PyPlover: A System for GPU-enabled Serverless Instances

Ryan Yang, Nathan Pemberton and Jichan Chung

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2020-74
May 28, 2020

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-74.pdf

Demand for GPUs has grown exponentially since the onset of machine learning workloads. However, the cost of an efficient GPU remains very high. For a machine without a GPU, one solution is to send the GPU workload to a dedicated cluster of GPU-enabled instances for processing. However, without the proper knowledge, this method turns out to be very inefficient due to improper load balancing and instance tuning. We propose PyPlover, a serverless GPU framework that allows the user to send kernels and inputs to a serverless provider without needing to worry about set-up costs and load balancing.

Advisor: Randy H. Katz


BibTeX citation:

@mastersthesis{Yang:EECS-2020-74,
    Author = {Yang, Ryan and Pemberton, Nathan and Chung, Jichan},
    Editor = {Katz, Randy H. and Gonzalez, Joseph},
    Title = {PyPlover: A System for GPU-enabled Serverless Instances},
    School = {EECS Department, University of California, Berkeley},
    Year = {2020},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-74.html},
    Number = {UCB/EECS-2020-74},
    Abstract = {Demand for GPUs has grown exponentially since the onset of machine learning workloads. However, the cost of an efficient GPU remains very high. For a machine without a GPU, one solution is to send the GPU workload to a dedicated cluster of GPU-enabled instances for processing. However, without the proper knowledge, this method turns out to be very inefficient due to improper load balancing and instance tuning. We propose PyPlover, a serverless GPU framework that allows the user to send kernels and inputs to a serverless provider  without needing to worry about set-up costs and load balancing.}
}

EndNote citation:

%0 Thesis
%A Yang, Ryan
%A Pemberton, Nathan
%A Chung, Jichan
%E Katz, Randy H.
%E Gonzalez, Joseph
%T PyPlover: A System for GPU-enabled Serverless Instances
%I EECS Department, University of California, Berkeley
%D 2020
%8 May 28
%@ UCB/EECS-2020-74
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-74.html
%F Yang:EECS-2020-74