Euro-Par, 2006 1 A Resource Allocation Approach for Supporting Time- Critical Applications in Grid Environments Qian Zhu and Gagan Agrawal Department of Computer Science and Engineering The Ohio State University IPDPS 2009 IPDPS 2009 Conference May 28 th , 2009 Rome, Italy
29
Embed
A Resource Allocation Approach for Supporting Time-Critical Applications in Grid Environments
A Resource Allocation Approach for Supporting Time-Critical Applications in Grid Environments. Qian Zhu and Gagan Agrawal Department of Computer Science and Engineering The Ohio State University. IPDPS 2009 Conference. May 28 th , 2009 Rome, Italy. IPDPS 2009. Context. - PowerPoint PPT Presentation
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Euro-Par, 2006 1
A Resource Allocation Approach for Supporting Time-Critical
Applications in Grid Environments
Qian Zhu and Gagan Agrawal
Department of Computer Science and Engineering
The Ohio State University
IPDPS 2009
IPDPS 2009 ConferenceMay 28th, 2009 Rome, Italy
Euro-Par, 2006
Context
• Ongoing research on supporting time-critical adaptive applications
• Fixed time, flexible computations– Maximize a QoS/Benefit function
• Flexibility: image quality, image size…• Time constraints
IPDPS 2009
Euro-Par, 2006 4
Motivating Application: Great Lake Nowcasting and Forecasting (GLFS)
Model
WeatherData
WaterQuality
20 km2
0 k
m
1 km
1 k
m
• Flexibility – Grid resolution
– Internal time step
– External time step
• Time Constraints
IPDPS 2009
Euro-Par, 2006 5
Summary of Application Needs
• Time-Critical Event Handling– Intense computation and communication– Time and resource constraints– Application-specific flexibility– benefit function
• The CPU/memory usage increases as ErrorTolerance value decreases or the ImageSize value increases.• The change in the value of ErrorTolerance has a more significant impact, compared to the ImageSize parameter.
IPDPS 2009
Euro-Par, 2006
Experimental Study: Great Lake Nowcasting and Forecasting
• The CPU usage changes as the values of ExternalTimeStep and InternalTimeStep vary.• The memory usage remains roughly the same.
IPDPS 2009
Euro-Par, 2006
Problem Description
• Heterogeneous and Dynamic Resources
• Different CPU, Memory, and/or Bandwidth Usage– Different service components– Different values of adjustable service parameters within
the same service component
• Schedule the Service Components to Maximize the Benefit Function Within the Time Constraint
• Each service component is deployed on a single node• Multiple processing round
WSTP TreeConstruction
Service
TemporalTree
ConstructionService
CompressionService
Unit ImageRendering
Service
DecompressionService
ImageComposition
Service
Error tolerance
Image size
Wavelet coefficient
Data Packet
IPDPS 2009
Euro-Par, 2006
Resource Allocation Approach Overview
• Allocate Heterogeneous Resources to Services to Maximize the Benefit Within the Time Constraint– Unique characteristics of resource usage
– Extra resource usage by varying the values of adaptive parameters
• Normal Execution Phase– Train rules for Efficiency Value estimation
– Assign service priority
• Event Handling Phase– Apply the learned rules to infer Efficiency Value
– Priority-based scheduling
IPDPS 2009
Euro-Par, 2006
Efficiency Value
• To capture the suitability of executing the Service on the Processing Node
• Definition– Benefit contribution
, where – Adaptation overhead
, where– Node status
• Weighted sum of standard deviation of the workload and resource variance every 30 seconds
iB
j,iV
j
iSjN
%100)(min
min
i
iicurr
i B
BBB
)(
)( minmin
icurr
icurr
ii
xfB
xfB
%100)(,min
,min
,
,
ji
jijicurr
ji V
VVV
),(
),(
2,
min1,min
icurrjk
jicurr
ijk
ji
xNRuleV
xNRuleV
IPDPS 2009
Euro-Par, 2006
Efficiency Value – Cont’d
01 21 , standard deviation of workload and resource variance
)T,T(e)V
B(sigmoidE j,i
compicomm2
)(
j,i
i1j,i
j
how efficient is for supporting parameter adaptation of for overall benefit optimization
jN iS
• Efficiency value estimation– Fuzzy logic rules
• Calculating Efficiency Value
IPDPS 2009
Euro-Par, 2006 16
Efficiency Value -- Example
IPDPS 2009
CPU=2.0GHzMem=800MB =0.2
CPU=1.2GHzMem=1.0GB =0.4
CPU=2.8GHzMem=2.5GB =0.6
CPU=1.0GHzMem=3.0GB =0.1
CPU=3.0GHzMem=2.0GB =0.05
S1 S2
N1 N2 N3 N4 N5
E1,1 E1,2 E1,3 E1,4 E1,5
0.92 0.46 0.52 0.35 0.72E2,1 E2,2 E2,3 E2,4 E2,5
0.96 0.78 0.56 0.28 0.30
Parameter:x1Priority = 8
Parameter:x2Priority = 4
CPUintensive
Memoryintensive
N1 N2 N3 N4N5 N1
N2N3
N4 N50%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
No
rmal
ized
Ben
efit
S1S2
Figure: Example of Efficiency Value Calculation: (a) Computed Values (b) Normalized Benefit with Different Allocations
• Assigning to and to yields the maximum benefit• Our definition of efficiency value captures the suitability of different nodes for different services
1S 2S1N 2N
Euro-Par, 2006
Scheduling Algorithm
• Greedy Scheduling– Service priority based
• Benefit Optimization and Meeting the Time Deadline– Adjust and 1 2