Adaptive monitoring and control framework in Application Service Management environment

Sikora, Tomasz Dawid (2023) Adaptive monitoring and control framework in Application Service Management environment. PhD thesis, Birkbeck, University of London.

Abstract

The economics of data centres and cloud computing services have pushed hardware and software requirements to the limits, leaving only very small performance overhead before systems get into saturation. For Application Service Management–ASM, this carries the growing risk of impacting the execution times of various processes. In order to deliver a stable service at times of great demand for computational power, enterprise data centres and cloud providers must implement fast and robust control mechanisms that are capable of adapting to changing operating conditions while satisfying service–level agreements. In ASM practice, there are normally two methods for dealing with increased load, namely increasing computational power or releasing load. The first approach typically involves allocating additional machines, which must be available, waiting idle, to deal with high demand situations. The second approach is implemented by terminating incoming actions that are less important to new activity demand patterns, throttling, or rescheduling jobs. Although most modern cloud platforms, or operating systems, do not allow adaptive/automatic termination of processes, tasks or actions, it is administrators’ common practice to manually end, or stop, tasks or actions at any level of the system, such as at the level of a node, function, or process, or kill a long session that is executing on a database server. In this context, adaptive control of actions termination remains a significantly underutilised subject of Application Service Management and deserves further consideration. For example, this approach may be eminently suitable for systems with harsh execution time Service Level Agreements, such as real–time systems, or systems running under conditions of hard pressure on power supplies, systems running under variable priority, or constraints set up by the green computing paradigm. Along this line of work, the thesis investigates the potential of dimension relevance and metrics signals decomposition as methods that would enable more efficient action termination. These methods are integrated in adaptive control emulators and actuators powered by neural networks that are used to adjust the operation of the system to better conditions in environments with established goals seen from both system performance and economics perspectives. The behaviour of the proposed control framework is evaluated using complex load and service agreements scenarios of systems compatible with the requirements of on–premises, elastic compute cloud deployments, server–less computing, and micro–services architectures.

Metadata

Statistics

DownloadsShow export options

Export as XMLJSONCSV

Activity Overview

6 month trend

67Downloads

6 month trend

85Hits

Additional statistics are available via IRStats2.

Archive Staff Only (login required)

Edit/View Item