Abstract
With the rapid growth of traffic volume in datacenter networks (DCNs), packet switches suffer from insufficient switching chip capacity and difficulties in transmission control, making it challenging to provide high goodput and low latency for emerging cloud applications. We present Zeropod, a future-proof DCN architecture featuring simplified zero-buffer packet switches inside the point-of-delivery (pod). Within each pod, traffic transmission is scheduled by a per-pod centralized scheduler for collision avoidance, enabling a highly simplified data plane, facilitating benefits like higher switching capacity and precise transmission control. Among the pods, buffered Core switches work as barriers and relay inter-pod data, limiting the scope of centralized scheduling and thus simplifying the control plane. Zeropod combines host-level and flow-level scheduling for high performance with low overhead. Evaluation results show that Zeropod consistently performs better or equivalent to traditional buffered DCN, particularly regarding flow completion time (FCT). When accounting for the increased switching capacity due to the removal of buffers, its performance is further improved. Zeropod explores an extreme end of the design spectrum, and we hope it can encourage further exploration in the DCN community.
Introduction
Datacenter traffic is exploding, doubling every year, while the capacity of switching chips only doubles every two years. This "SerDes gap" creates a significant bottleneck. Traditional switches use on-chip buffers to manage flow competition, but at high line rates (400Gbps+), these buffers lead to "bufferbloat," packet drops, and complex management overhead. Zeropod proposes a radical shift: removing buffers entirely from switches within a "pod" and replacing them with a centralized scheduler to prevent collisions before they occur.
The Zeropod Architecture
Zeropod organizes the network into zero-buffer pods connected by traditional buffered core switches. Within each pod, switches are simplified by removing buffers and forwarding tables, allowing manufacturers to reallocate chip area toward raw switching capacity.
The system relies on three main pillars:
- Centralized Scheduling: A per-pod scheduler manages "host-level" collisions by deciding exactly when and on which path every host sends data in fixed-size time chunks called "epochs."
- Source Routing: To eliminate forwarding tables, the scheduler provides specific paths to hosts, which are attached to packet headers for hop-by-hop routing.
- Hierarchical Control: While the scheduler handles host-level timing, endhosts manage "flow-level" priority, ensuring latency-sensitive "mice flows" are prioritized within allocated slots.
Buffered core switches act as barriers, limiting the scope of centralized scheduling to a single pod, which ensures scalability and prevents a "God-scheduler" bottleneck.
Performance and Results
The researchers evaluated Zeropod using the YAPS packet-level simulator against standard datacenter workloads like WebSearch and microbenchmarks like Incast (many-to-one traffic).
Key findings include:
- Mice Flow Optimization: Zeropod reduces the 99th-percentile Flow Completion Time (FCT) for small flows by approximately 50% compared to traditional buffered networks (using DCTCP) at high loads.
- Capacity Gains: By removing buffers, Zeropod can effectively double switching capacity on the same silicon area. Simulations of "Zeropod (2x)" show performance approaching the theoretical "Ideal" upper bound.
- Incast Resilience: The centralized scheduler prevents the buffer overflows typical of many-to-one traffic patterns, leading to near-ideal performance in Incast scenarios.
Discussion and Future Work
While Zeropod offers significant performance and hardware simplification benefits, it introduces new challenges. It requires extremely tight time synchronization between hosts and low-latency scheduler processing (targeted at 15µs). Furthermore, while the pod is bufferless, the architecture still relies on core-layer buffers. Future work will focus on physical ASIC design to validate area-saving theories in silicon and exploring fully end-to-end zero-buffer designs.
Summary updated