Resource Allocation in Cloud

By Geetha Priya Balasubramanian

Overview Introduction

Challenges

Mechanisms for allocation

Concept I

Concept II

Conclusion

Introduction Cloud computing

◦Complex system◦Shared resources

Why resource management is important?

Challenges

Lean allocation of resources

Unpredictable requests

Shared resources

Resource usage – time variant◦Dynamic availability of resources

Mechanisms for allocation

Allocation techniques Control theory

Feedback mechanism to guarantee system stability, predict transient behavior

Machine learning No performance model

Utility-based Performance model to correlate user-

level performance with cost

DYNAMIC RESOURCE ALLOCATION IN COMPUTING CLOUDS THROUGH DISTRIBUTED MULTIPLE CRITERIA DECISIONANALYSISYa gız Onat Yazır, Chris Matthews, Roozbeh FarahbodStephen Neville, Adel Guitouni, Sudhakar Ganti and Yvonne Coady.

ConceptCentralized global arbiter Two level architecture

◦Application agents mapping between performance level and

resource level requirements per application environment

◦Node agents Configuration changes in the resource

requirements Local re-distribution of the resources Moving suitable components to other

computational units

Node program and Task life cycle

PROMETHEE Method

Results and conclusion

FROM DATA CENTER RESOURCE ALLOCATION TO CONTROL THEORY AND BACK

Xavier Dutreilh, Nicolas Rivierre, Aur´elien Moreau, Jacques Malenfant and Isis Truck

ConceptResource allocation and policies

◦threshold-based policies, where upper and lower bounds on the performance trigger adaptations, where some amount of resources are allocated or deallocated

◦sequential decision policies based on Markovian decision processes (MDP) models and computed using, for example, reinforcement learning.

Experiment

ObservationsResource allocation as automatic control1. Measure patterns of evolution of the

performance against time since the start of the adaptation action

2. Adaptations at a faster rhythm than the time required to stabilize the performance of the system leads to instability

3. Analyze the workload patterns of variation to so that

1. New stabilization performance has small cope up time

2. Maximum performance with adaption actions

ObservationsFinding good resource allocation policies1) Adaptation actions, measure the new

stabilized performance after crossing lower threshold, and make sure the upper (resp. lower) threshold is strictly less than (resp. greater than) this measure.

2) From the maximal time ts to stabilize the performance after any adaptation action, compute the difference between the two thresholds so that the time to pass from one threshold to another is larger than t s even for the maximal slope in the workload variation.