A 10G NetFPGA Prototype for In-Network Aggregation

An emerging class of distributed applications such as graph processing and key-value store require fine grained non-uniform memory access and communication, which commodity networking infrastructure is ill-equipped to handle. By batching messages on the end host we can increase bandwidth utilization at the cost of latency and memory. However, by batching messages in the network instead of exclusively at the end hosts, it is possible to reduce these memory and latency penalties. We introduce a 10G NetFPGA prototype that performs in-network message aggregation which for commodity systems can offer increased throughput at minimal latency and memory increases.