Title: A Grid-Based Middleware for Processing Distributed Data Streams
1A Grid-Based Middleware for Processing
Distributed Data Streams
- Liang Chen
- Advisor Gagan Agrawal
- Computer Science Engineering
2Roadmap
- Introduction
- Motivation
- Our approach and challenges
- System Overview and Initial Evaluation
- Introduce system architecture and design
- Discuss the self-adaptation function
- Self-Adaptation Algorithm
- Explain the algorithm
- Evaluate the system by using two data mining
applications - Resource Allocation Schemes
- Dynamic Migration
- Motivation
- Light-weight summary structure (LSS)
- How applications utilize the dynamic migration
- Evaluation
- Adaptive Volume Rendering
- Related work
- Conclusion and Future work
3Introduction-Motivation
- What is data steam
- Data stream data arrive continuously
- Enormous volume and must be processed online
- Need to be processed in real-time
- Data sources could be distributed
- Data Stream Applications
- Online network intrusion detection
- Sensor networks
- Network Fault Management system for
telecommunication network elements
4Introduction-Motivation
Network Fault Management System (NFM) analyzing
distributed alarm streams
Switch Network
NFM (Network Fault Management) System
5Introduction-Motivation
- Challenges
- Data and/or computation intensive
- System can be easily overloaded
6Introduction-Motivation
- Possible solutions
- Grid computing technologies
- Automatically adjust processing rate
7Introduction-Motivation
- The needs for processing distributed data streams
- A middleware running in Grid
- Allocate Grid resources
- Provide self-adaptation function
8Introduction-Our Approach
- We implemented a middleware to meet the needs
- Five contributions of our work
- 1. Utilizing existing grid standards
- Liang Chen, K. Reddy and G. Agrawal
GATES A Grid-Based Middleware for Processing
Distributed Data Streams.HPDC, 2004. - 2. Providing self-Adaptation functionality
- Liang Chen and G. Agrawal Supporting
Self-Adaptation in Streaming Data Mining
Applications. IPDPS, 2006. - 3. Supporting automatic resource allocation
- Liang Chen and G. Agrawal A Static
Resource Allocation Framework for Grid-Based
Streaming Applications. Concurrency Computation
Practice and Experience Journal, Volume 18, Issue
6 , Pages 653 - 666. - 4. Supporting efficient dynamic migration
- Liang Chen, Q. Zhu and G. Agrawal A
Supporting Dynamic Migration in Tightly Coupled
Grid Applications. SC 2006. - 5. Studying adaptive rendering application
9Roadmap
- Introduction
- Motivation
- Our approach and challenges
- System Overview and Initial Evaluation
- Introduce system architecture and design
- Discuss the self-adaptation algorithms
- Self-Adaptation Algorithm
- Introduce the algorithm
- Evaluate the system by using two data mining
applications - Resource Allocation Schemes
- Dynamic Migration
- Motivation
- Light-weight summary structure (LSS)
- How applications utilize the dynamic migration
- Evaluation
- Adaptive Volume Rendering
- Related work
- Conclusion and Future work
10System Architecture and Design(Architecture)
- Use Globus Toolkit 3.0, built on OGSA
- Allows users to specify their algorithms
implemented in Java - Take care of plugging user-defined algorithms
into the system and running them in Grid. - Applications need be broken down into a number of
pipelined stages
11System Architecture and Design(Architecture)
Application
Stage A
Stage B
Stage C
12System Architecture and Design(GATES API
Functions)
- Public class Second-Stage implements
StreamProcessing -
-
- void work(buffer in, buffer out)
-
-
- while(true)
-
- DATA GATES.getFromInputBuffer(in)
- Inter-Results Processing(Data)
- GATES.putToOutputBuffer (out,
Inter-Results) -
-
-
13Adaptation Parameter
- Definition
- A parameter in an application
- Changing the parameters value can change
processing rate of the application, also impact
accuracy of the processing - Two kinds of adaptation parameters
- Performance parameter
- Accuracy parameter
- Example
- Sampling rate is an accuracy parameter
14Pseudo Codes Again with Self-adaptation API
Functions
Public class Second-Stage implements
StreamProcessing //Initialize
sampling-rate Sampling-rate (Max Min)/2
void work(buffer in, buffer out)
GATES.specifyAccuracyPara(Sampling-rate, Max,
Min) while(true) DATA
GATES.getFromInputBuffer(in)
Inter-Results Processing(Data,
Sampling-rate) GATES.putToOutputBuffer
(out, Inter-Results) Sampling-rate
GATES.getSuggestedValue()
15Roadmap
- Introduction
- Motivation
- Our approach and challenges
- System Overview and Initial Evaluation
- Introduce system architecture and design
- Discuss the self-adaptation function
- Self-Adaptation Algorithm
- Explain the algorithm
- Evaluate the system by using two data mining
