Browse Prior Art Database

Autonomic Performance Tuning System(APTS) - Using Unsupervised Machine Learning Algorithms.

IP.com Disclosure Number: IPCOM000249506D
Publication Date: 2017-Mar-02
Document File: 4 page(s) / 32K

Publishing Venue

The IP.com Prior Art Database

Abstract

In this research publication, the discussion is about architecture for Autonomic Performance Tuning System(APTS) framework which is built on Machine Learning algorithms. The publications also aims to discuss performance enhancement with an example on APTS using data analysis.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 35% of the total text.

1

Autonomic Performance Tuning System (APTS) - Using Unsupervised Machine Learning Algorithms.

In this research publication, the discussion is about architecture for Autonomic Performance Tuning System(APTS). The publications also aims to discuss performance enhancement with an example on APTS using data analysis.

In today's big data world, a variety of applications like database, data analytic, data processing, data mining, and many more run on the system in an on-demand computing and distributed environment. There are variety of machine learning algorithms being developed and used for computer vision, robotics, insights into customer buying's or even for social sentiment analysis during election campaign. These algorithms are capable of handling huge loosely defined data set and helps to drive a meaningful insights into them. However, there is a very little research is done in the industry on how these new-age algorithms could be used to analyze the low level system performance metrics on given workload and provide insights into work load characteristics.

Also now a days, virtualization is heavily used to increase system utilization which make it as a multi-tenant environment where each workload has different performance requirements. In addition workload characteristics of the same workload will change from time to time like a a state machine. So, a static performance tuning of the workload may not give over all improvement in all states. The performance tuning should be specific to each state of workload and should be applied when the workload enters the state and should be removed when it leaves the state. Autonomic Performance Tuning System(APTS) provides a better solution by applying workload specific performance improvements on the fly based on big data analysis which may not be possible for system administrator as the data is too huge for analysis.

Autonomic Performance Tuning System(APTS) uses unsupervised machine learning algorithms to build a state machine based on various characteristics of given workload and generate patterns. It uses system level metrics data(for example a processor level trace on instruction execution and memory access patterns on workload would generate millions system level metrics) to get insights into workload characteristics. Then it would identify possible patterns using algorithms like K-Means clustering and identify performance tunables for those patterns. Then it will apply those tunable when the pattern re-occurs.This logic would be embedded into each subsystem of the Operating System and framework proposed is for utilizing the machine learning algorithms and accessing the knowledge base. This way, subsystem specific performance tuning can be easily achieved.

Consider an example, where applications make plenty of user-land to kernel cross over resulting in system-calls heap. In such an environment applications exchanging data essentially perform many operations and execute millions and millions of co...