DGSF: Disaggregated GPUs for serverless functions

Transparently enable serverless functions to use GPUs through CUDA APIs.

Meta Info

Presented in IPDPS 2022.

Understanding the paper

TL;DR

This paper presents DGSF, a platform that transparently enables serverless functions to use GPUs through CUDA APIs.

GPU disaggregation challenges

• Support API transparency.

• Hide the latency of communication with remote GPUs.

• Load-balancing access to heavily shared GPUs.

Implementation

• The prototype uses OpenFaaS v0.21.1 as serverless platform.

• Assume all functions are invoked without a cold start.

Evaluation

• The cost of API remoting?

• The impact of DGSF's optimization?

• Utilization increase and performance gains when consolidating functions?

• The overhead of migration?