Window-based Parallel Operator Execution with In-Network Computing
| Autor: | Bochra Boughzala, Christoph Gärtner, Boris Koldehofe |
|---|---|
| Přispěvatelé: | Distributed Systems |
| Jazyk: | angličtina |
| Rok vydání: | 2022 |
| Předmět: | |
| Zdroj: | Proceedings of the 16th ACM International Conference on Distributed and Event-based Systems (DEBS '22), 91-96 STARTPAGE=91;ENDPAGE=96;TITLE=Proceedings of the 16th ACM International Conference on Distributed and Event-based Systems (DEBS '22) |
| Popis: | Data parallel processing is a key concept to increase the scalability and elasticity in event streaming systems. Often data parallelism is accomplished in a splitter-merger architecture where the splitter divides incoming streams into partitions and forwards them to parallel operator instances. The splitter performance is a limiting factor to the system throughput and the parallelization degree.This work studies how to leverage novel methods of in-network computing to accelerate the splitter functionality by implementing it as an in-network function. While dedicated hardware for in-network computing has a high potential to enhance the splitter performance, in-network programming models like the P4 language are also highly limited in their expressiveness to support corresponding parallelization models. We propose P4 Splitter Switch (P4SS) which supports overlapping and non-overlapping count-based windows for multiple independent data streams and parallelizes them to a dynamically configurable number of operator instances. We validate in the context of a prototypical implementation our splitting strategy and its scalability in terms of switch resource consumption. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|