Computer Science Thesis Proposal

— 1:30pm

In Person and Virtual - ET - Reddy Conference Room, Gates Hillman 4405

HUGO DE FREITAS SIQUEIRA SADOK MENNA BARRETO, Ph.D. Student, Computer Science Department, Carnegie Mellon University

Streaming Abstractions for Intra-Host Communication

Hosts have historically been designed around the CPU, with peripheral devices playing an auxiliary role. Today, however, the rise of specialized computing architectures has shifted the role of peripheral devices. Accelerators and SmartNICs now perform a significant fraction of computation and can work independently from the CPU. Yet, existing abstractions for intra-host communication still assume CPU control. This abstraction mismatch forces the CPU to be on the datapath; compromising performance and scalability. These problems stem from three fundamental issues with the existing abstractions. First, existing abstractions need ad-hoc and complex routing logic between devices, requiring the CPU to be involved in the routing of data, even when the data does not need to be processed on the CPU. Second, existing abstractions make data accesses unpredictable, making it hard to mask access latencies. Finally, existing abstractions impose fixed and application-specific data formats, requiring the CPU to reformat the data in order to glue different devices and applications.

The main claim in this work is that streaming abstractions allow intra-host communication to be more scalable and performant. Regarding routing, streaming allows the CPU to be removed from the datapath, only being used to make coarse-grained decisions that can then be implemented in the dataplane outside the CPU. Regarding access latencies, streaming makes data accesses predictable, this can be leveraged by CPUs and accelerators, allowing them to fetch the next input ahead of time. Regarding data format, streaming does not impose data boundaries, allowing the same interface to be used for different kinds of functionalities with minimal glue logic.

We show the benefits of streaming for intra-host communication with Ensō, a streaming interface designed for communication between NICs and the CPU. We then discuss our proposal for bringing the benefits of streaming to other devices through a programmable host interconnect.

Thesis Committee 

Justine Sherry (Chair)
David G. Andersen
James C. Hoe
Arvind Krishnamurthy (University of Washington)
Aurojit Panda (New York University)

Additional Information

Add event to Google
Add event to iCal