Characterizing Object Stores For Serverless Systems
Open Access
- Author:
- Satish, Chirag
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- November 02, 2020
- Committee Members:
- Aasheesh Kolli, Thesis Advisor/Co-Advisor
Wang-Chien Lee, Thesis Advisor/Co-Advisor
Chitaranjan Das, Program Head/Chair - Keywords:
- Serverless systems
Object Stores
Characterization - Abstract:
- Serverless systems, a new cloud offering, enables users to deploy applications by just writing code, with the cloud provider performing the deployment, storage, authentication, scaling, security and billing by provisioning a large number of auxiliary backend services. With the increased adoption of these systems, it is critical that these platforms can support a wide range of applications and provide performance guarantees. Due to a large number of interacting backend components, these systems often show significant divergence in performance not only from conventional cloud deployments but also from the cloud providers' own expectations. To this end, it is vital to understand the holistic system performance-cost trade-offs for cloud providers, for any serverless research/effort. In this work, we have focused primarily on object stores and the properties they exhibit under different configurations, to understand what works best for serverless systems. We have compiled a serverless workload set and deployed it for characterization on a serverless testbed built from tuned open-source production-grade tools. We find that the choice of the backing file system and traffic to a serverless application can impact overall latency. We also find that close physical placement of serverless functions can yield a significant improvement in latency. Finally, we observe that right-sizing of object stores is key to stable performance with increasing load. We believe these findings will help cloud providers better schedule serverless functions using object stores and provision backing file-systems appropriately for better serverless performance.