applications - Resource Allocation Schemes
- Dynamic Migration
- Motivation
- Light-weight summary structure (LSS)
- How applications utilize the dynamic migration
- Evaluation
- Adaptive Volume Rendering
- Related work
- Conclusion and Future work
16Self-Adaptation Algorithm
- View the system as a pipeline
- To ensure real-time processing, a balanced
pipeline is needed - When average queue length is too small or too
large, queue is under or over loaded. Pipeline is
not balanced.
- Measure the average lengths of the queues in the
pipeline
- When GATES.getSuggestedValue() is invoked, use
the heuristic way to determine a new value for
the adaptation parameter according to the
measured lengths
17Self-adaptation Algorithm
- The way we measure average queue length
- the heuristic way to adjust an adaptation
parameter - Should the adaptation parameter be modified, and
if so, in which direction? - How to find a new value (update the value) of the
adaptation parameter
18Self-adaptation Algorithm
- Should the adaptation parameter be modified, and
if so, in which direction? - The answer is related to the pipelines load
state.
19Self-adaptation Algorithm
Convergent States
Non-Convergent States
20Self-adaptation Algorithm
Summary of Load States
21Self-adaptation Algorithm
- How to determine a new value for the adaptation
parameter - Linear update increase or decrease by a
fixed value - Hard to find a proper fixed value
- Binary search
22Self-adaptation Algorithm
Left Border
Current Value
Right Border
New Value
23(No Transcript)
24Self-adaptation Algorithm
- Two Data mining applications
- Clustream Clustering data-points in streams
25Data Mining Applications System Evaluation
- Dist-Freq-Counting finding frequent itemsets
from distributed streams
26Data Mining Applications System Evaluation
27Data Mining Applications System Evaluation
28Data Mining Applications System Evaluation
29Data Mining Applications System Evaluation
30Data Mining Applications System Evaluation
31Data Mining Applications System Evaluation
Data Mining Applications System Evaluation
32Data Mining Applications System Evaluation
Data Mining Applications System Evaluation
33Data Mining Applications System Evaluation
Data Mining Applications System Evaluation
34Roadmap
- Introduction
- Motivation
- Our approach and challenges
- System Overview and Initial Evaluation
- Introduce system architecture and design
- Discuss the self-adaptation algorithms
- Self-Adaptation Algorithm
- Explain the algorithm
- Evaluate the system by using two data mining
applications - Resource Allocation Schemes
- Dynamic Migration
- Motivation
- Light-weight summary structure (LSS)
- How applications utilize the dynamic migration
- Evaluation
- Adaptive Volume Rendering
- Related work
- Conclusion and Future work
35Resource Allocation Schemes
- Problem Definition
- Grid resource allocation for pipelined
applications that process distributed streaming
data in real-time is challenging - The scheme consists of two parts
- Static Part allocate resources before an
application runs - Dynamic Part re-allocate resources in run-time
- A framework to monitor resources and support
dynamic resource allocation
36Static Allocation Scheme
- Static allocation problem determining a
deployment configuration - Objective Automatically generate a deployment
configuration according to the information of
available resources
- The number of data sources and their location
- The number of stages consisting of a pipeline
- The number of instances of each stage
- How the instances connect to each other
- The node where each instance is placed
37Roadmap
- Introduction
- Motivation
- Our approach and challenges
- System Overview and Initial Evaluation
- Introduce system architecture and design
- Discuss the self-adaptation algorithms
- Improved Self-Adaptation
- self-adaptation algorithm
- Evaluate the system by using two data mining
applications - Resource Allocation Schemes
- Dynamic Migration
- Motivation
- Light-weight summary structure (LSS)
- How applications utilize the dynamic migration
- Evaluation
- Adaptive Volume Rendering
- Related work
- Conclusion and Future work
38Dynamic Migration-Motivation
- Grid resources vary frequently
- Dynamically allocating new resources and
migrating applications to the new resources
improve performance - Checkpointing is a classic method to support
dynamic migration - A snapshot of systems running state
- Transmit to a remote site
- Restore execution context and restart processes
- Disadvantages of checkpointing
- Platform dependent
- Inefficient
- Involve lots of implementation efforts
- Our approach is base on Light-weight Summary
Structure (LSS)
39Dynamic Migration-LSS
- Processing Structure
- ...
