NegotiaToR: Towards A Simple Yet Effective On-demand Reconfigurable Datacenter Network

Title: NegotiaToR: Towards A Simple Yet Effective On-demand Reconfigurable Datacenter Network

Authors: Cong Liang, Xiangli Song, Jing Cheng (Tsinghua University); Mowei Wang, Yashe Liu, Zhenhua Liu (Huawei Technologies Co., Ltd); Shizhen Zhao (Shanghai Jiao Tong University); Yong Cui (Tsinghua University)

Scribe: Xiaoqiang Zheng(Xiamen University)

Introduction:
With more and more applications running in the cloud, traffic demands in datacenter networks(DCN) increase continuously. Among the applications, tasks like high performance computing (HPC) are greatly affected by both goodput and latency, putting stringent requirements on the network. However, existing packet-switched DCNs struggle to meet these requirements due to inadequate capacity of switching chips. To address this, the development of optical switching technology especially fast optical switching has led researchers to turn to reconfigurable optical DCNs, which provide higher capacity as well as lower cost compared with packet switching.

Key idea and contribution:
The authors present NegotiaToR, a simple reconfigurable DCN architecture with scalable on-demand scheduling for flat topologies with fast optical switching enabled. With in-band distributed scheduling, it dynamically adapts the optical links among top-of-rack (ToR) switches to traffic demands, and sends data directly to destinations through one-hop paths. Beyond the scheduled connections, NegotiaToR also provides unscheduled connections, mitigating the impact of scheduling delays even under incasts. Utilizing existing arrayed waveguide grating routers (AWGR) and fast-tunable lasers, NegotiaToR is compatible with prevalent flat topologies, achieving a better performance than the state-of-the-art traffic-oblivious scheme on the same hardware with similar complexity.

NegotiaToR comprises the distributed NegtotiaToR Matching scheduling algorithm, the two-phase epoch serving as in-band control plane and data plane, and a mechanism to bypass scheduling delays even under incasts. Furthermore, the authors explore the possibilities to trade off simplicity for better performance, and show that extra complexity does not necessarily translate into proportionate performance gains.

Evaluation:
The authors investigate the effectiveness of NegotiaToR’s design elements, and evaluate its performance through large-scale simulations using a packet-level simulator, YAPS. The evaluation shows that NegotiaToR achieves better mice flow FCT at all loads regardless of topologies. With priority queues enabled, NegotiaToR’s FCT is consistently one to two orders of magnitude better. Even without priority queues, such significant advantages still exist at lighter loads. Regarding goodput, NegotiaToR remarkably outperforms the traffic-oblivious scheme on both topologies at high loads. In contrast, NegotiaToR’s on-demand scheduling can better utilize the network’s bandwidth and reduce bandwidth waste. In conclusion, with similar complexity, NegotiaToR outperforms the state-of-the-art traffic-oblivious design in both FCT and goodput.

Q1: Hi, very nice work. I’m wondering if your NegotiaToR protocol will write into consistency issues because it supports multiple requests at the same time and has a default schedule. So could you comment on the consistency issue? I mean, for example, what if you come up with several schedules that conflict with each other? Or in the worser case, can they run into deadlocks?

A1: Normally this kind of thing won’t happen. And because one design in NegotiaToR is that it’s stateless. It’s just a process that all the messages are received and then send the results back. It doesn’t care about states.

Q2: Thank you, nice work. I have one simple question like if a tour sends multiple requests, is it possible that this tour receives zero reply?

A2: It is possible under unlucky circumstances. For example, all requests are rejected by destination.

Personal thoughts:
This paper presents NegotiaToR, an on-demand reconfigurable datacenter network (DCN) architecture with a simple design. I appreciate the design because it utilizes the fast optional switching technology, which is able to meet the high goodput and low latency requirements of DCNs. I believe that NegotiaToR can facilitate the development of next-generation reconfigurable DCNs in the near future.