Mohammad Alizadeh

Dr. Mohammad Alizadeh has fundamentally advanced how datacenters communicate efficiently in transporting data. One key contribution is the control of datacenter network congestion and packet loss with a groundbreaking datacenter transport control protocol, called DCTCP. DCTCP significantly increases performance in datacenter environments where state-of-the-art TCP protocols fall short. The theoretical foundation upon which DCTCP is built and the empirical analyses, novel algorithms, and explicit congestion notification techniques it leverages enable data packets to circumvent congestion at network routers and allow network endpoints to modify their transmission windows to mitigate congestion effectively while using significantly less buffer space. In essence, the processing of congestion signals is moved from within the network fabric to the network endpoints, enabling traffic bursts to be tolerated better and leading to reduced transport latency, high data throughput, and greater network utilization.

Alizadeh's key contribution led to important new techniques to balance network load in datacenters (CONGA and Letflow), to reduce flow completion times (pFabric), and to minimize packet transport latency of short messages on commodity hardware (Homa and HULL) while utilizing network capacity efficiently. Aspects of DCTCP are in the RDMA over Converged Ethernet (RoCE) standard, and DCTCP is in mainline Linux and supported by many network interface and switch manufacturers. Datacenters and products from industry-leading companies (Alibaba, Cisco, Google, Intel, Mellonox, Meta/Facebook, Microsoft, and Morgan Stanley, among others) have adopted various techniques Alizadeh pioneered, and some of his proposed techniques serve as the basis from which other datacenter network protocols are derived, including a transport protocol for remote direct memory access networks (HPCC). Moreover, highly-cited works coming from his key contributions have been recognized by best paper and test-of-time awards, and many of his techniques serve as the basis or standards by which new schemes proposed by other researchers in the field oftentimes are compared.

His key contribution and the works derived from it will continue to influence significantly how researchers in academia and industry think about, design, build, and operate high-performance datacenter networks and networking protocols deployable in current and future systems.