Research

Let an application use a large portion of a shared cluster for a very short time, so big-data work finishes in seconds at no extra cost.

Systems that keep expensive GPUs busy — burst-parallel training, GPU multiplexing, and cost-efficient MoE serving.

Nu and Quicksand split applications into tiny migratable units, so datacenters can harvest stranded resources at microsecond timescales.

Trading massive scale-out on efficient hardware for datacenter energy savings, including SmartNIC-hosted clusters.

Consistent replication in 1 RTT by replicating commutative operations before ordering them.

Using synchronized clocks to cut EPaxos conflict rates by up to 50% in wide-area replication.

Asynchronous BFT consensus with near-optimal throughput on variable-bandwidth networks.

Distributed sorting redesigned from scratch to finish within a 1-millisecond time budget.

A DRAM-based key-value store aggregating thousands of servers, with 5 µs reads and durable 13.5 µs writes.

A general-purpose layer that turns at-least-once RPCs into exactly-once, linearizable operations — adding only 530 ns.

One storage view over mix-and-match cloud storage services, with consistent client-side caching and 1-RTT transactions.

Server-driven admission control (Breakwater) and lock-contention-aware overload control (Protego) for µs-scale services.

A logging system with 8–18 ns log invocations — 1–2 orders of magnitude faster than Log4j2 or spdlog.

An efficient latency debugging tool that helps developers find the causes of latency spikes in multithreaded datacenter applications.