Offloading the Training of an I/O Access Pattern Detector to the Cloud

Abstract

I/O operations are a bottleneck for numerous applications, so optimizing the performance of these operations is of paramount importance. Many techniques explore and apply optimizations to different layers of the I/O stack to improve performance. The difficulty that arises is that the workload changes constantly. So detecting access patterns correctly, at runtime, becomes essential for systems that seek to self-adjust their parameters. Furthermore, the I/O pattern detection techniques should represent minimal overhead and should be able to perform detection as quickly as possible. This paper approaches a machine learning technique for detecting the I/O access patterns and proposes offloading the local training workload to the cloud using a TPU accelerator. Such an approach does not interfere with classifier accuracy (reaching up to 99% accuracy). Still, it allows the training to be asynchronous, enabling the local machine to allocate its computing resources to scientific applications while the model is trained or updated in the cloud.