Test performance of EBS volumes as home directory storage

[sgibson91]

Context

The main question we want to answer is:

Is the default of 3000 IO/s and 125MB/s we get (with gp3 volumes) enough to provide users with their expected performance?

For comparison, EFS provides 90,000-250,000 IO/s and 10-60 GB/s (read) 1-5 GB/s (write) throughput (technically) so we will be almost an order of magnitude down using EBS and should understand how this will affect users.

Relevant links

• https://aws.amazon.com/ebs/pricing/ (has performace and iops numbers for EBS)
• https://docs.aws.amazon.com/efs/latest/ug/performance.html#performance-overview (has numbers for EFS)
• https://fio.readthedocs.io/en/latest/fio_doc.html (alternative tooling)

Task list

Tasks

Beta Give feedback

1. Decide what this 'standard activity' is: An example is "time taken for a git clone https://github.com/torvalds/linux to complete". We should put the repo in a local path in a container and clone from there to isolate confounding factors such as a network.
2. Design and implement infrastructure that allows us to to provide something to test the standard activity on and scale it to replicate many users
3. Record measurements on the NFS server (EBS)
4. Do 'standard activity' with increasing number of parallel clients, and record both how long the activity takes, and measurements on the server

Definition of Done

• Simulate N number of users performing some standard activity on home directories to understand at what point the standard activity becomes unbearably slow.