FlacIO: Flat and Collective I/O for Container Image Service

This conference talk from USENIX FAST '25 presents "FlacIO: Flat and Collective I/O for Container Image Service," a novel approach to addressing I/O bottlenecks in container image services. Discover how researchers from Huawei Technologies identified that existing container image solutions suffer from high I/O amplification and excessive network traffic due to storage-oriented and global-oriented container image abstraction. Learn about their proposed memory-oriented and service-oriented "runtime image" abstraction that represents the memory state of the container service's root file system, enabling more efficient network transfer and faster root file system construction. The presentation details the design and implementation of FlacIO, an I/O accelerator based on this runtime image concept that works with a runtime page cache on host nodes. Evaluation results demonstrate impressive performance improvements, with FlacIO reducing container cold startup latency by up to 23 times compared to full image solutions and 4.6 times compared to lazy loading approaches. The talk also covers real-world application benefits, showing up to 2.25 times performance speedup in object storage scenarios and 1.7 times improvement in machine learning training environments.