Managed by

Self-managing Grid Services for Efficient Data Management

Anastasios Gounaris

(University of Cyprus, University of Manchester)

CoreGRID Summer School

Budapest, September 2007

2

Web/Grid Services play an increasingly important role in large-scale data management applications

E.g., SkyQuery (figure taken from [2])

Or more generic WSMSs that compile, optimize and evaluate workflows composed of multiple calls to remote WSs in order to analyze data [1]

[1] U. Srivastava, K. Munagala, J. Widom, and R. Motwani.Query optimization over web services. In VLDB, pages 355–366, 2006.[2] Tanu Malik et al. SkyQuery: A Web Service Approach to Federate Databases. In CIDR 2003

3

At the middleware level: OGSA-DAI in brief

• Different types of data resources - including relational, XML and files - can be exposed via web services onto Grids.

• A number of popular data resource products are supported. • Through OGSA-DAI, data resources can become grid-

enabled in a clean, non-intrusive, straightforward way.

• “Perform” documents can describe/include data creation, query, modification and delivery tasks

For more information, visit the OGSA-DAI website: http://www.ogsadai.org.uk

4

Services For Accessing Multiple Grid Data Sources and Computational Resources

…I’d like to find similar proteins

to those proteins that have

interacted with any other protein

in more than X cases...

ProteinID1 ProteinID2

…. ….

ProteinID Sequence

…. ….

ProteinID Sequence

…. ….

Subquery2

Subquery1

BLAST

5

OGSA-DQP in brief

• Provides services that enable nodes to participate in the evaluation of a query

• Additionally, it provides a coordinator service. • Presents to the user a unified view of the available (computational,

analysis and OGSA-DAI) resources in a location-transparent way.

OGSA-DQP ethos:• Light-weight, loosely-coupled integration of data and computational

resources.• Wrapper-mediator plus:

– Allocation of query sub-plans to arbitrary grid nodes.– Partitioned and pipelined parallelism.

• Service-based in that:– Queries run over services.– Query processor is itself implemented as a collection of services.

For more information, visit the OGSA-DQP website: http://www.ogsadai.org.uk/dqp

6

Services For Semantic Integration

…I’d like to find authors of papers from the database

community…

Author Field

…. Database

Professor School

…. Database

Researcher Topic

…. Database

- Extend OGSA-DQP with semantic integration capabilities

Subquery3

Subquery2

Subquery1

[3] Gounaris, Comito, Sakellariou and Talia. A Service-Oriented System to Support Data Integration on Data Grids. CCGrid 2007

Semantic Map

Topic↔Field ↔SchoolAuthor ↔ Researcher↔Professor

7

BUT, there are some problems

- continuing the example of slide 4

- IF, machines have the same capacity, connection speed and load, execution time will be t

- BUT, IF the load of one machine doubles, then the execution time will be 2* t, whereas the optimum is 4/3 * t

- ALSO, IF a third machine becomes available, we must be able to employ it

- THUS, we need ADAPTIVE, SELF-MANAGING SOLUTIONS

BLAST1

BLAST2

0.5

0.5

8

- Concave-like performance graphs are very common

- BUT, finding the optimal point is challenging, since the volatility of the environment and the noise insert local optima and non-linearities

- In addition, we may have functions of the form F(x,y,z,...) instead of F(x)

Service tuning is non-trivial

F(x)

x

9

Outline of (the rest of) this talk

- Component-based adaptivity solutions - Monitoring- Architecture- An example from OGSA-DQP

- Control theoretical approaches

10

Monitoring with external component…

… such as R-GMA:+ RDBMS technology

allows for expressive queries

- Centralized- Cannot cope well with

rapidly evolving properties

- Cannot cover application specific metric (e.g., per tuple join cost, selectivity of an operator)

R-GMA

11

Self-monitoring

- Can monitor both at machine level (e.g., current load, memory available, etc.) and at application level (e.g., per tuple cost of subquery, selectivities, etc.)

- No unnecessary information dissemination - … which results in significantly low overheads

[4] Anastasios Gounaris, Norman W. Paton, Alvaro A. A. Fernandes, Rizos Sakellariou: "Self-monitoring Query Execution for Adaptive Query Processing". Data Knowl. Eng. 51(3): 325-348 (2004)

12

Going one step further: Self-managing

According to [5], an adaptive, self-managing element is as follows:

[5] Jeffrey O. Kephart, David M. Chess: The Vision of Autonomic Computing. IEEE Computer 36(1): 41-50 (2003).For more information visit: http://www-306.ibm.com/autonomic

Main advantages:• components can be re-used• easier combinations of different adaptivity flavors• clean engineering• easier configuration

13

A Corresponding Adaptivity Framework for Grid Query Services (extension to OGSA-DQP)

Qu

ery

Evalu

ato

r

query

Monitoring

Response

Assessment

Resource

pool

Resource

pool

Executioninfo

Resource info

Monitoring events

Issues with current executionPrompt foradaptation

Monitoring ComponentMonitoring Component

Assessment ComponentAssessment Component

Response ComponentResponse Component

Adaptivity phases Adaptivity components

WS boundaries

[6] Gounaris, Paton, Sakellariou, Fernandes, Smith, Watson: "Modular Adaptive Query Processing for Service-Based Grids ". Proc. of the 3rd IEEE International Conference on Autonomic Computing, 2006[7] A Gounaris Resource Aware Query Processing on the Grid, PhD thesis, 2005

