Your browser does not support JavaScript!

Doctoral theses

Current Record: 7 of 114

Back to Results Previous page
Next page
Add to Basket
[Add to Basket]
Identifier 000457103
Title QoS-based resource management for increasing server utilization in the cloud
Alternative Title Διαχείριση πόρων νέφους για την επίτευξη υψηλής χρήσης της υποδομής και καλής απόδοσης των εφαρμογών
Author Σφακιανάκης, Ιωάννης Μ.
Thesis advisor Μπίλας, Άγγελος
Reviewer Κατεβαίνης, Μανώλης
Πρατικάκης, Πολύβιος
Μαγκούτης, Κωνσταντίνος
Αναστασιάδης, Στέργιος
Τσουμάκος, Δημήτριος
Ξύδης, Σωτήριος
Abstract Cloud computing is compelling because it simplifies the management of the infrastructure and allows elastic scaling of resources for applications. As a result, more and more professionals and enterprises rely on the cloud to process and store their information. This increasing demand for cloud resources forces providers to constantly expand their infrastructure and maintain more servers, significantly increasing costs. At the same time, cloud users care for application performance and tend to overprovision applications, which results in a significant portion of the cloud resources being unused. Therefore, providers must handle the tradeoff between application performance and resource utilization to reduce the cost of the infrastructure and, at the same time, keep users satisfied. In this dissertation, we propose techniques for resource management that increase utilization while, at the same time, maintaining application performance above a userdefined level. The system estimates the resources required to achieve a certain level of performance and then places the application appropriately in the infrastructure. First, we design and implement a profile-based approach in a sandboxed environment and create performance models for each application. The system uses these profiles to accurately correlate the allocated resources to application performance and minimize the unused resources of the infrastructure. However, profile-based approaches have limitations: (1) the system cannot handle “unknown” applications and (2) the runs required to generate the profiles can be overwhelming. Next, we address these limitations using a reactive approach, generating performance models on-the-fly during the application execution. The system gradually learns the behavior of each application using a feedback loop controller and constantly improves the estimations about the required resources. Therefore, the reactive approach can handle applications without prior knowledge of their performance profiles and adapt to workload changes. The limitation of reactive approaches is that they cannot handle well workloads with rapid changes in their load. Finally, we propose applying the reactive approach to serverless computing, where we adjust the required resources based on the current changes in the workload. These predictions help the system adjust quickly to sudden changes in the load and maintain the tail latency of the application. We evaluate the proposed system using synthetic workloads that resemble real data center workloads. To achieve that, we develop a methodology that: (1) processes data center traces from major providers, (2) extracts their most important characteristics, (3) scales the workload to match the underlying infrastructure, and (4) executes the workload using well-known cloud applications that match the ones used in the trace. Our results show significant improvement in the utilization of the infrastructure without a visible drop in application performance.
Language English
Subject Cloud computing
Cloud deployment
Resource allocation
Resource assignment
Task scheduling
Workload generation
Ανάθεση πόρων
Ανάπτυξη νέφους
Δημιουργία φόρτου εργασίας
Κατανομή πόρων
Προγραμματισμός εργασιών
Υπολογισμός νέφους
Issue date 2023-07-21
Collection   School/Department--School of Sciences and Engineering--Department of Computer Science--Doctoral theses
  Type of Work--Doctoral theses
Permanent Link https://elocus.lib.uoc.gr//dlib/e/6/9/metadata-dlib-1689066533-706047-30897.tkl Bookmark and Share
Views 789

Digital Documents
No preview available

Download document
View document
Views : 2