Repository for data and configuration for modelling scheduling on Cirrus CPU partition using ElastiSim.
This repository contains data for the January 2024 modelling period. There is a correpsonding repository for the September 2024 modelling period.
- algorithm/ - Subdirectory containing the simple, FIFO scheduling algorithm used by Elastisim simulations
- analysis/ - Scripts and analysed data from measured job load and simulations
- cirrus-cpu-measured/ - Measured job data from Cirrus
- cirrus-cpu-* - Various scheduler simulations (see below for details)
- emissions/ - Emissions modelling data and output
- utils/ - Utility scripts
| Directory | Description |
|---|---|
| cirrus-cpu-R/ | Cirrus jobs only, Cirrus jobs rigid |
| cirrus-cpu-M/ | Cirrus jobs only, Cirrus jobs moldable |
| cirrus-cpu-M_archer2-R-10pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 10% of Cirrus resource use |
| cirrus-cpu-M_archer2-R-20pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 20% of Cirrus resource use |
| cirrus-cpu-M_archer2-R-30pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 30% of Cirrus resource use |
| cirrus-cpu-M_archer2-R-40pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 40% of Cirrus resource use |
| cirrus-cpu-M_archer2-R-50pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 50% of Cirrus resource use |
| cirrus-cpu-M_archer2-R-60pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 60% of Cirrus resource use |
| cirrus-cpu-M_archer2-R-70pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 70% of Cirrus resource use |
| cirrus-cpu-M_archer2-R-80pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 80% of Cirrus resource use |
| cirrus-cpu-M_archer2-R-90pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 90% of Cirrus resource use |
| cirrus-cpu-M_archer2-R-100pc/ | Cirrus jobs moldable. ARCHER2 mobile jobs: 100% of Cirrus resource use |
## Model description
The Elastisim scheduler model consists of a number of different components:
- Scheduler algorithm definition
- Cirrus cluster definition
- Application performance model(s)
- Job list to simulate
The file [algorithm/algorithm.py] describes how jobs are scheduled. The algorithm is:
- Loop over jobs in the queue in order of submission:
- If a moldable job:
- Loop from smallest to largest job size and try to fit into available slots
- If it cannot be fit into available slots, move on to next queued job
-If a rigid job:
-Try to fit into available slots
- If it cannot be fit into available slots, move onto next queued job
### Cirrus cluster definition