The inference workloads of Large Language Models (LLMs) pose significant latency and cost challenges due to increasing model sizes and demand for real-time responses. Existing cluster schedulers for multi-instance LLM serving primarily focus on load balancing to optimize memory usage, which is insufficient for workloads with diverse request characteristics. In such cases, the compute layout — the arrangement of tokens across iterations within each instance—plays a crucial role in determining latency. We propose AdaGen, a workload-adaptive cluster scheduler that minimizes latency and thus maximizes SLO attainment by optimizing compute layouts across instances. AdaGen employs a multi-step scheduling strategy: it first classifies requests based on prefill and decode lengths, then balances load, and finally performs selective distributed execution across instances. Each step incrementally refines the scheduling based on the compute layouts derived from the decision of the previous step. To avoid the overhead of actual execution to generate the layouts, AdaGen introduces a novel simulation-based estimator. Extensive experiments using production workloads show that AdaGen achieves up to 3.6× higher SLO attainment and 2× better cost-efficiency compared to the existing systems, while ensuring scalability.

Collective communication operations in AI and HPC workloads generate heavy network traffic. Offloading these operations to network switches reduces latency, but performing arithmetic and replication at line rate is difficult, especially as port counts and link speeds grow. Existing in-network approaches rely on accumulation buffers that not only limit throughput but also require complex state management to handle stragglers and congestion. We present CCSwitch, a modular switching fabric built from 4×4 non-blocking Collective Engines (CEs). Each CE combines spatial and temporal parallelism to perform reductions without accumulation buffers. CEs compose into k-ary n-tree topologies, scaling to 32- and 256-port switches while preserving non-blocking throughput. Source routing and flit-level synchronization keep per-switch state minimal. Our FPGA implementation shows that CCSwitch's quaternary-tree reduction fabric uses up to 23% fewer LUTs and 12–30% fewer flip-flops than a comparable Clos-based design at equal throughput. Enabling the full feature set—source routing, replication, and time-multiplexed VCs—uses 1.4–1.8× more LUTs than the circuit-switched baseline, well below the 3–5× overhead typical of packet-switched NoC routers, while supporting concurrent collectives on shared links.

Serverless computing offers resource efficiency but suffers from a heavyweight data plane. We present Palladium, a DPU-offloaded serverless data plane enabling distributed zero-copy communication. Palladium uses two-sided RDMA and cross-processor shared memory to mitigate limitations of wimpy DPU cores. Its DPU-enabled network engine (DNE) isolates RDMA resources and manages flows across tenants. By converting HTTP/TCP to RDMA at ingress, Palladium reduces protocol overhead on the critical path.

Dataplane languages like P4 enable flexible and efficient packet-processing using domain-specific primitives such as programmable parsers and match-action tables. Unfortunately, P4 programs tend to be monolithic and tightly coupled to the hardware architecture, which makes it hard to write programs in a portable and modular way---e.g., by composing reusable libraries of standard protocols.

New applications where anyone can broadcast video are becoming very popular on smartphones. With the advent of high definition video, ISP providers may take the opportunity to propose new high quality broadcast services to their clients. Because of its centralized control plane, Software Defined Networking (SDN) seems an ideal way to deploy such a service in a flexible and bandwidth-efficient way. But deploying large scale multicast services on SDN requires smart group membership management and a bandwidth reservation mechanism to support QoS guarantees that should neither waste bandwidth nor impact too severely best effort traffic. In this paper, we propose a Network Function Virtualization based solution for Software Defined ISP networks to implement scalable multicast group management. Then, we propose the Lazy Load balancing Multicast (L2BM) routing algorithm for sharing the network capacity in a friendly way between guaranteed-bandwidth multicast traffic and best-effort traffic. Our implementation of the framework made on Floodlight controllers and Open vSwitches is used to study the performance of L2BM.

New applications where anyone can broadcast high quality video are becoming very popular. Internet services providers (ISPs) may take the opportunity to propose new high quality multicast services to their clients. Because of its centralized control plane, software defined networking (SDN) enables the deployment of such a service in a flexible and bandwidth-efficient way. But deploying large-scale multicast services on SDN requires smart group membership management and a bandwidth reservation mechanism with QoS guarantees that should neither waste bandwidth nor impact too severely best effort traffic. In this paper, we propose: 1) a scalable multicast group management mechanism based on a network function virtualization approach for software defined ISP networks to implement and deploy multicast services on the network edge and 2) the lazy load balancing multicast (L2BM) routing algorithm for sharing the core network capacity in a friendly way between guaranteed-bandwidth multicast traffic and best-effort traffic and that does not require costly real-time monitoring of link utilization. We have implemented the mechanism and algorithm, and evaluated them both in a simulator and a testbed. In the testbed, we experimented the group management at the edge and L2BM in the core with an open vSwitch-based QoS framework and evaluated the performance of L2BM with an exhaustive set of experiments on various realistic scenarios. The results show that L2BM outperforms other state-of-the art algorithms by being less aggressive with best-effort traffic and accepting about 5%-15% more guaranteed-bandwidth multicast join requests.

We propose to demonstrate Direct Code Execution (DCE), a framework that enables to execute nearly unmodified applications and Linux Kernel code jointly with the ns-3 simulator. DCE allows therefore fully deterministic reproducibility of network experiments. DCE also supports larger scale scenarios than real-time emulators by using simulation time dilatation. In this demonstration, we will showcase two main scenarios: (1) a basic example describing how to integrate in DCE the Data Center TCP (DCTCP) Linux kernel patch, and then how to customize this protocol and run it on different scenarios; (2) a more advanced use case demonstrating how to benefit from DCE to build a rich and realistic evaluation environment for Software Defined Wireless Networks based on Open vSwitch and the NOX SDN controller.