As computation and data are moved to the "cloud", there is a need to significantly increase the capacity and performance of the underlying data centers through improved hardware and communication infrastructure. A key component of the communication infrastructure is the network implementation, currently based upon Ethernet. Technological trends suggest that next-generation data center networks will be based upon RDMA (Remote Direct Memory Access), a powerful communciation technology used today in high performance computer systems. While RDMA offers enormous performance potential, current cloud infrastructures are not designed to exploit this potential. This project will investigate RDMA-based system designs for distributed storage systems and in-memory applications.
Specifically, the project research agenda consists of (a) building a high performance distributed key-value store and a B-tree based store (b) developing several applications in image search and deep learning whose distribution is made feasible by RDMA's ultra-low-latency and high throughput; (c) evaluating the performance of the resulting systems and applications at scale on the NSF-funded PRObE testbed.