The main contribution of this work is the discovery and study of four performance bugs in the Linux scheduler. These bugs cause the scheduler to leave cores idle while runnable threads are waiting for their turn to run. Resulting performance degradations are in the range 13-24% for typical Linux workloads, and reach 138x in some corner cases.

Here’s a graph of how many instructions per clock (IPC) a Core i7 achieves on Xalan, a sub-benchmark from SPEC CPU.

ask “why are things breakable” rather than “why did things break”