Microsoft is proud to be a part of SIGCOMM 2026. Check out our sessions, covering network traffic engineering, new load balancing strategy, adaptive photonic switching schedules, AI network simulation, and more. https://t.co/mjAzIdLJuJ https://t.co/25Aw2CNYou

Microsoft at SIGCOMM 2026

Source: https://www.linkedin.com/pulse/microsoft-sigcomm-2026-microsoftresearch-lzoje Microsoft is proud to sponsorSIGCOMM 2026. This year’s conference takes place August 19-21 in Denver, bringing together professionals from academic and industrial backgrounds who research or deploy communication networks and networked systems. We hope to see you at the event. We invite you to explore the breadth ofsystems and networking research at Microsoftand to browse our conferencesessions, which are outlined below.

Wednesday, August 19, 2026

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Frules-offload-engine-roe-accelerating-host-sdn-policy-evaluation%2F&urlhash=Zap6&trk=article-ssr-frontend-pulse_publishing-image-blockExperience Session 4: Cloud Data Planes & Network Virtualization

Cloud service providers rely on software-defined networking (SDN) policies to implement complex network architectures and security semantics in multi-tenant environments. SDN policy evaluation is a mandatory step for each network packet in a virtual machine hosted on public cloud services. Evaluating SDN policies requires up to several hundred microseconds, which severely limits performance.

This research introduces a networking optimized architecture that embeds packet semantics directly into instructions. This eliminates expensive data caches and load-store units, which makes FPGA implementation more efficient. The authors design the ROE-Core, a multi-core, multi-threaded engine for SDN policy evaluation, using read-only instructions to deliver high throughput and low latency connection establishments. They also deploy an end-to-end hardware/software co-designed system that evaluates SDN policies, supports connection lifecycle management, and seamlessly interacts with existing accelerators on Azure to achieve line-rate performance.

Thursday, August 20, 2026

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Fnear-optimal-online-traffic-engineering%2F&urlhash=RGJo&trk=article-ssr-frontend-pulse_publishing-image-blockResearch Session 10: Internet Measurement & Routing

Most WAN traffic engineering systems rely on a centralized controller that periodically collects traffic data, computes routes, and updates the network. This approach often takes several minutes to react to changes at scale and may produce suboptimal results.

OnlineTEdelivers near-optimal traffic engineering decisions within seconds of network or traffic changes, significantly outperforming state-of-the-art solutions. Using distributed optimization, each switch solves part of the problem while a central coordinator synchronizes progress, with compute requirements well within the capabilities of modern switches. OnlineTE also enables edge-based traffic engineering at scale, which can use network resources more efficiently than conventional approaches.

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Fhoney-i-shrunk-the-headers-with-flow-zip%2F&urlhash=d6-k&trk=article-ssr-frontend-pulse_publishing-image-blockResearch Session 12: Host Networking & Packet Processing

Every packet of data sent across a network includes header information that helps route and manage traffic. While necessary, these headers consume bandwidth without carrying useful application data, reducing goodput (the amount of useful data delivered), and can increase flow completion times that slow down application performance. Modern networking techniques such as tunneling increases the header size and add more such overhead.

While it is possible to compress these headers, existing methods require specialized hardware on every hop to compress and decompress the packet to and from custom header formats.Flow.ZIPis a backward-compatible header compression mechanism designed for existing data center networks. It leverages a combination of last-hop network offload and MPLS support, both of which are ubiquitous in modern data center deployments.Flow.ZIPovercomes scalability limitations in these components by selectively coordinating compression for a subset of flows. As a result,Flow.ZIPachieves up to 58% reduction in average flow completion time on real-world data center workloads.

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Fstorm-enabling-traffic-scheduling-for-rdma%2F&urlhash=wMkc&trk=article-ssr-frontend-pulse_publishing-image-blockResearch Session 13: Datacenter Transport: RDMA & Congestion Control

Remote direct memory access (RDMA) is increasingly used across datacenter workloads with different scheduling needs, but in practice most traffic relies on simple fair sharing. This research presents STORM, a network interface card (NIC)-level scheduler for RDMA workloads that uses information already available to the NIC. STORM assigns a small number of network priority levels and prioritizes requests that are near completion or blocking other work. Prototyped on an FPGA NIC with negligible overhead, STORM reduced LLM training iteration time by up to 12% and cut average and P99 flow completion slowdown by up to 90% compared to fair scheduling.

Friday, August 21, 2026

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Fcapybara-dynamic-load-balancing-with-microsecond-scale-tcp-migration%2F&urlhash=L0Sk&trk=article-ssr-frontend-pulse_publishing-image-blockResearch Session 15: Programmable Switches & Data-Plane Hardware

Layer-4 load balancers are widely used to support scalable cloud services, but they perform poorly under unpredictable skewed workloads.Capybarais a new load balancer architecture that enables dynamic rebalancing of established connections. It divides load balancing responsibility into a fast L4 load balancer, a host-switch co-designed connection migration protocol, and a transport interface for application-level connection state migration. Capybara leverages programmable switches and kernel-bypass to efficiently implement connection migration without disruption, while maintaining transparency to clients. Under realistic workloads, Capybara achieves up to 149× lower tail latency and more than 2× higher throughput for scale-out services compared to state-of-the-art load balancing approaches.

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Fharvest-adaptive-photonic-switching-schedules-for-collective-communication-in-scale-up-domains%2F&urlhash=rtvf&trk=article-ssr-frontend-pulse_publishing-image-blockResearch Session 17: Congestion, Rate & Schedule Control

Silicon photonic interconnects are widely used to increase bandwidth and energy efficiency in today’s computer networks. While direct optical connections can reduce congestion and latency, frequent reconfiguration introduces overhead that can outweigh the benefits.

Harvestaddresses this challenge by automatically generating topology reconfiguration schedules that minimize collective communication time. Given a communication algorithm and schedule, Harvest determines how the interconnect should evolve over time, balancing reconfiguration costs against congestion and propagation delays. This approach adapts to different photonic technologies by accounting for their reconfiguration characteristics.

Through simulation and hardware emulation on commercial GPUs, Harvest significantly reduces collective completion time across multiple communication algorithms compared with both static interconnects and reconfigure-every-step approaches.

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Fnuwa-efficient-generative-control-plane-for-ai-network-simulation%2F&urlhash=O2to&trk=article-ssr-frontend-pulse_publishing-image-blockResearch Session 18: Learning-Based & Data-Driven Network Systems

Extremely large AI supercomputers support the rapid evolution of AI but training them is very expensive. Network simulation helps improve the efficiency of large-scale AI training simulation. However, high-fidelity network simulation is very slow at scale.

This research shows that the efficiency of the control plane is a major bottleneck in the simulation of large-scale training clusters. It introducesNüwa,an efficient generative control plane for AI network simulation. Nüwa leverages the layered network architecture of an AI network to express routing information in a formula for each layer. Evaluations show that Nüwa can reduce routing table generation from hours to only 20 seconds for 64K nodes. For data plane execution, Nüwa can reduce the overall simulation time over 100x by almost eliminating the forwarding calculation.