- while(true)
-
- read_data_from_streams()
- process_data()
- accumulate_intermediate_results()
- reset_auxiliary_structures()
-
- ...
- Data structure storing summary information is
Light-weight summary structure - Others are Auxiliary structures
40Dynamic Migration-LSS
- Two observations with respect to LSS
- The size of LSS is much smaller than that of the
total memory - Auxiliary structures are usually reset at the end
of each loop. Unnecessary to migrate auxiliary
structures when migration occurs at the end of a
loop - LSS can be used to support dynamic migration
- GAETS provides an API function to allocate a
block of memory to be LSS - An application stores summary information to LSS
- transmit only LSS at the end of the loop to a new
node and restore the LSS at the new node
41Dynamic Migrationsupported by GATES
42Dynamic Migration
- Advantages of using LSS
- Efficient, only LSS is migrated
- Not impact the accuracy of processing
- Support migration across heterogeneous platforms
- Reduce application developers efforts on making
application capable of migration
43Dynamic Migration
44Dynamic Migration
- Evaluation
- Three applications
- Counting sample
- LSS stores intermediate top M frequently
occurring numbers - Clustream, clustering data points in streams
- LSS stores micro-clusters computed at the second
stage - Dist-Freq-Counting, finding frequent itemsets in
distributed streams. - LSS stores unprocessed itemsets
45Dynamic Migration
46Dynamic Migration
- Migration using LSS is efficient
47Dynamic Migration
- Migration using LSS is efficient
48Dynamic Migration
- Benefits of migration in a dyamic environment
49Dynamic Migration
50Dynamic Migration
- Migration using LSS is efficient
51Dynamic Migration
- Migration using LSS is efficient
52Dynamic Migration
- Benefits of migration in a dynamic environment
53Dynamic Migration
- LSS migration does not impact processing accuracy
- The counting sample application was used
- Compared the average accuracy of the processing
results from the non-migration and the migration
versions, they are 97.28 and 97.51 accurate
54Roadmap
- Introduction
- Motivation
- Our approach and challenges
- System Overview and Initial Evaluation
- Introduce system architecture and design
- Discuss the self-adaptation algorithms
- Self-Adaptation Algorithm
- Explain the algorithm
- Evaluate the system by using two data mining
applications - Resource Allocation Schemes
- Dynamic Migration
- Motivation
- Light-weight summary structure (LSS)
- How applications utilize the dynamic migration
- Evaluation
- Adaptive Volume Rendering
- Related work
- Conclusion and Future work
55Adaptive Volume Rendering
- Motivation Grid computing is needed
- Visualization involves large volumes of dataset
- We focus on streaming volume data
- Interactively visualizing volume data in
real-time is needed - Computationally intensive
- Resources consumed
- Real-time processing can not be guaranteed
- The places where data are generated are
distributed - Typical client-server architecture is not
scalable - Network bandwidths of wide-area networks are low
- Computing capability of normal desktop is not
enough - Grid techniques would be a good solution
- Divide the procedure into stages organized in a
pipeline - Allocate nodes close to data source to
pre-process volume data - The size of intermediate results is much smaller
56Adaptive Volume Rendering
- Motivation GATES is desirable
- Automatic adaptation is desirable
- Volume rendering algorithms running on a grid
need to be highly adaptive - Adaptation usually achieved by manually adjusting
adaptation parameters - Such manual parameter adaptation is very
challenging in a grid environment - Automatic resource allocation is desirable
- Grid environment is highly changeable
