DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
Ph.D. Thesis Defense
Speaker : Mr. Sheshadri Kalkunte Ramachandra
S.R. Number : 06-18-02-10-12-18-1-16336
Title : “Designing Quality of Service aware Serverless Platforms”
Research Supervisor: Dr. J. Lakshmi
Thesis examiner : Dr. Nitin Auluck, IIT-Ropar
Date & Time : December 04, 2024 (Wednesday) at 10:30 AM
Venue : The Thesis Colloquium will be held on HYBRID Mode
# 102 CDS Seminar Hall /MICROSOFT TEAMS.
Please click on the following link to join the Thesis Colloquium:
MS Teams link
ABSTRACT
Serverless computing is a widely used Cloud computing service offering that provides users with managed runtimes to develop their business logic as functions. It supports event-driven execution of functions and a powerful “pay-as-you-use” billing model that
bills the functions based on the duration of execution for the resources assigned. The service provider through the serverless platform performs application deployment, resource management, provisioning, and system administration tasks. Though this automatic
resource management results in ease of use for the user, the platform internally performs many interactions to orchestrate the function execution, leading to significant overheads. Platform overheads coupled with the prevalence of hardware heterogeneity in
the Cloud and resource specification that is agnostic to application characteristics cause non-determinism in function executions. As a result applications do not have any guarantees from the platform to meet their expectation; and existing serverless platforms
provide no way for the users to express their application requirements to the platform.
Quality of Service(QoS) is a generic mechanism for the application to describe its expectation from the execution in a platform neutral way. The serverless platform can use the QoS specifications to make decisions about resource allocation and provisioning to modulate the application behaviour in accordance to the desired QoS. The challenge in designing for QoS support is its specificity to application domain. Different domains exhibit different requirements and it may not be feasible or realistic to design one QoS model that suits every application. To understand and evaluate the design challenges and impact of QoS on application behaviour on serverless platforms, the thesis identifies three prevalent application domains which give a fair representation of challenges across different application domains that have adopted serverless computing. In the context of these identified applications, the thesis proposes how serverless platforms can be designed to address application requirements while making resource provisioning and deployment decisions.
The first contribution explores the domain of Internet of Things. The applications in this domain are typically deployed across Edge and Cloud layers and have varying image and video data as input; they are triggered based on events. Performance characteristic, response latency is the primary QoS attribute considered for this domain. This work addresses the challenges of deducing resource specifications and performing function deployment in a heterogeneous resource setting. The proposed solution “QoS aware serverless platform for Edge-Cloud continuum” uses application characteristics and QoS requirements to deploy functions across Edge and Cloud with an objective of meeting user-specified QoS requirements. It also exploits data locality in function placement to improve function execution latency. Proposed design is evaluated using representative application scenario, and results analysed for efficacy, improvements and pitfalls.
The second contribution explores the domain of Network Function Virtualization (NFV). NFV allows network functions to be realized as software entities; they are stitched together with other network functions to form a Network Service or a Service Function Chain (SFC), and are deployed on Cloud as an integral part of the infrastructure. SFCs are used to enforce organizational policies on the incoming and outgoing network traffic, typically for a tenant. Network traffic is akin to streaming data; it requires persistent services to have responsive processing of traffic to meet application requirements. In this work, end-to-end latency and ordering are the application requirements. The existing solutions that deploy the SFCs on the Cloud using Infrastructure as a Service (IaaS) or Function as a Service (FaaS) suffer from the problems of poor scalability and high platform overheads, respectively. In order to address these shortcomings, a Hybrid Serverless Platform (HSP) is proposed that comprises the best of both worlds, IaaS and FaaS. The proposed solution uses IaaS for steady state workloads, and FaaS to service the network workload in phases of scaling IaaS or failures of IaaS instances to avoid loss of service. In addition, a succinct way to specify tenant specific requirements to the platform is also discussed. Experiments based on real-world workloads show the benefits of the Hybrid serverless design and QoS-based resource management.
The third contribution focuses on the domain of Inferencing which constitutes an important class of event-driven applications. The applications in this domain are designed as a workflow of functions that process images and texts as input; they are typically deployed across Edge and Cloud layers. Response Latency is the QoS attribute under consideration. This work addresses the problem of platform overheads in enacting inferencing workflows. Existing solutions suggest fusing functions to avoid overheads. However, they result in high execution costs and do not address the case of function fusion across heterogeneous resource layers. To address this, RightFusion, a novel fusion criteria that performs QoS aware Function Fusion across Edge and Cloud is proposed; it uses application characteristics and QoS requirements to perform low-cost function executions that result in significant resource savings compared to existing solutions. Results analysis also discuss the usefulness and shortcomings of the design to bring out its applicability.
The work described in the thesis demonstrates the value of considering QoS requirements in serverless platforms for both the user and provider. It enables the users to convey expected or desired application behavior, thereby bringing in a sense of guarantee during its execution on serverless platforms. It also enhances the platform’s response in consideration of the desired requirements of the application, thereby improving provider credibility. QoS specification also enables smarter resource management in serverless platforms, leading to resource savings for the provider and reduced cost for the user. Though QoS varies from one application domain to another, the requirements can be addressed by bringing in necessary design customization to different components in the existing serverless architecture; these changes are applicable to serverless platforms in general, irrespective of the application domain. While latency was the primary QoS attribute in our study, other QoS features, such as security, privacy, and reliability, present interesting research directions to explore.
ALL ARE WELCOME
