Microsoft at NSDI 2026: Advances in Large-Scale Network Systems

Microsoft announced that 11 papers targeting large-scale network systems were accepted at NSDI 2026. This year's achievements span data centers, wide-area networks, AI systems, and cloud infrastructure, with a focus on jointly improving throughput, stability, and efficiency.

• In LLM and AI systems, DroidSpeak shares KV cache across fine-tuned models with the same architecture, achieving up to 4x throughput and faster responses while minimizing output quality degradation. Eywa uses an LLM to automatically generate protocol models from natural language documentation, enabling model-based testing, and found 33 bugs. Of these, 16 were newly discovered issues. AVA combines an event knowledge graph with VLM-based agentic retrieval to support open-ended, ultra-long video analysis, and on AVA-100, achieved 75.8% accuracy based on 8 videos over 10 hours long and 120 manually annotated Q&A pairs.

• In the infrastructure and memory area, Octopus uses switchless CXL memory pods to lower costs and target multi-rack scaling. In a 3-server prototype, RPC was 3.2x faster than in-rack RDMA and 2.4x faster than a CXL switch. Pyrocumulus implements low-overhead live migration of storage-optimized VMs using an FPGA SmartNIC and LM protocol. ForestColl generates broadcast/aggregation schedules in polynomial time on heterogeneous network fabrics, securing both theoretical optimality and scalability.

• In operations, security, and efficiency, HEDGE accounts for probabilistic link capacity to absorb per-wavelength failures in optical networks, combining link-local and global resilience to maintain existing system-level throughput while reducing network disruption. MetaEase analyzes worst-case performance of heuristics using only source code, revealing performance gaps in real systems. HarvestContainers utilizes up to 75% of idle CPU without any application or OS modifications, while keeping tail latency within 4% of standalone. The community award-winning SONiC DASH SmartSwitch, deployed at scale on Azure, improves power and space efficiency through a hardware-friendly pipeline, unified switch architecture, and open development model. KRAKENGUARD uses load-time symbolic execution to finely isolate eBPF programs, blocking malicious behavior and also finding vulnerabilities.

Microsoft was a sponsor again this year, and its researchers also served on the program committee and steering committee.