Block device layer IO scheduling method and system for artificial intelligence model training

Abstract

This application proposes a block device layer I / O scheduling method and system for training artificial intelligence models. The method includes: identifying checkpoint write flows at the block device layer; calculating an adaptive sharding threshold based on device characteristics and business requirements; submitting large I / O requests generated by merging the checkpoint write flows to a kernel delayed work queue for asynchronous splitting to generate an ordered I / O request shard chain; traversing the I / O request shard chain; performing a scheduling checkpoint for each small I / O request retrieved; and then distributing the small I / O request to the block device. This application reduces the latency of inference read I / O while ensuring checkpoint write throughput, without increasing the CPU overhead of I / O hot paths.

Images