AI training and inference are driving cloud networks toward terabit-per-second (Tbps) bandwidth per server, challenging the scalability and efficiency of today's cloud network architectures. A prevalent design scales bandwidth by stacking monolithic Data Processing Units (DPUs), but this approach tightly couples control and data plane resources, leading to excessive cost, power consumption, and operational complexity. We identify a fundamental control-data plane divergence in AI clouds: while data plane bandwidth demand grows rapidly, control plane demand remains largely flat due to the dominance of elephant flows. As a result, monolithic DPUs become systematically over-provisioned when used as bandwidth scaling primitives.

The increasing popularity of large models and datasets has highlighted the significance of distributed training networks. As gradient synchronization generates substantial traffic, in-network aggregation (INA) has emerged as a solution to offload aggregation onto the switch, alleviating network congestion and accelerating distributed training. However, the limited memory capacity of the INA switch becomes a potential bottleneck as computation shifts into the network, especially in multi-tenant scenarios. To address this bottleneck and enhance network throughput, we propose the Aggregation with Innetwork Resource Pooling (AIRP) framework. Unlike existing approaches that optimize individual switches in a localized manner, AIRP takes a holistic view and efficiently pools switch memory resources across the entire network, allocating them to multiple tenants. Evaluation using the ns-3 simulator and P4 testbed demonstrates that AIRP can accelerate the training of various models, including computer vision and language models. The experimental results show that AIRP outperforms existing INA approaches by up to 7 times in terms of network throughput in multi-tenant scenarios, while also achieving great flexibility and efficiency in deployment.

Interconnection network in HPC is becoming a bottleneck due to increasing traffic load. We model adaptive routing mechanisms and prove that even with advanced adaptive routing, static networks like Dragonfly cannot handle non-uniform traffic efficiently, let alone the frequently changing non-uniform traffic. Therefore, it requires architectural changes for network-wide improvements, e.g., reconfigurable networks.Existing reconfigurable networks hardly support agile reaction to traffic changes with little impact on network. Therefore, we propose MUSE1, a Dragonfly-based runtime incrementally reconfigurable network to enable a small number of link adjustments for agility and little impact on transmitting flows during every reconfiguration with optical circuit switch (OCS).Simulations with both synthetic traffic and real-world workloads prove that MUSE can prevent saturation under typical traffic patterns that cause congestion in static Dragonfly. MUSE is 30-55% better than static Dragonfly and Flexfly w.r.t commonly used performance metrics like flow completion time (FCT). We also build a MUSE prototype and demonstrate that MUSE enables 20-30% less application finish time (AFT).

Network experiments are essential to network-related scientific research (e.g., congestion control, QoS, network topology design, and traffic engineering). However, (re)configuring various topologies on a real testbed is expensive, time-consuming, and error-prone. In this paper, we propose Software Defined Topology Testbed (SDT), a method for constructing a user-defined network topology using a few commodity switches. SDT is low-cost, deployment-friendly, and reconfigurable, which can run multiple sets of experiments under different topologies by simply using different topology configuration files at the controller we designed. We implement a prototype of SDT and conduct numerous experiments. Evaluations show that SDT only introduces at most 2% extra overhead than full testbeds on multi-hop latency and is far more efficient than software simulators (reducing the evaluation time by up to 2899x). SDT is more cost-effective and scalable than existing Topology Projection (TP) solutions. Further experiments show that SDT can support various network research experiments at a low cost on topics including but not limited to topology design, congestion control, and traffic engineering.

More and more web applications are deployed on the cloud storage services that store data objects of the web applications in the geo-distributed datacenters belonging to Cloud Service Providers (CSPs). In order to provide low request latency to the web application users, in the previous work, the web application developers need to store more data object replicas in a large number of datacenters or send redundant requests to multiple datacenters (e.g., closest datacenters), both of which increase monetary cost. In this paper, we conducted request latency measurement from a GENI server (as a client) to AWS S3 datacenters for one month, and our observations lay the foundation for our proposed system called GeoCol, a geo-distributed cloud storage system with low cost and latency using reinforcement learning (RL). To achieve the optimal tradeoff between the monetary cost and the request latency, GeoCol encompasses a request split method and a storage planning method. The request split method uses the SARIMA machine learning (ML) technique to predict the request latency as an input to an RL model to determine the number of sub-requests and the datacenter for each sub-request for a request in order to enable the parallel transmissions for a data object. In the storage planning method, each datacenter uses RL to determine whether each data object should be stored and the storage type of each stored data object. Our trace-driven experiment on AWS S3 and GENI platform shows that GeoCol outperforms other comparison methods in monetary cost with 32 % reduction and data object request latency with 51 % reduction.