SelfConfiguring Heterogeneous Server Clusters

1
Self-Configuring Heterogeneous Server Clusters

• Taliver Heath, Bruno Diniz, Enrique Carrera,
  Wagner Meira Jr., and Ricardo Bianchini
• (Presented by me)

2
Conserving Energy for Clusters

• Homogeneous clusters
• Solution Leave fewest number of nodes on to
  satisfy requests
• Heterogeneous clusters
• Which nodes?
• Request distributions?

3
Myth of Homogenous Clusters

• Clusters purchased for expandability
• Upgrades
• Blades for power concerns
• Repairs
• Differentiation

4
Heterogeneity in Clusters

• PC
• 800 MHz
• 70W idle
• 94W busy
• SCSI
• RDRAM
• FE links

• Blade
• 1200 MHz
• 40W idle
• First - 150W
• 46W busy
• Laptop IDE
• SDRAM
• GE and FE links

5
Approach

• Model application
• Performance
• Component utilizations
• Component power consumptions
• Optimize metric
• Ratio of power to performance
• Transform application
• Distribute requests, resources
• Dynamically reconfigure

6
Model Overview

• Models throughput by bottleneck analysis
• Models power with linear function
• Solves across all resources simultaneously ?
  non-linear, multi-variable optimization

7
Model Variables
8
Client to Server Distribution

• Vector for distribution of requests

9
Probability of node i using resource r on machine
j
Machine 1 uses own resource
Machine 2 uses resource on machine 3
10
Utilization

• Per resource (r), per machine (i)

11
Solving for Throughput
12
Solving for Power
13
Micro-benchmarking

• For each component we need
• Maximum throughput
• Needs workload characteristics
• Power consumption for a given utilization
• We use several micro-benchmarks combined with
  Least-Squares fitting

14
Using the Models

• Micro-benchmark
• Optimize metric using simulated annealing
• Generate distribution lookup table
• Throughput ? Configuration

15
Instantiation of Models

• Flash-like Web Server
• No cooperation
• Static requests

16
Validation of the Models

• 8 node cluster
• 4 - 800 MHz traditional PC
• 4 1.2 MHz blade server
• Modified distribution to nodes
• 100 Mbps ethernet connections

17
Validation
18
Experiment Setup

• 2 day trace from World Cup 98
• Accelerated by a factor of 20

19
Experiments Conducted

• Energy Oblivious
• All servers on
• Adaptive
• Assumes all machines identical
• Adjusts to match demand
• Model Adaptive
• Recognizes differences in nodes
• Adjusts to optimize metric (P/T)

20
Heterogeneous Oblivious
21
Heterogeneous Adaptive
22
Heterogeneous Model Adaptive
23
Results
24
Why Modeling Wins
25
Future Directions

• Dynamic Content
• CGI distributions
• Automatic corrections for errors
• Errors in machine specification
• Errors in application specification
• Best future node addition
• Different budget constraints
• Component-level control

26
Related Work

• Cluster power and energy Pinheiro01
• Request distribution Carrera01
• Load balancing in heterogeneous systems Zhou93
• Modeling clusters Maggs95
• Resource management in clusters Aron00

27
Conclusions

• Models are accurate
• Server configuration saves more energy
• Minor impact on performance
• Possibly design heterogeneous clusters from the
  beginning

28
Questions?
29
For more information

• www.darklab.rutgers.edu

30
Modeling

• Knowledge required for accurate modeling
• Cluster details
• Application information
• Trace specifics

31
Combining Resources
32
Finding Resource Utilization
Utilization Vector
Non-local Utilization
33
PRESS server

• Cooperative web server
• Locality Conscious
• Developed by Carrera
• For this paper, cooperative nature disabled
• Distribution changes by alerting clients

34
Homogeneous Oblivious
35
Homogeneous Online Adaptive
36
Homogeneous Model Adaptive
37
Modeling for Clusters
38
Dynamic PRESS

• Idea evolves from Pinheiro01
• Algorithm

Periodically 1. Find required throughput of
cluster 2. Find optimal cluster configuration
for throughput 3. Reconfigure cluster