Global Distributed Client-side Cache for HPC and AI Workloads

Explore the design and implementation of an efficient distributed client-side caching framework for HPC and AI workloads in this 51-minute conference presentation. Learn how massive datasets and complex computations at exascale speeds create communication overhead bottlenecks in distributed storage environments, and discover why traditional server-side and node-local client-side caching approaches fall short. Examine the limitations of server-side caching due to fixed memory and network bandwidth constraints, and understand how traditional client-side caching leads to inefficiencies through redundant data duplication and lack of global synchronization. Analyze the advantages of global distributed client-side caching over high-speed interconnects, which leverages aggregate resources including DRAM, local SSDs, network bandwidth, and RDMA capabilities across client nodes while scaling independently of server nodes. Delve into the specific challenges of implementing such systems and explore the carefully tuned policies required to manage valuable resources effectively. Gain insights into communication overhead challenges in distributed storage systems, examine the benefits of leveraging aggregate resources with high-speed interconnects, and study the key components of the proposed distributed client-side caching framework that addresses microsecond latency requirements for time-critical HPC and AI applications.