RingStitch improves resource use in fragmented TPU clusters
RingStitch: Demand-Aware Optical Stitching for Fragmented TPU Clusters
Networking and Internet Architecture
Summary
Large AI training centers use special optical connections to link many processors together. However, smaller jobs often leave unused processor space scattered across different racks, which is hard to use efficiently. The authors introduce RingStitch, a scheduler that cleverly connects these scattered resources, making it easier to run new jobs without wasting capacity. Their simulations show better job scheduling compared to usual methods.
What this means in practice
- •For data center operators: Improve scheduling efficiency by connecting fragmented TPU resources across racks to better utilize partial free capacity for diverse jobs.
- •For network infrastructure teams: Design optical interconnect configurations that minimize cross-rack communication costs while using flexible logical rings for resource aggregation.
Authors
Huiru Ao, Fan Yang, Binglei Wang, Bo Liu, Jialong Li
Abstract
Large-scale AI training clusters increasingly use optical circuit switching (OCS) to reconfigure rack-level interconnects and create elastic accelerator slices. In multi-tenant TPU-style clusters, however, small and medium jobs often leave partial free capacity stranded inside racks. Although the aggregate free capacity may be sufficient for a new job, it cannot be used by local placement or coarse full-rack stitching. This paper presents RingStitch, an OCS-based defragmentation scheduler that turns fragmented rack capacity into schedulable resources. RingStitch follows a local-first policy, stitches compact cross-rack fragments only when needed, and orders the selected racks into a low-cost logical ring. Simulations on a TPU 8t-like SuperPod model show that RingStitch improves schedulability over local and full-rack baselines.