Tools
Tools
Singh, Ankit Kumar (2021) An efficient Serverless Architecture to support fine-grained coordination and state sharing for Stateful Applications. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (1MB) | Preview |
Preview |
PDF (Configuration manual)
Download (801kB) | Preview |
Abstract
In today’s cloud driven work culture, serverless infrastructure is widely adopted due to its pay-as-you-go benefits. In serverless resource allocation is fully managed by the cloud provider and the developer can solely concentrate on application development. To take to benefit of serverless infrastructure many organizations are migrating their work on serverless. A range of applications and domains including image, text, and speech processing have led to machine learning (ML) becoming a widely deployed technology across many IT industries. It is difficult to implement ML algorithms workflows on a serverless platform. There are clearly defined user interactions during various steps in ML workflows, such as data preprocessing, data training, and data fine-tuning. The user may execute this frequently and will require low latency and the provision of resources automatically. A serverless platform must be designed efficiently to support a stateful application workflow. There is a need for distributed shared memory layer to manage synchronization and fine-grained coordination between intermediate shared data. The proposed architecture provides an efficient workflow for stateful application with low network latency and the ability to share the mutable object with fine grained coordination and synchronization. I have deployed the architecture on Amazon Web Services lambda functions. I have also validated the architecture using micro-benchmarks like latency, throughput, and parallelism. Further to state the fine-grained state management in ML application on serverless I have compared its performance with AWS Spark Clusters for k-means clustering algorithm using 30 GB of data generated using spark-pref. According to the results, it was found that the proposed architecture optimizes ML application’s performance on serverless in terms of time.
| Item Type: | Thesis (Masters) |
|---|---|
| Subjects: | Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Cloud computing |
| Divisions: | School of Computing > Master of Science in Cloud Computing |
| Depositing User: | Clara Chan |
| Date Deposited: | 14 Oct 2021 10:28 |
| Last Modified: | 14 Oct 2021 10:28 |
| URI: | https://norma.ncirl.ie/id/eprint/5092 |
Actions (login required)
 |
View Item |