Setup

Pageserver setup

• i3en.3xlarge
• The same as in 3aa4fc593e573f897bb8bb18590e5a186a2815a1
  • 40 GiB (!) page cache
  • 500k file descriptors
• (Disabled materialized page cache lookups to stress walredo in the benchmarks)
  • Baseline: problame/one-runtime--benchmark-2024-03-23--baseline
  • This PR: problame/one-runtime--benchmark-2024-03-23

Pagebench runs on an m7i.2xlarge (oversized) in the same AZ.

Pagebench Workloads:

• 128 clients against 1 tenant, fits into PageCache
• 3000 tenants, 1 client each, 10x larger amount of layers than page cache
• 20 tenants, 10 clients each, fits into Page cache

admin@neon-loadgenerator:~$
./pagebench  get-page-latest-lsn --mgmt-api-endpoint <http://srv:9898> --page-service-connstring postgresql://srv:64000 --limit-to-first-n-targets 1 --num-clients 128 --runtime 2m
./pagebench  get-page-latest-lsn --mgmt-api-endpoint <http://srv:9898> --page-service-connstring postgresql://srv:64000 --limit-to-first-n-targets 20 --num-clients 10 --runtime 2m
./pagebench  get-page-latest-lsn --mgmt-api-endpoint <http://srv:9898> --page-service-connstring postgresql://srv:64000 --limit-to-first-n-targets 3000 --num-clients 1 --runtime 2m

Manually ensure we can warm up

Results