Towards efficient processing of latency-sensitive serverless DAGs at the edge

Xiaosu Lyu, Ludmila Cherkasova, Robert Aitken, Gabriel Parmer, Timothy Wood
2022 Proceedings of the 5th International Workshop on Edge Systems, Analytics and Networking  
Many emerging novel applications expect "near real-time" processing and responses, which can not be guaranteed by today's Cloud and would require processing at the Edge. Serverless computing is a particularly promising architecture for edge environments since it offers to improve efficiency by precisely scaling resources to meet application needs. As the edge applications grow more complex and get composed from a subset of simpler functions or microservices, there is a need to support more
more » ... icated function topologies which can be represented as directed acyclic graphs (DAGs). However, running DAG functions on a serverless platform poses new challenges related to interconnecting, instantiating, and scheduling function sandboxes. In this paper 1 , we explore how Sledge, a Wasm-based serverless runtime, can be extended to support DAG functions. Sledge's unique design allows for extremely lightweight sandbox instantiation -a new sandbox can be started for each function invocation in under 30𝜇sec -which mitigates the cold start problems that can be especially detrimental to DAGs. Rather than relying on expensive coordination via shared storage, the enhanced Sledge framework provides a fast memory communication channel to propagate data through the DAG. We consider the DAGs with service level objectives, defined by their execution deadlines. To ensure the DAGs meet their performance requirements, we consider, analyze, and compare two deadline-aware pluggable schedulers (that we implemented in Sledge) on a variety of realistic workloads. CCS CONCEPTS • Computing methodologies → Distributed computing methodologies; • Computer systems organization → Real-time system architecture.
doi:10.1145/3517206.3526274 fatcat:45p6zzdng5bspeea2h4dxvpcfi