Optimization of checkpoint operations for deep learning computing

摘要

Systems and methods are provided to optimize checkpoint operations for deep learning (DL) model training tasks. For example, a distributed DL model training process is executed to train a DL model using multiple accelerator devices residing on one or more server nodes, and a checkpoint operation is performed to generate and store a checkpoint of an intermediate DL model. A checkpoint operation includes compressing a checkpoint of an intermediate DL model stored in memory of a given accelerator device to generate a compressed checkpoint, and scheduling a time to perform a memory copy operation to transfer a copy of the compressed checkpoint from the memory of the given accelerator device to a host system memory. The scheduling is performed based on information regarding bandwidth usage of a communication link to be utilized to transfer the compressed checkpoint to perform the memory copy operation, wherein the memory copy operation is performed at the scheduled time.