AURA: Adaptive Unified Runtime for Anomaly-Resilient Linear Algebra Operations

While PARALiA (2023) made significant strides in performance-aware autotuning for multi-GPU systems, it still relies on pre-computed models that struggle with unexpected workload deviations. This research proposes AURA, a runtime system that continuously learns from performance anomalies as they occur. Unlike PARALiA's model-based approach, AURA would implement online learning algorithms to detect when operations deviate from expected performance patterns, automatically triggering alternative execution strategies. The system would maintain a dynamic knowledge base of anomaly patterns and their effective mitigations, essentially learning from "mistakes" in real-time. This builds on FT-PBLAS's (2020) fault tolerance concepts but extends them to performance faults rather than just system failures. The key innovation is treating performance anomalies not as errors to be avoided, but as learning opportunities to improve future execution. This could be particularly valuable for exascale systems where workload characteristics are increasingly unpredictable.

References:

1. PARALiA: A Performance Aware Runtime for Auto-tuning Linear Algebra on Heterogeneous Systems. Petros Anastasiadis, Nikela Papadopoulou, G. Goumas, N. Koziris, Dennis Hoppe, Li Zhong (2023). ACM Transactions on Architecture and Code Optimization (TACO).
2. FT-PBLAS: PBLAS-Based Fault-Tolerant Linear Algebra Computation on High-performance Computing Systems. Yanchao Zhu, Yi Liu, Guozhen Zhang (2020). IEEE Access.