Customized Dynamic Load Balancing for a Network of Workstations

Taken from work done by:

Mohammed Javeed Zaki, Wei Li, Srinivasan Parthasarathy

Computer Science Department, University of Rochester

June 1997

Presenter:Jacqueline Ewell

Static vs. Dynamic Load Balancing

Static Load Balancing:• allows the programmer to delegate work before runtime• can accommodate for heterogeneous processor and non-uniform loops• avoids runtime scheduling overheads• needs to know all information about Workstations ahead of time

Dynamic Load Balancing:• ability to delegate work based on runtime performance of a Network of Workstations (NOW)• transient external loads by multiple-users, heterogeneous processors, memory availability, network bandwidths and contentions, and software leads to a more logical choice of dynamic load balancing

Dynamic Load Balancing Strategies• Task Queue Model: a centralized queue of work

• Diffusion Model: • all work is delegated to each processor, • when an imbalance is detected between it and its neighbor, work is moved

• Predict future performance from past performance: Exchange of performance information

• Global Distributed Scheme• Global Centralized Scheme• Local Distributed Scheme• Local Centralized Scheme

Work queue

Local Global

Distributed

Centralized

Dynamic Load Balancing Strategies

Global - all load balancing is done on a global scale

Centralized - the load balancer is located on one processor

Local - processors are divided into groups (size = K) and load balancing decisions are done within a group

Distributed - the load balancer is replicated on every processor

Dynamic Load Balancing Strategies

Global Centralized

Load Balancer

P1 P2 P3 Pn...

Global Distributed

Local Distributed

P1 P2 P3 Pn...

Load Balancer

Load Balancer

Load Balancer

Load Balancer

Local Centralized

Load Balancer

P1 P2 P3 Pn...P4

G1 G2

Pn...

Load Balancer

P3

Load Balancer

P3

Load Balancer

P1 P2

Load Balancer

Load Balancer

G1

G2

Strategy Tradeoffs

Global vs. Local

• Global information is available at synchronization time; therefore work distribution is optimal• Global scheme - synchronization and communication cost is much higher• Local scheme - groups may sit idle while other groups are overloaded

Centralized vs. Distributed

• Centralized scheme - one load balancer will hurt scalability• Centralized scheme - distribution calculations are on one processor; therefore, done sequentially• Distributed - “all-to-all” exchange of performance profile; therefore, network contention could be a problem

DLB Modeling & Decision ProcessModeling Parameters:

• number of processors• normalized processor speed• number of neighbors

• data size• number of loop iterations• work per iteration• # of bytes to be comm./iteration• time per iteration

• network latency & bandwidth• network topology

• maximum load• duration of persistence of load

ProcessorParameters

ProgramParameters

NetworkParameters

External LoadModeling

DLB Modeling & Decision Process (cont.)

Total DLB Cost: Synchronization Cost + Cost of Calculating New Distribution +Cost of Sending Instructions* +Cost of Data Movement

*only applies to centralized schemes

DLB Modeling & Decision Process (cont.)Synchronization Cost:

• GCDLB: one-to-all(P) + all-to-one(P)• GDDLB: one-to-all(P) + all-to-all(P )• LCDLB: one-to-all(K) + all-to-one(K)• LDDLB: one-to-all(K) + all-to-all(K )

2

2

Cost of Calculating New Distribution: Usually very small

Cost of Sending Instructions: Number of send Messages * Latency

Cost of Data Movement: Number of Message * Latency +Number of Iterations Moved *Number of Bytes that need to be communicated per iteration /Bandwidth

DLB Modeling & Decision Process (cont.)

Initially: work will be divided equally among all processors

Synchronization: 1/Pth work has been done load function is known average effective speed is know

Performance Metric: (number of iteration per second)load function and other parameters are plugged into the model to select the beststrategy

Work Movement: if amount of work to be moved is above a threshold

Profitability Analysis - move work only if there is a 10% improvement in execution time

Experiment• Global Schemes are best ; computation/communication ratio is high • More Processors -> More Synchronization Cost ; favors Local Scheme

• Global is still better at 16-processors

• Centralized master, sequential redistribution, instruction sends, and delay factors add sufficient overheads to Centralized scheme

Experiment• Amount of work/iteration is small; Local Distributed is favored

• As data size increases; Global Distributed does better

• On 16-processors, Local Distributed is the best

• Local is better than Global; since computation/comm. Ratio is small

• Distributed is better than Centralized

Modeling Results

Conclusions• Different Schemes are best for different applications• Customized Dynamic Load Balancing is essential when

transient external loads are introduced• Given the model, it is possible to select a good scheduling

scheme

Future Work• Other Dynamic Load Balancing Schemes need to be incorporate into the model (not lying on the extremes)

• Instead of Local Central, have one master per group• Local schemes, work should be exchanged between different groups• Dynamic Group memberships