Work

Enabling Quality-of-Service in Emerging Workloads

Year: 2024
Type: dissertation
Abstract: The domain of computer systems has become diverse and complex with the emergence of various technologies. At first, data-centers and cloud computing introduced the notion of resource virtualization and allowed for the emergence of Serverless Computing as the ultimate cloud offering; Then came the rise of deep learning which allowed for the large and complex model training generating workloads at enormous scale, and finally to accommodate such high scale workloads, a Low Latency Low Loss Scalable Throughput (L4S) technology has been created to reduce queuing delays, ensuring low latency to IP flows with a high throughput performance, for the users accessing the aforementioned cloud services. Each of these domains are highly active in research communities with a large amount of work directed towards the application and management of their workloads. Despite all this, only limited work has been done towards ensuring quality-of-service for the clients in these emerging fields. This thesis aims to provide a stable quality-of-service and workload management for the cloud services such as Serverless computing and deep learning, along with the network services which are used for access by users for performance and fairness in multi-tenant environments. In this dissertation, we first introduce our system Sequoia, the first ever platform in the Serverless architecture that ensures quality-of-service for the users. After that, we turn our focus towards machine learning using virtualized resources, intending to improve resource effi- ciency. We develop Nearchus, an adaptive job accelerator for the distributed machine learning jobs which provides efficient resource allocation to improve the training speed for machine learning jobs. Finally, while ensuring quality-of-service on the application layer, the network infrastructure over the internet remains decades old and tends to be lacking for these emerging workloads. We plan to look into traditional networking solutions and new Low Latency Low Loss Scalable Throughput (L4S) technology which has shown promise. There has been limited work towards client fairness in deployment of such technologies in real-world scenarios which has become an important factor for ensuring quality-of-service in multi-tenants environments. Hence, the final part of the dissertation intends to evaluate the effects of multi-bottleneck scenarios and mixing L4S with non-L4S network traffic. The goal is to reduce any unfairness to ensure individual quality-of-service for the users. These improvements will allow the end-users to enjoy better quality-of-experience for their applications by using adequate network controls for access, combined with smart and efficient resource management on the cloud.
Author: Tariq, Ali Raza
Language: English
Open Access status: green
OpenAlex ID: W7009555542
Cites: 179
Cited by: 1
Related to: 10