Compare performance/features of different Rucio RSE access modes

Agreement:

• test workload: unsure

TODO: are there example "data mesh" workflows we can test?

The deadlines for ensuring budget is in place for v0.2 expansion for SRCNet are rapidly approaching (at least they are in the UK, exact dates still TBC). Within the UK we would like to make use of various different supercomputing resources within SRCNet, but as yet, its unclear on how to deal with the connecting to SRCNet controlled storage.

Given a large proportion of the early SRCNet compute recourses are likely to be spend on SDP pipeline development, we will focus on imaging pipelines for these tests, likely looking at existing LOFAR pipelines, assuming no SKA specific pipelines are available. (TODO - what other test workloads are a better fit here?)

This work is to evaluate various possible modes of operation including:

• Benchmark accessing data both within RAL, and remotely to Cambridge using using xroot (and stretch compare to S3) vs a local high speed scratch storage system within the HPC system (at Cambridge)
• POC and document the practicalities of adding an isolated high speed scratch within an HPC centre as a Rucio RSE, and copying data in/out of the HPC centre as required using FTS, Rucio CLI, or similar (i.e. buffering like SDP architecture describes, with copy in buffer, in use buffer and a copy out buffer).

AC1: Document the performance of xrootd vs S3 APIs between RAL and Cambridge in the UK, compared with using a high speed local scratch filesystem.

AC2: Document findings from POC of copying between RAL cloud and Cambridge HPC clusters using Rucio/FTS