COSMA facilities

COSMA7
COSMA is comprised of:
COSMA also includes a number of other facilities including:
- the DINE BlueField cluster
- the Cosmological Database
- GPU systems
- Other facilities
Storage includes:
- /cosma5 - 2.4PB
- /cosma6 - 2.5PB (to be retired in Autumn 2023)
- /cosma7 - 3.1PB
- /cosma8 - 5.3PB
- /snap7 - 440TB for fast temporary storage (e.g. checkpointing)
- /snap8 - 1.2PB for fast temporary storage (e.g. checkpointing)
- /cosma/home - 37TB for user homespace (10GB quota)
- /cosma/local - 37TB for program/module storage
Users have access to /cosma/home/PROJECT/USERNAME, and one or more of /cosmaN/data/PROJECT/USERNAME.
Scratch space
COSMA7 and COSMA8 both contains the /snap[7,8]/ fast I/O storage space, which offers an achievable 180GB/s-400GB/s bandwidth. At installation in 2018, /snap7 this was the fastest storage in Europe.
To make the most of alloted time on COSMA, researchers often pause their jobs e.g. every 12-24 hours, and create checkpoints which capture all the data in memory. Using the checkpointing /snap storage, these checkpoints are many tens of times faster than without. For example, if a checkpoint for a 5 month simulation takes three hours on COSMA6, made every 12 hours, researchers will lose 1.8 million core hours (13% of available time) on a 4096 core job. On COSMA7, it will take only 10 minutes to create a checkpoint, and so only 0.2% of available time is lost.
COSMA6-8 network map (mixed InfiniBand fabric)
Resources
- COSMA network map (last modified: 6 May 2020)
- COSMA6-8 fabric (last modified: 15 November 2021)
