A study of Scheduler Scalability

This presentation examines the scalability of the current scheduler under heavy load and increased CPU numbers. We show that the current scheduler design around a single run queue and single lock can become a bottleneck due to lock contention and thus limits scalability. We examine different implementations such as multi-queue and table-based priority schedulers that can reduce lock contention while maintaining the same functional behavior as the current scheduler. These implementations demonstrate better overall scheduling performance over a wide spectrum of loads and system configurations. We also investigate how performance can be further improved by relaxing the functional equivalence requirement with the current scheduler.