Presentation
HELM: Characterizing Unified Memory Accesses to Improve GPU Performance Under Memory Oversubscription
DescriptionUnified memory (UM) technologies simplify memory management across CPU and GPU domains in GPU-accelerated heterogeneous architectures through transparent data migration. However, the default migration mechanism can severely degrade performance when applications oversubscribe GPU memory. Existing approaches to mitigating this performance degradation often fail to generalize, as they target specific application types, require specialized hardware, or integrate opaque classification methods.
We introduce HEterogeneous Locality Metrics (HELM), a novel set of semantically meaningful metrics designed to characterize UM access patterns across diverse applications. These metrics are quantified using readily accessible UM driver telemetry data, providing users with tractable and interpretable UM memory characterizations. Such insight is critical for selecting optimal UM migration and placement policies under oversubscription. We demonstrate HELM’s accuracy and interpretability through access pattern analysis across various UM workloads. Experimental results on real systems show that HELM effectively guides policy selection, which outperforms default UM behavior by 3.5X on average.
We introduce HEterogeneous Locality Metrics (HELM), a novel set of semantically meaningful metrics designed to characterize UM access patterns across diverse applications. These metrics are quantified using readily accessible UM driver telemetry data, providing users with tractable and interpretable UM memory characterizations. Such insight is critical for selecting optimal UM migration and placement policies under oversubscription. We demonstrate HELM’s accuracy and interpretability through access pattern analysis across various UM workloads. Experimental results on real systems show that HELM effectively guides policy selection, which outperforms default UM behavior by 3.5X on average.
Event Type
Paper
TimeTuesday, 18 November 20251:52pm - 2:14pm CST
Location260-267
System Software and Cloud Computing