- The GATES middleware could fulfill the needs
- Grid-based
- Provide the self-adaptation function to
applications - Automatically allocate Grid resources
57Adaptive Volume Rendering
- Overall design
- Two pipelined steps the first step
- Build octrees from volume data
- Octree is a tree data structure, in which each
internal node has up to 8 children - Here, we use an octree to represent
multiresolution information for a volume - Procedure to build an octree for a volume is as
follows - Divide volume space into 8 subvolumes and create
8 children nodes - For each subvolume, calculate standard deviation
of all voxels in the subvolume, and store the
deviation to the corresponding child node - If the deviation is larger than a pre-defined
value, divide the subvolume, repeat the above
procedure. Otherwise, stop
58Adaptive Volume Rendering
- Overall design
- Two pipelined steps the second step
- Use an octree and its corresponding volume to
render images - Provided an error tolerance (or user-defined
resolution), use DFS to traverse the octree and
stop at the nodes where the deviation is less
than the resolution or error tolerance. - Project the corresponding 3D-subvolumes to an
image
59Adaptive Volume Rendering
60Adaptive Volume Rendering
- Make the rendering self-adaptive
- Two adaptation parameters used in the third stage
- Error Tolerance performance parameter
- Image Size accuracy parameter
- Only one adaptation parameter can be adjusted by
GATES. So we fix one and adjust the other
61(No Transcript)
62Adaptive Volume Rendering
63Adaptive Volume Rendering
150kbps
200kbps
250kbps
64Adaptive Volume Rendering
65Adaptive Volume Rendering
- Experiment 3 compare the performance of two
implementations - Java-imple
- C-imple
66Adaptive Volume Rendering
- Experiment 3 compare the performance of two
implementations
67Related Work
- Middleware for data stream processing
- Data cutter, Stampede
- Differences in a cluster, no self-adaptation, no
specifically for real-time processing - Continuous query systems
- STREAM, dQUOB, TelegraphCQ, NiagraCQ
- Differences centralized, no adaptation supports
- Distributed continuous query systems
- Aurora, Medusa, Borealis
- Differences continuous queries, not in Grid
environment - In-Network aggregation in sensor network
- Stream-based overlay networks
68Related work
- Grid Resource Allocation
- Condor, Realtor, ACDS
- Main Differences our work focus on Grid resource
allocation for workflow applications - Adaptation Through a Middleware
- Cheng et al.s adaptation framework, SWiFT,
Conductor, DART, ROAM - Main Differences our work focus on general
supports for adaptation in run-time - Dynamic Migration in Grid Environment
- Condor, XCATS, Charm
- Main Differences our work use LSS
69Conclusion
- Grid computing could be an effective solution for
distributed data stream processing - GATES
- Distributed processing
- Exploit grid web services
- Self-adaptation to meet the real-time constraints
- Grid resource allocation schemes and dynamic
migration
70Future Work
- CPU cycles and Network bandwidths
- Currently, only network bandwidth is considered a
constraint when scheduling Grid resources - Few related work proposes a metric to integrate
both for pipelined appliations - Port GATES from GT3 to GT4
- Support fault-tolerance and high availability
- Further relieve programming burdens from
application develops - Specify meta-data
- Support distributed continuous queries
- Specify a set of query operators
71Acknowledgements
- My advisor, Prof. Agrawal, proposed the idea of
implementing the middleware, and gave lots
advices for the directions of my research - Prof. Shen gave lots of helps on implementing the
render application, and provided lots of write-up
for the chapter 7
72Questions?
- No more questions? Thanks!