Ziming Mao

Modern Mixture-of-Experts (MoE) workloads rely on expert parallelism (EP) to achieve high GPU efficiency. State-of-the-art EP communication libraries, such as DeepEP, rely on GPU-initiated RDMA communication. Although performant, they have poor portability across heterogeneous GPU and NIC hardwar...

Fast-evolving machine learning (ML) workloads have increasing requirements for networking. However, host network transport on RDMA NICs is hard to evolve, causing problems for ML workloads. For example, single-path RDMA traffic is prone to flow collisions that severely degrade collective communic...

This paper presents a set of in-process and remote distributed caches for datacenter environments, CLINK and CRINK respectively, that provide linearizable reads entirely from memory without contacting storage. These caches remain loosely coupled to the storage layer and achieve high performance,...