Implementing Ultra-Low-Latency Datacenter Services with Programmable Logic

Datacenters require many low-level network services to implement high-level applications. Key-value store (KVS) is a critical service that associates values with keys and lets machines share these associations over a network. Client machines send a key search request over Ethernet, and KVS servers return values associated with these keys. Critical metrics for the KVS service in the datacenter are latency, throughput, and power. Low latency is desired to minimize the time required for a client to wait after it sends a key to a server and the value is returned. High throughput is sought to maximize the rate at which each server performs lookups. Power-efficient implementations are desired to reduce the amount of energy required to operate the service. Most existing datacenters run KVS services using software on microprocessor cores. In this work, the authors implemented KVS in field-programmable gate array (FPGA) logic and found it to be 14 times faster while using 13 times less power than the best software implementation using kernel bypass. A latency below a half-microsecond (0.5 μs) and throughput of 150 million searches per second (MSPS) was achieved.