Abstract
Modern systems-on-chip employ numerous accelerators to scale performance and efficiency, with many user applications seeing significant gains. However, as we move to accelerate other parts of our system stack, particularly runtime and kernel system services, we find that existing operating systems lack mechanisms for seamless accelerator integration beyond offloading.
As a solution, we present a new operating system design, the Blended OS. Going beyond offloading and kernel bypass, the Blended OS enables the programmer to reshape the software stack by defining their bespoke system graph of user, runtime, and system components to communicate directly with each other. To enable this, we identify three critical design principles, Liberty, Equality and Collaboration, and provide a hardware system substrate enabling standard queue-based APIs and direct data structure access by services.
To evaluate our design, we build and run experiments on Pengwing, a Blended OS prototype. Pengwing is a 9 tile SoC that includes 3 RISC-V cores and 6 accelerators, including a hardware network stack, memory allocator, garbage collector, key-value store, and encryption units. Our evaluation on several workloads demonstrates both speedups and new system capabilities.
Short bio
Nazerke Turtayeva is a 4th-year PhD student at the University of California, Santa Barbara, advised by Jonathan Balkind. Her research focuses on HW-SW co-design for heterogeneous SoCs and domain-specific accelerators. Apart from her main research, she has interned at Microsoft Research and was a student researcher at Google DeepMind.Speakers
Speaker: Nazerke Turtayeva, PhD Student, UC Santa Barbara