Improving Scalability of Xen
summary
Xen is extended to handle thousands of guests: a patch replaces select with poll/epoll to remove the 124‑FD console limit, a three‑level event API prototype lifts event channels from 4K toward 256K, and a 3,000‑VM run on an 8‑socket, 80‑core machine proves the changes work (instances run for days and deliver events). Measured bottlenecks are slow domain creation (seconds to hours at scale), Linux bridge 1,024 interface limit, PV backend buffer starvation, IO slowness and CPU contention requiring reserved physical CPUs.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 2:02Open Source
- 2:37Architecture
- 5:06Atrium Guest
- 5:46Current scalability status
- 8:11Simple calculation
- 10:30Backend Domains
- 13:04Scalability Issues
- 16:11To Set Limit
- 17:51Event Channel Limit
- 18:58Default ABI
- 21:09Designers Follow ABI
- 22:36Memory Footprint
- 23:33Pros and Cons
- 24:02People API
- 27:19Link View Design
- 28:04Link View Pros Cons
- 28:44Community Decision
- 36:06Mini OSP
- 38:34Hydramonster
- 43:20Observations
