4/2/24

Issues
• Challenging
– Unpredictable events (like Requests, jobs, peaks etc)
Resource Management – No global state info
– Complex with very large number of resources
– Resources are shared
– Sometimes, no centralized control

1 2

Resource Allocation Aspects Load Balancing

Admission Control

A B C D
•
– Preventing from violations of system policies
– Might require global state information
Capacity Allocation
80% 60% 40% 20%
•
– Allocation of resources for individual instances
– Might require a search in very large search space
Loads
– State of the individual systems changes rapidly !!

A B C D
• Load Balancing
– Local load balancing à optimization is possible to some extent
– Global load balancing à difficult

50% 50% 50% 50%

– Might affect the cost of service provisioning
• Energy Optimization
– Recently, a big challenge
Balanced
– Focus of investment changed !!!
– DVFS
• QoS guarantees
– ???
– Bandwidth Scheduling
– SLA
– Other resources etc

3 4

1
 4/2/24

Energy – Performance Trade-off Load Balancing & Energy Efficiency

Example:
A B C D
Loads 80% 60% 40% 20%
A B C D
Balanced 50% 50% 50% 50%
A B C D
Energy 100% 100% 0% 0%
5 6

Workloads Application & System Goals

• Interactive Vs. Non-interactive • System Owner
• Heterogeneous workloads (hybrid !!) – Maximize Resource Utilization
• Non-interactive • Application Owner
• Interactive – Minimize Makespan for application
– Peak ?? (Spike !!)
– Minimize Makespan for workload
– Minimize Cost

7 8

2
 4/2/24

Resource Management: Trade-off Application Design

• Scaling Map Applications to Resources • Decompose monolithic application into
• Throughput discrete tasks onto different VMs
– Identify degree of parallelism
• Response Time Makespan Decomposition • Divide and Conquer
• Energy • Job Queue Time
– Independent tasks àNo. of processors/VMs
– Identify dependencies and constraints
• QoS Guarantees • Compute Time (CPU) • Order of execution
• Fairness • I/O time (disk) • Decompose input, output and intermediate
• N/W time data onto different VMs
– Operate on different data at the same time
• System overhead

9 10

Points to be considered !!! Other Issues

Issue Scheme Key Problems
• Shared Vs. Independent data Job Priority Non-Preemptive Delay of High-Priority Jobs
– Locks, read locks, and write locks Preemptive Overhead, implementation
• Maximize speed-up Resource Required Static Load imbalance
• Maximize concurrency and reduce overheads Dynamic Overhead, implementation
• Reduce time between results Resource Sharing Dedicated Poor Utilization

• Synchronous Vs. Asynchronous Space Shared Tiling, large job

pipelining Scheduling Time Shared Process-based job control
• Blocking Vs. Non-Blocking with context-switch
• Dependency overhead
– Control Vs. Data Independent Severe slowdown
• Data Locality Gang Scheduling Implementation difficulty
– Spatial Vs. Temporal Competition with foreign Stay Local job slowdown
(local) jobs Migrate Migration threshold,
Migration Overhead

11 12

3
 4/2/24

Resource Management: Trade-off Allocation Techniques

• Control Theory
• Scaling y
eor
– Control theory uses the feedback to guarantee system stability and predict
Th transient behavior
• Throughput tro
l ng
rni
Con a – but can be used only to predict local rather than global behavior
L e
• Response Time ne – Filters are /were designed
achi ase
d
• Machine Learning
• Energy M B
lity ed
Uti r ient – could be applied for coordination of several autonomic system managers
• QoS guarantees t-o – Do not need any performance model
a rke
• Fairness M • Utility-based
– Performance Vs. Cost
– Performance Model is to be designed.
• Optimal Vs. Near-optimal • Market-oriented
– Based on combinatorial Auctions & Bundling of the resources
– No performance Model

13 14

QoS

Penalty levels changes according to SLA

High
Penalty

Average Penalty
High reward

Average reward

15