Design Trade-off for Cloud Storage

Author:

Shahidi, Narges

Graduate Program:

Computer Science and Engineering

Degree:

Doctor of Philosophy

Document Type:

Dissertation

Date of Defense:

September 01, 2022

Committee Members:

George Kesidis, Major Field Member
Mahmut Kandemir, Chair & Dissertation Advisor
Bhuvan Urgaonkar, Major Field Member
Dinghao Wu, Outside Unit & Field Member
Chitaranjan Das, Program Head/Chair

Keywords:

Cloud
Cloud Storage
Storage Design
Cloud Storage Design
Serverless
SSD
NAND Flash Memories
Serverless Computing
Cloud Computing
Solid State Drives

Abstract:

Cloud services are growing rapidly in all dimensions as they serve as a platform to support the needs of enormous amounts of small businesses, to deploy the complicated enterprise data services. Several Cloud architecture models have been offered recently by the Cloud vendors to allow vast resources available in the Cloud to be utilized in a more programmable way. On the other hand, in order to maximize their benefits, Cloud vendors have the arduous task of optimizing for the allocation of the Cloud resources such as CPU and memory, in the normal and in the peak days. Serverless computing has become popular recently as a highly scalable and efficient way to program the radical growth of the Cloud to allow versatile usage of these resources. With the rapid increase of data intensive computation, the Cloud services are becoming more and more dependent on different layers of storage design. This is adding another dimension to the already complicated and multi-objective optimization problem of the Cloud resources. The objective of this dissertation is to optimize the design trade-off for Cloud storage design for performance and utilization to allow lower cost of ownership and higher manageability. Storage solutions such as Solid State Drives are one of the main resources in the Cloud that needs careful optimizations and planning of high utilization due to the high cost of ownership. In this dissertation, we first discuss the challenges that Cloud storage providers are facing in order to provide a cost efficient, meanwhile viable storage solution. We looked at local storage consolidation solutions and discussed the approaches that Cloud providers can take to reduce the workload interference. We proposed a workload dependent design for the SSD management layer which allows higher utilization of these devices without sacrificing performance isolation. Then in the same area, we looked at novel approaches to increase SSD utilization to the benefit of reducing tail latency, which is one of the main service level objectives for most of the Cloud customers. In the second part of this dissertation, we focused on the serverless computing area and explored the serverless design spectrum and the challenges of data driven applications to fit into this popular model. We provided a comprehensive study of the proposed stateful designs by the serverless vendors. We implemented machine learning and video processing applications with serverless architecture and showed the inefficiency of the current stateful models proposed by major vendors, for these highly used applications. Then, we designed a serverless model based on the Object Oriented programming language to allow easier deployment of the Cloud applications. At the end, we proposed a design architecture for the proposed Cloud model and showed the benefit of the model through several experiments on the real world applications. The model we proposed allows us to leverage storage layers in order to fit a wider range of applications. We discussed how a good programming model can make the code deployment to the Cloud much easier. and at the same time provides flexibility for the Cloud vendors for better optimization of Cloud resources in order to increase the elasticity of the serverless design.