CVMFS Catchup

Friday 23 Aug 2024, 10:30 → 11:30 Europe/London

Teams

Adam Parker (High Performance Data Analytics), Shaun de Witt

Agenda and minutes

Previous actions: (new actions and updates in red)

#	Action	Owner	Status	Due
01	Enquire about Azure cloud resources for hosting CVMFS stratum-0 data server	SdW	Asked John Galley. awaiting response. Required resources: 2 cores, 4GB RAM, 500GB storage
02	Request VM on STFC IRIS for initial benchmarking tests	AP	Received the application form but not sent yet.	19/09/24
03	Demonstrate CVMFS client installation and data access on CUMULUS login nodes (unprivileged user account)	SD	Done	19/09/24
04	Demonstrate CVMFS client installation and data access on CUMULUS compute nodes (no internet access)	SD, AP	Blocked by 05 - requires squid proxy setup first	19/09/24
05	Put in ticket for VM on the cumulus noad previously known as login-03. Find out suitable resources requirements for request beforehand.	SD		19/09/24
06	Find out how versioning works on CVMFS	SD		19/09/24
07	Demonstrate CVMFS client installation and data access on CSD3 compute nodes ( internet access is available)	SD, AP		19/09/24

 

Agenda: (notes from meeting in red)

• Status of external server requests for initial benchmarking tests (Azure and IRIS)
  Note: external VM farm will be available at UKAEA in near future, but probably not available for use for another ~3 months. 
• Progress updates
• (MAIN EVENT) Discuss architecture for offline compute nodes 
  Target squid proxy for cumulus and use direct access to stratum-0 on CSD3. One cumulus node has been converted from login to a VM server, SD to raise ticket for a squid VM for CVMFS.
• Next steps and assign actions
  Initial discussions on benchmarking and testing aims:
  
  • compare data access performance between uda and CVMFS on CSD3
  • compare data access performance between Sam's mastapp (zarr files on S3), using s5cmd, and CVMFS 
  • depending on data transfer performance results, can consider using cvmfs for metadata only and using s5cmd instead for data access.

1 Review of actions and summary of recent progress

See action list in event minutes section

Speakers: Adam Parker (High Performance Data Analytics), Shaun de Witt, Stephen Dixon (The UDA man)

2 HPC client architecture and strategy discussion

General options for CVMFS client access on offline compute nodes are:
- Squid proxy on dedicated VM on CUMULUS network. This accepts http requests from compute nodes and forwards to stratum-1 CVMFS server over internet.
- Local stratum-1 server on CUMULUS network. Similar requirements to squid proxy, but maintains full copy of data instead of caching some recently-used subset.
- pre-fetch and cache all required data on shared storage on CUMULUS (through login nodes?) before running compute jobs

Aim of discussion to decide which configuration we wish to target.

Speakers: Adam Parker (High Performance Data Analytics), Shaun de Witt, Stephen Dixon (The UDA man)

3 Benchmark testing plan

What are we hoping to measure and what test cases can we think of

Speakers: Adam Parker (High Performance Data Analytics), Shaun de Witt, Stephen Dixon (The UDA man)

4 Next steps and summary of new actions

Aims for next meeting and assigning actions

Speakers: Adam Parker (High Performance Data Analytics), Shaun de Witt, Stephen Dixon (The UDA man)