Mowei Wang

Recent advances in fast optical switching show promise in meeting the high goodput and low latency requirements of datacenter networks. We present NegotiaToR, a simple network architecture for optical reconfigurable DCNs that utilizes on-demand scheduling to handle dynamic traffic. In NegotiaToR, racks exchange scheduling messages through an in-band control plane and distributedly calculate non-conflicting paths from binary traffic demand information. Optimized for incasts, it also provides opportunities to bypass scheduling delays. NegotiaToR is compatible with prevalent flat topologies, and is tailored towards a minimalist design for on-demand reconfigurable DCNs, enhancing practicality. Through large-scale simulations, we show that NegotiaToR achieves both small mice flow completion time and high goodput on two representative flat topologies, especially under heavy loads. Particularly, the flow completion time of mice flows is one to two orders of magnitude better than the state-of-the-art traffic-oblivious reconfigurable DCN design.

Recent advances in fast optical switching technology show promise in meeting the high goodput and low latency requirements of datacenter networks (DCN). We present NegotiaToR, a simple network architecture for optical reconfigurable DCNs that utilizes on-demand scheduling to handle dynamic traffic. In NegotiaToR, racks exchange scheduling messages through an in-band control plane and distributedly calculate non-conflicting paths from binary traffic demand information. Optimized for incasts, it also provides opportunities to bypass scheduling delays. NegotiaToR is compatible with prevalent flat topologies, and is tailored towards a minimalist design for on-demand reconfigurable DCNs, enhancing practicality. Through large-scale simulations, we show that NegotiaToR achieves both small mice flow completion time (FCT) and high goodput on two representative flat topologies, especially under heavy loads. Particularly, the FCT of mice flows is one to two orders of magnitude better than the state-of-the-art traffic-oblivious reconfigurable DCN design.

Converged Ethernet employs Priority-based Flow Control (PFC) to provide a lossless network. However, issues caused by PFC, including victim flow, congestion spreading, and deadlock, impede its large-scale deployment in production systems. The fine-grained experimental observations on switch buffer occupancy find that the root cause of these performance problems is a mismatch of sending rates between end-to-end congestion control and hop-by-hop flow control. Resolving this mismatch requires the switch to provide an additional buffer, which is not supported by the classic dynamic threshold (DT) policy in current shared-buffer commercial switches. In this paper, we propose Selective-PFC (SPFC), a practical buffer management scheme that handles such mismatch. Specifically, SPFC incrementally modifies DT by proactively detecting port traffic and adjusting buffer allocation accordingly to trigger PFC PAUSE frames selectively. Extensive case studies demonstrate that SPFC can reduce the number of PFC PAUSEs on non-bursty ports by up to 69.0%, and reduce the average flow completion time by up to 83.5% for large victim flows.

Traffic modeling of Datacenter Network (DCN) today is over-simplified, deviating from the ground truth. Adopted by numerous researchers, the common practice relies on the assumptions of traffic homogeneity and independence for ease of use. Based on our investigation of a real-world traffic dataset, we disprove these assumptions and point out the severe fidelity issue of the common practice that could invalidate many motivations and conclusions from influential research works. In this paper, we present Encore, a DCN traic modeling framework for ine-grained traic modeling and high-fidelity synthetic traffic generation. Leveraging machine learning techniques, Encore effectively extracts and preserves essential distribution and sequential features from raw traic. Preliminary experiments demonstrate that the traic generated by Encore not only restores the key features of real traffic but also achieves high consistency when used to evaluate network performance. We envision further expanding Encore to full-process traffic modeling and generation, and expect these critical improvements in traffic models can facilitate the DCN performance evaluation and optimization.

Today’s network is notorious for its complexity and uncertainty. Network operators often rely on network models for efficient network planning, operation, and optimization. The network model is responsible for understanding the complex relationships between network performance metrics (e.g., delay and jitter) and network characteristics (e.g., traffic and configuration). However, we still lack a systematic approach to developing accurate and lightweight network models that are aware of the impact of network configurations (i.e., expressiveness) and provide fine-grained flow-level temporal predictions (i.e., granularity). In this paper, we propose xNet, a data-driven network modeling framework based on graph neural networks (GNN). It is worth noting that xNet is not a dedicated network model designed for a specific network scenario with constraint considerations. On the contrary, xNet provides a general approach to modeling the network characteristics of concern with relation graph representations and configurable GNN blocks. xNet learns the state transition functions between time steps and rolls them out to obtain the full fine-grained prediction trajectory. We implement and instantiate xNet with three use cases. The experimental results show that xNet can accurately predict different performance metrics (i.e. temporal and steady-state QoS) in different scenarios, with performance comparable to state-of-the-art domain-specific models. Compared with traditional packet-level simulators, xNet achieves a speed improvement of more than two orders of magnitude, demonstrating its promising application in real-time optimization of network configurations.

Switch buffer serves an important role in the modern internet. To achieve efficiency, today's switches often use on-chip shared memory. Shared memory switches rely on buffer management policies to allocate buffer among ports. To avoid waste of buffer resources or excessive buffer occupation by a few ports, existing policies tend to maximize overall buffer utilization and pursue queue length fairness. However, blind pursuit of utilization and misleading fairness definition based on queue length lead to buffer occupation with no benefit to throughput but extends queuing delay and undermines burst absorption of other ports. With analysis of current dynamic threshold policies, we demonstrate that meaningless buffer occupation can potentially impair the absorption capability of shared buffer, whereas none of the existing policies have addressed this problem. We contend that a buffer management policy should proactively detect port traffic and adjust buffer allocation accordingly. In this paper, we propose Traffic-aware Dynamic Threshold (TDT) policy. On the basis of the classic dynamic threshold policy, TDT proactively raises or lowers port threshold to absorb burst traffic or evacuate meaningless buffer occupation. We present detailed designs of port control state transition and state decision module that detect real-time traffic and change port thresholds accordingly. Simulation and DPDK-based real testbed demonstrate that TDT simultaneously optimizes for throughput, loss and delay, and reduces up to 50% flow completion time.

The increasingly complicated and diverse applications have distinct network performance demands, e.g., some desire high throughput while others require low latency. Traditional congestion controls (CC) have no perception of these demands. Consequently, literatures have explored the objective-specific algorithms, which are based on either offline training or online learning, to adapt to certain application demands. However, once generated, such algorithms are tailored to a specific performance objective function. Newly emerged performance demands in a changeable network environment require either expensive retraining (in the case of offline training), or manually redesigning a new objective function (in the case of online learning). To address this problem, we propose a novel architecture, DeepCC. It generates a CC agent that is generically applicable to a wide range of application requirements and network conditions. The key idea of DeepCC is to leverage both offline deep reinforcement learning and online fine-tuning. In the offline phase, instead of training towards a specific objective function, DeepCC trains its deep neural network model using multi-objective optimization. With the trained model, DeepCC offers near Pareto optimal policies w.r.t different user-specified trade-offs between throughput, delay, and loss rate without any redesigning or retraining. In addition, a quick online fine-tuning phase further helps DeepCC achieve the application-specific demands under dynamic network conditions. The simulation and real-world experiments show that DeepCC outperforms state-of-the-art schemes in a wide range of settings. DeepCC gains a higher target completion ratio of application requirements up to 67.4% than that of other schemes, even in an untrained environment.