Node specs @ Piz Daint
- up to 150 nodes (shared with ATLAS and CSCS users, for the Tier-0)
- dual socket Xeon E5-2695 v4 @ 2.10GHz
- 68 schedulable cores (HT enabled), for a total of ~10'000 cores
- 128 GB RAM/node (no memory limits per job, small swap is available but not recommended), which means ~2 GB RAM/core
- Scratch shared with Tier-2 (has been recently reinforced with an SSD layer) of 700 TB
Middleware @ Piz Daint (dedicated to Tier-0)
- 4 ARC servers (2 for submission, all 4 for data staging)
- Queues not published on BDII
- Accounting not pushed to APEL
CMS proposed Reconstruction workload
- 8 thread processes (16 GB RAM)
- For 1000 cores, input cache needed is ~20 TB big (~1 week buffer) and read @ 500 MB/s
- Mostly pure streaming (push data to Scratch, process, send 50% of the data back to other sites)
- 8-12 hour jobs
- CentOS & Singularity needed
Integration @ CSCS
- ATLAS is few (2-3 days per week) high-priority workloads
- CMS is background tasks that can take 24x7 if possible
- Up to 150 nodes are available (can be less, depending on the load)
- CMS can potentially use all the nodes, and is scheduled out when ATLAS workloads show up
- All managed by the scheduler
