Apple XNU: Clutch Scheduler
Apple (via Hacker News):
The traditional Mach scheduler attempts to achieve these goals by expecting all threads in the system to be tagged with a priority number and treating high priority threads as interactive threads and low priority threads as batch threads. It then uses a timesharing model based on priority decay to penalize threads as they use CPU to achieve fairshare and starvation avoidance. This approach however loses the relationship between threads and higher level user workloads, making it impossible for the scheduler to reason about the workload as a whole which is what the end user cares about. One artifact of this thread based timesharing approach is that threads at the same priority level are treated similarly irrespective of which user workload they are servicing, which often leads to non-optimal decisions. It ultimately leads to priority inflation across the platform with individual subsystems raising their priority to avoid starvation and timesharing with other unrelated threads. The traditional thread level scheduling model also suffers from the following issues:
- Inaccurate accounting: CPU accounting at the thread level incentivizes creating more threads on the system. Also in the world of GCD and workqueues where threads are created and destroyed rapidly, thread level accounting is inaccurate and allows excessive CPU usage.
- Poor isolation: In the Mach scheduler, timesharing is achieved by decaying the priority of threads depending on global system load. This property could lead to a burst of activity at the same or lower priority band causing decay for the App/UI thread leading to poor performance and responsiveness. The scheduler offers very limited isolation between threads working on latency sensitive UI workloads and threads performing bulk non-latency sensitive operations.
The clutch scheduler is the timesharing algorithm for threads on a single cluster. The Edge scheduler extends on the clutch scheduler design to support multiple clusters of different performance and efficiency charecterstics. The Edge scheduler uses the clutch timesharing per cluster and adds other multi-cluster features such as thread placement, migration, round-robining etc.
