs 3.5 days
- Approximate computation: reduce serial part by trading off accuracy for improved performance
- trick 1: calculate profile only every x turns
- trick 2: assume local profile is proportional to the global profile
- Approximate computation is case dependent and should be evaluated! Process monitoring the error
- Error estimates underway
There are minutes attached to this event.
Show them.