Christiaan Banister

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2022-144

May 19, 2022

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-144.pdf

System-on-a-Chip designs with specialized processors and domain-specific accelerators have grown in popularity over the last decade to meet ever-increasing compute demands. This is due to the superior performance-per-watt that can be obtained from such designs as well as the inherent limitations in scaling up traditional out-of-order superscalar CPUs. However, a heavy reliance on microarchitectural details and a lack of coordination between software writers and hardware designers makes targeting these specialized IPs highly non-trivial. In addition, offloading tasks from the CPU to an accelerator introduces overheads that may nullify the advantage of utilizing it in the first place. This paper investigates a transparent method for allowing programmers to run user code on digital signal processing units (DSPs) and quantifies the impact of such an offload.

We propose a lightweight framework at the system software level in the style of message passing Asymmetric Multiprocessing kernels. By pairing a large symmetric multiprocessing kernel with a lightweight real-time operating system, we find that we can achieve low-latency accelerator communication while maintaining a relatively straightforward programming model.

Advisors: Krste Asanović

---

BibTeX citation:

@mastersthesis{Banister:EECS-2022-144,
    Author= {Banister, Christiaan},
    Title= {Low Overhead Remote Procedure Call System for Saturn DSP},
    School= {EECS Department, University of California, Berkeley},
    Year= {2022},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-144.html},
    Number= {UCB/EECS-2022-144},
    Abstract= {System-on-a-Chip designs with specialized processors and domain-specific accelerators have grown in popularity over the last decade to meet ever-increasing compute demands. This is due to the superior performance-per-watt that can be obtained from such designs as well as the inherent limitations in scaling up traditional out-of-order superscalar CPUs. However, a heavy reliance on microarchitectural details and a lack of coordination between software writers and hardware designers makes targeting these specialized IPs highly non-trivial. In addition, offloading tasks from the CPU to an accelerator introduces overheads that may nullify the advantage of utilizing it in the first place. This paper investigates a transparent method for allowing programmers to run user code on digital signal processing units (DSPs) and quantifies the impact of such an offload.

We propose a lightweight framework at the system software level in the style of message passing Asymmetric Multiprocessing kernels. By pairing a large symmetric multiprocessing kernel with a lightweight real-time operating system, we find that we can achieve low-latency accelerator communication while maintaining a relatively straightforward programming model.},
}

EndNote citation:

%0 Thesis
%A Banister, Christiaan 
%T Low Overhead Remote Procedure Call System for Saturn DSP
%I EECS Department, University of California, Berkeley
%D 2022
%8 May 19
%@ UCB/EECS-2022-144
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-144.html
%F Banister:EECS-2022-144