Authors: Zhuo Huang, Hao Fan, Kun Wang, Hang Huang, Lin Gu, Chen Yu, Song Wu, and Hai Jin (Huazhong University of Science and Technology (HUST))
Abstract: HPC workflows are growing in scale and complexity, demanding elastic resource provisioning and fine-grained parallel execution. While serverless computing is conceptually well-suited due to its on-demand model, existing frameworks are primarily designed for lightweight, event-driven workloads, leading to fundamental mismatches in highly concurrent HPC scenarios.
In such workloads, large numbers of function instances must be created simultaneously, introducing significant startup and communication overhead. Kernel-dependent mechanisms—such as cgroups, namespaces, and the network stack—become bottlenecks: instance initialization triggers heavy lock contention, while cross-node communication incurs additional costs from data copying, serialization, and context switching.
To address these issues, we propose eFaaS, a serverless architecture tailored for HPC. By leveraging eBPF, eFaaS enables lock-free isolation and zero-copy communication, minimizing kernel overhead. Experimental results show that eFaaS maintains stable performance under extreme concurrency, significantly reduces long-tail latency, and achieves end-to-end speedups of 1.27×–3.74× across representative HPC workflows.
Presentation: file