Performance issue on high concurrency
Bug #1157261 reported by
Zhou Yuan
This bug affects 4 people
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
Mirantis OpenStack |
Invalid
|
Critical
|
Aleksandr Shaposhnikov | ||
6.0.x |
Invalid
|
Critical
|
Aleksandr Shaposhnikov | ||
6.1.x |
Invalid
|
Critical
|
Aleksandr Shaposhnikov | ||
OpenStack Identity (keystone) |
Invalid
|
Wishlist
|
Unassigned |
Bug Description
Im currently using Keystone as the auth-server in Swift.
We have done performance test on our solution and we found that performance of keystone+swift is quite low, only 200 ops/s. While with the same setup the throughput can be improved to 7k ops/s by replacing keystone with swauth
We found the keystone process is fully saturating one sandy bridge core. This looks like a scalability issue.
Would be good if keystone could scale up well on a multi-core server.
Changed in mos: | |
importance: | Undecided → Critical |
milestone: | none → 6.0 |
Changed in mos: | |
assignee: | nobody → Aleksandr Shaposhnikov (alashai8) |
summary: |
- performance issue on high concurrency + Performance issue on high concurrency |
tags: | added: performance |
To post a comment you must log in.
100% agree; these benchmarks fall in line with my own.
I believe the biggest bottleneck is caused by our driver architecture and the division we maintain between them at the controller level (e.g., cross-driver calls are incredibly expensive for trivial lookups that could be done in SQL joins). I'm sure there are a lot of significant improvements we could make before addressing that issue, though (more specific & thus efficient driver calls; paging & filter hints at the driver level, etc).