Service Clustering for Autonomic Clouds Using Random Forest

Service Clustering for AutonomicClouds Using Random Forest

Rafael Brundo UriarteIMT Lucca

Sotirios Tsaftaris Francesco TiezziIMT Lucca University of Camerino

CCGrid - 7th May 2015 - Shenzhen, China

Contents

1 Introduction

2 Requirements and Existing Solutions

3 RF+PAM

4 Evaluation

5 Conclusions

Uriarte, Tsaftaris and Tiezzi 1/29

Introduction

Introduction Uriarte, Tsaftaris and Tiezzi 2/29

Cloud Computing

I Everything-as-a-Service

I Dynamism

I Heterogeneity

I Virtualization

I Large-Scale

Introduction Uriarte, Tsaftaris and Tiezzi 3/29

Autonomic Computing

Introduction Uriarte, Tsaftaris and Tiezzi 4/29

Autonomic Clouds

I Restricted Knowledge

I Approaches to alleviate the problem:

I Machine LearningI Service Clustering

Introduction Uriarte, Tsaftaris and Tiezzi 5/29

Applications in the Domain

I Anomalous Behaviour Detection

I Service Scheduling

I Application Profiling

I SLA Risk Assessment

Introduction Uriarte, Tsaftaris and Tiezzi 6/29

Requirements and Existing Solutions

Requirements and Existing Solutions Uriarte, Tsaftaris and Tiezzi 7/29

Requirements

Characteristics Requirements

Security, Heterogeneity,Dynamism

Mixed Types ofFeatures

Large-Scale, Dynamism On-line Prediction

Large-Scale, Multi-AgentLoosely-Coupled

Parallelism

HeterogeneityLarge Number of

Features

Requirements and Existing Solutions Uriarte, Tsaftaris and Tiezzi 8/29

Existing Approaches

I Solutions which handle mixed data types usually are notscalable (e.g. HClustream)

I Expert intervention is not feasible due to the dynamism

I Distance Metric Learning Approaches require labelled dataor are computationally expensive.

Requirements and Existing Solutions Uriarte, Tsaftaris and Tiezzi 9/29

RF+PAM

RF+PAM Uriarte, Tsaftaris and Tiezzi 10/29

Random Forest

I Mixed Features

I Large Number of Features

I Efficient and Scales Well

I Easily Parallelizable

RF+PAM Uriarte, Tsaftaris and Tiezzi 11/29

Random Forest

Clustering with Random Forest

I Originally Developed for Classification

I On-Line Random Forest

I Intrinsic Measure of Similarity

I Clustering Algorithm (e.g. PAM)

RF+PAM Uriarte, Tsaftaris and Tiezzi 12/29

Similarity Using RF: Criteria

RF+PAM Uriarte, Tsaftaris and Tiezzi 13/29

Problems

I Similarity Matrix (Big Memory Footprint)

I Re-cluster on Every New Observation

RF+PAM Uriarte, Tsaftaris and Tiezzi 14/29

Solution: RF+PAM

I Off-line Training and On-line Prediction

I Similarity Learning and Standard Clustering

RF+PAM Uriarte, Tsaftaris and Tiezzi 15/29

Solution: RF+PAM

Build Forest, Calculate Similarities, Cluster, Selectthe medoids and Store the references of the leaves.

RF+PAM Uriarte, Tsaftaris and Tiezzi 16/29

Solution: RF+PAM

Parse service and Assign the cluster of the mostsimilar medoid to it.

RF+PAM Uriarte, Tsaftaris and Tiezzi 17/29

Evaluation

Evaluation Uriarte, Tsaftaris and Tiezzi 18/29

Experiments

1. Cluster Quality

2. On-Line Prediction

3. Use Case

Evaluation Uriarte, Tsaftaris and Tiezzi 19/29

Cluster Quality

I Clustering quality compared to 2 otherapproaches (same dataset)

I Better results in all criteria

I Connectivity - Connectedness of the clustersI Dunn Index - Cluster density and SeparationI Silhouette - Confidence in the assignment

Evaluation Uriarte, Tsaftaris and Tiezzi 20/29

On-line Prediction

I On-Line vs Batch Mode

I K-Fold Cross-Validation

I Compared the Adjusted Rand Index (ARI) for 2datasets:

I Monitoring data of Google’s productionclouds - 12500 servers

I Requests of a grid of the Dutch UniversitiesResearch Testbed (DAS-2) - 200 servers

Evaluation Uriarte, Tsaftaris and Tiezzi 21/29

Results: ARI

K Google DAS-2

100 0.81 (0.32) 0.70 (0.23)

50 0.75 (0.19) 0.68 (0.17)

20 0.73 (0.09) 0.67 (0.11)

10 0.70 (0.06) 0.63 (0.09)

5 0.69 (0.05) 0.61 (0.07)

Evaluation Uriarte, Tsaftaris and Tiezzi 22/29

Use Case

I Schedules according to the Dissimilarity

I Similar services separated

I Algorithms:

1. Random2. Dissimilarity3. Isolated

Evaluation Uriarte, Tsaftaris and Tiezzi 23/29

Use Case

I 9 VMs

I Arrival Rates

I Types of Service

I Services’ SLA

Evaluation Uriarte, Tsaftaris and Tiezzi 24/29

Results

Evaluation Uriarte, Tsaftaris and Tiezzi 25/29

Conclusions

Conclusions Uriarte, Tsaftaris and Tiezzi 26/29

Summary

I We propose RF+PAM to alleviate the problemof limited knowledge in AC

I Validated RF+PAM with 3 Experiments

I Scheduling Algorithm

Conclusions Uriarte, Tsaftaris and Tiezzi 27/29

Future Works

I More Use Cases

I Better Implementation

Conclusions Uriarte, Tsaftaris and Tiezzi 28/29

Thank you!Questions?

Rafael Brundo [email protected]

Conclusions Uriarte, Tsaftaris and Tiezzi 29/29

Prune Trees

I Parsing is very fast and efficient

I Prune requires analysis (time consuming)

Conclusions Uriarte, Tsaftaris and Tiezzi 29/29

Retraining

Ratio of predictions/training services (user defined):

I Parallel training

I Trade-off between updating/prediction

Other solutions:

I Dissimilarity to Medoids

I On-line Clustering (Current Limitations andPrediction Speed)

Conclusions Uriarte, Tsaftaris and Tiezzi 29/29