Raptor: Mitigating CPU-GPU False Sharing Under Unified Memory Systems

The introduction of Unified Memory (UM) technology has greatly increased the programmability of CPU-GPU heterogeneous systems. At the same time, Unified Memory systems have given rise to new performance challenges. Achieving the desired performance and energy efficiency on such systems requires careful consideration of data allocation and migration. This paper looks at the problem of false sharing under UM. We present Raptor, a system for fast and accurate detection of page-level false sharing in heterogeneous applications. The system employs binary code instrumentation and leverages hardware performance counters to track UM allocations and data access patterns and pinpoint energy inefficiencies created by the occurrence of false sharing. Experiments on a suite of heterogeneous applications show false sharing can be a common occurrence in collaborative design paradigms with tight coupling of CPU-GPU tasks. When false sharing is eliminated via a padding scheme, applications are able to achieve higher performance at lower clock frequencies, leading to improved energy efficiency by as much as 2.96× and by 1.62× and 1.47× on average on two contemporary CPU-GPU platforms.