HSToR: Enhancing Scalability for Optical Circuit Switched Networks with Recirculation

Combining optical circuit switches (OCSes) and electrical packet switches (EPSes) has emerged as an appealing architecture for data centers. However, in this hybrid data center network architecture with widespread application, challenges persist in data packet scheduling and buffer management for OCS-based matching patterns due to spatial and temporal traffic pattern constraints. Given the demand-agnostic way for OCS configurations, where each switch can only provide matching between input and output ports at any given time. In other words, the availability of optical circuits is intermittent, resulting in potential blockage of flows between them when optical circuits are unavailable. Traditionally, in the event of flow blockage, multiple queue buffers are maintained at Top-of- Rack (ToR) switches, with the number of queues corresponding to the number of destinations. Unfortunately, the “per-destination queue” approach suffers from poor scalability. To address this issue, we propose HSToR, a novel design for ToR switches, aimed at enhancing scalability and data transfer efficiency in optical circuit switched networks. Experimental results demonstrate that HSToR achieves high scalability, maintains lower latency as the network scales, better addresses packet loss during buffering, and sustains high throughput and link utilization.