Next-generation IoT workloads, such as machine learning and image recognition, are demanding more computation resources and challenging the way a cloud is designed today. There is an emerging trend to use hardware accelerators, such as GPU and FPGA, to accelerate specific applications, including networking, security and machine learning. Data centers and cloud providers are investing in these efficient hardware accelerators solution and allowing users to access these optimized resources on demand. Given the range of users, from the hardware-savvy to the newcomers, enabling easy access to specialized accelerators while getting the best performance in a cloud environment can be a challenging task.
This presentation will discuss recent developments for hardware accelerator solutions in the open source community, with a focus on FPGAs and Smart NICs based on FPGAs.
We will briefly introduce FPGAs and delineate the cloud usage models for FPGAs. We will then do a deep dive into cloud orchestration frameworks (OpenStack and Kubernetes) and discuss how these frameworks handle devices and resources (for example, OpenStack Nova and Cyborg, Kubernetes Device Plugin). We will identify the challenges and gaps in both orchestration frameworks, which need to be addressed in the future.