Active Pages 20 Years Later: Active Storage for the Cloud

Todays data center and cloud architectures decouple compute and storage resources for better scalability. While this enables elastic scale-out, it also introduces additional data movement. Even though existing systems already offload processing into the storage layer, they struggle to strike a balance between compute capacity and the bandwidth of the network or storage device. Solutions based on traditional servers suffer from a bandwidth mismatch between storage, CPU, and network. At the same time, they often overprovision compute capacity. Network storage devices are better balanced in terms of bandwidth but provide limited processing capabilities. With Caribou we explore an alternative design that offers rich offloading functionality while balancing the bandwidth requirements of storage and network. Caribou can process data originating from the storage at a multi-GB/s rate and consume data from the network at line-rate, all while not having overprovisioned computing capacity.