As data center workloads evolve to become increasingly compute-intensive, there is a growing need for accelerators. There are a wide variety of accelerators, spanning GPUs, FPGAs, ASICs, and workload-specific ones such as TPUs. The Cyborg project in OpenStack aims to ease the adoption and lifecycle management of these diverse accelerator types.
Cyborg and Nova developers have put together an architecture to enable offload to various accelerators. The architecture includes FPGAs, which have unique needs for programming and bitstream management. In this presentation, we will look at use cases for offloads to devices in general, programming models for FPGAs, and the representation of devices (including FPGAs) in Placement. We will take a close look at the scheduling of instances that need accelerators. We will detail the archictecture of os-acc, a library for Nova compute to interact with Cyborg. Finally, we will present the current status of Cyborg development.
The attendees will learn about the need for accelerators, use cases for accelerators (including FPGAs) in the cloud, and the architecture of Cyborg/Nova to handle accelerators.