Buggy Execution Greedy Coherence Preventing the SC Violation x x y ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅

Atomicity violations and violations of sequential consistency are two broad classes of concurrency errors that reduce the reliability of software. These failures often occur when different processors interleave their memory accesses at a fine grain. Greedy Coherence (GreCo) is a simple hardware technique that delays responses to some coherence requests to make this kind of sharing coarser, separating different processors’ accesses to the same memory location in time. These delays avoid operation interleavings that lead to some atomicity violations and sequential consistency (SC) violations, making software more robust to these types of faults. We describe two implementations of GreCo: one that monitors coherence protocol messages and uses additional hardware structures to identify and avoid likely atomicity and SC violations; and a lowercomplexity implementation that leverages the existing processor write buffer and focuses mainly on avoiding SC violations. Simulation results show that GreCo avoids failures in several bug kernel programs and incurs a negligible performance impact—less than 2%—on programs from the PARSEC benchmark suite.