Presenter: David Novo

Abstract: Data storage and movement is a fundamental bottleneck in modern computing systems, greatly limiting their overall performance and energy efficiency. This presentation introduces two new techniques to achieve smarter data movement across the memory hierarchy. The first technique uses perceptron-based prediction to identify off-chip load requests using multiple program features. When the load is predicted to go off-chip, it issues a speculative request directly to the memory controller. As a result, the on-chip cache access latency is removed from the critical path when the prediction is correct. The second technique uses reinforcement learning for data placement in hybrid storage systems. These systems use multiple different storage devices to provide high and scalable storage capacity at high performance. The new data placement technique observes different features of the running workload and the storage devices to make system-aware data placement decisions. Both techniques leverage machine learning algorithms to optimize data movement, significantly outperforming state-of-the-art solutions.

Conference: MPSoC

Link: http://www.mpsoc-forum.org/archive/2023/

Date: June 30, 2023

Location: Fort Collins, Colorado, USA