Dynamic Load Sharing and Balancing

Dynamic Load Sharing and Balancing

Sig Freund

Outline

Introduction

Distributed vs. Traditional

scheduling

Process Interaction models

Distributed

Systems

Research Studies

Guess about the Future?

Questions

Introduction

Load Sharing

Load Balancing

Static vs. Dynamic

Distributed vs. Traditional scheduling differences

Communication overhead is non-negligible

Effects of underlying hardware cannot be ignored

Dynamic behavior of system must be addressed

Process Interaction models

Precedence process model

Static Process Scheduling

Precedence Process Scheduling

Extended List Scheduling

Earliest Task First

Process Interaction models

Communication process model

Process Interaction models

Disjoint process model

Distributed System Difficulties

Assumption of prior knowledge not realistic

Use disjoint process model since we don’t know how they interact

Simple Heuristic Strategy

Central Controller Process

Maintain queue size of each processor

Process arrive and depart asynchronously

Assignment made to processor with shortest queue

At completion, processor sends queue info

Process Migration

Monitors processors assigned to processes

Schedule is initiated by process departure instead of arrival

Moves processes to achieve fairness

Two common models

Sender-initiated algorithm

Receiver-initiated algorithm

Sender-initiated Algorithm

Process wishing to off-load initiates

Migrates process from heavily loaded processor to lightly loaded

Requires three decisions

Transfer policy

Selection policy

Location policy

Location Policy

Random selection

May cause chain effect of transfers

Probe for candidate

Probe limit

Multicast

Receiver-initiated Algorithm

Lightly loaded processor requests more work

Same three decisions

Transfer policy

Location policy

Selection

policy

Benefits

More stable

Perform better

Research Studies

Dynamic load balancing in distributed computer systems with star

topology[1995]

Performance of dynamic load balancing algorithm on cluster of workstations and

PCs[2002]

Sender-initiated Algorithm

Compare

load balancing policies

for effective

dynamic load balancing on the basis

of decision

strategy for job transfer and

destination decision.

FT

RT

ALBCI

ALBCII

[1995]

Sender-initiated Algorithm

Sender-initiated Algorithm

Hybrid Load Balancing Algorithm

Compare load balancing scheme for

parallel depth

first search on two systems.

6 Sun Workstations running Solaris

10 PC’s running Linux

Round-robin work sharing

Completely distributed system

[2002]

Hybrid Load Balancing Algorithm

Hybrid Load Balancing Algorithm

Current Challenges

High reliance on central node

High load systems require most communication resources

Heterogeneous systems greatly complicate load balancing

Future System[2008]

Board of Directors

Small group of controlling nodes

Number is dynamic

Capable of splitting or combining

Maintain shared historical data for better initial process assignment

Worker nodes

Specialized functions

Reconfigurable (Time slice/Com polling)

References

1.Randy Chow and Theodore

Johnson (1997),DISTRIBUTED OPERATING SYSTEMS AND ALGORITHMS

2.

Xinda, M. I. (2002). PERFORMANCE OF LOAD BALANCING ALGORITHM ON CLUSTER OF WORKSTATIONA AND PCS.

Algorithms and Architectures for Parallel Processing, 2002. Proceedings. Fifth International Conference on

, (pp. 44-47

).

3.

Chong-Gun, L. K.-S. (1995). Dynamic load balancing in distributed computer systems with star topology.

Distributed Computing Systems, 1995., Proceedings of the Fifth IEEE Computer Society Workshop on Future Trends of

, (pp. 514-517).

4. Abraham Silberschatz, Peter Galvin, and Greg Gagne (2005) OPERATING SYSTEM CONCEPTS

5. Sig Freund (2008) FUTURE DISTRIBUTED OPERATING SYSTEM CONCEPT

Questions?