14

Additional info

• How can these inter-operate?– Through publish subscribe

• What is the exact role of the adaptivity components– Depends on the adaptivity scenario that is to

be supported

15

A dynamic balancing example

table_scan(protein)

exchange

op_call(Blast)

project

Site C

Site A, B

Evaluator

Detector

Site A Site B

Coordinator

Evaluator

Detector

Diagnoser

Responder

1

subscribe

The Detector, Diagnoser and Responder are specific Instantiations of the Monitoring, Assessment and Response Components, respectively

16

A dynamic balancing example

table_scan(protein)

exchange

op_call(Blast)

project

Site C

Site A, B

Evaluator

Detector

Site A Site B

Coordinator

Evaluator

Detector

Diagnoser

Responder

2

Evaluators pass on monitoring info to detectors

17

A dynamic balancing example

table_scan(protein)

exchange

op_call(Blast)

project

Site C

Site A, B

Evaluator

Detector

Site A Site B

Coordinator

Evaluator

Detector

Diagnoser

Responder

3

Detectorsnotify the diagnoserof the actual behaviour

18

A dynamic balancing example

table_scan(protein)

exchange

op_call(Blast)

project

Site C

Site A, B

Evaluator

Detector

Site A Site B

Coordinator

Evaluator

Detector

Diagnoser

Responder

4

Diagnosernotifiesresponder

19

A dynamic balancing example

table_scan(protein)

exchange

op_call(Blast)

project

Site C

Site A, B

Evaluator

Detector

Site A Site B

Coordinator

Evaluator

Detector

Diagnoser

Responder

5

Responderchecks progress

20

A dynamic balancing example

table_scan(protein)

exchange

op_call(Blast)

project

Site C

Site A, B

Evaluator

DetectorSite C

Coordinator

Diagnoser

Responder

6Responderenforcesredistribution

21

Categories of Monitoring

M1: Notifications on the cost of processing a tuple in a subquery (tuple processing cost, selectivity, time waiting for input).

M2: Notifications about communication costs (cost of sending a buffer, recipient of buffer, size of buffer).

22

Monitoring: Parameters

M1 notifications:– Events produced by evaluator every 10 tuples; – Averages are calculated for last 25 events;– Threshold for sending notification is a 20%

change in cost per tuple.

M2 notifications are for every buffer.

23

Assessment

• Based o the monitoring info, calculate the throughput for each subquery.

• The assessment component seeks to identify where there is a workload imbalance.– Existing Balance Vector W = (w1, …, wn)– Proposed Balance Vector W’ = (w’1, …, w’n)

• Notify response component if:– (| wi- w’i|) / wi) > threshold (20%) for some i.

24

Response

• If evaluation not close to completion and If progress since last adaptation > x% then apply new distribution vector W’

• response can be retrospective: resends tuples to consumer based on W’.

• or prospective: no tuples are resent, but future tuples are sent to consumers based on W’.

25

Such a framework is general

E.g.,

- by extending the RC, it can adapt in a different way (e.g., by replacing a slow machine), while re-using the same MC and AC.

- By extending the AC, it can solve a different problem (e.g., checking also for vertical imbalances)

- By deploying a hierarchy of components, problems can be solved locally and at different levels in order to attain scalability

[8] Anastasios Gounaris, Jim Smith, Norman W. Paton, Rizos Sakellariou, Alvaro A.A. Fernandes, Paul Watson: "Adapting to Changing Resource Performance in Grid Query Processing". Proc. of the 1st International Workshop on Data Management in Grids, DMG'05 (held in conjunction with VLDB 2005).

26

Outline of (the rest of) this talk

- Component-based adaptivity solutions - Monitoring- Architecture- An example from OGSA-DQP

- Control theoretical approaches- Runtime optimization (hill climbing)- Extremum control- System identification

27

A real problem

•We want to retrieve a large dataset from a WS•It is widely accepted that due to performance limitations of XML parsers, SOAP, etc., we must split the data into several chunks•The performance is shown in the figure

0

50000

100000

150000

200000

0 2000 4000 6000 8000 10000 12000 14000

block size

resp

on

se t

ime

28

Main observations

The optimal block size depends on– The actual connection– The data itself (avg. tuple size)

Main challenges:– No model– Significant noise

• Local minima

– Requirement for fast convergence

29

Runtime optimization

Newton-based approach.• Let x be the block size and y=f(x) be the transfer time.• One chunk transfer corresponds to one adaptivity step

• Performance: LOW! (Such a technique is known to be sensitive to noise, and works better for quadratic models)

30

On extremum control

Objective: find the (optimal) value of the control variable(s) that yields the maximum (minimum) value of a process output (or performance measure) in the presence of noise.

y = f(x)

Extremum controller

ykxk

31

Switching extremum control

Let y be the performance metric and xk the block size at the kth step. Then

xk= xk-1 – gain* sign(Δx*Δy)Rationale:

– detect the side of the optimum point where the current block size resides on.

Intuition behind: – the next block size must be greater than the previous one, if,

in the last step, an increase has lead to performance improvement, or a decrease has lead to performance degradation.

– Otherwise, the block size must become smaller.

32

Extensions

• Linear models are not suitable, thus it is better the gain to be adaptive

gain = b* Δx*Δy/y• The impact of noise must be mitigated. To this

end, the values of x,y are the average over a window

• To enable continuous search of the block size space, add a dither signal

33

Evaluation

• We used 5 queries over data from the TPC-H benchmark

34

Results

Starting point: 1000 tuples, b = 15

The performance degradation drops by an order of magnitude or less

35

Intra-query behavior

•The technique is characterized by stability and quick convergence

•However, in some cases, overshooting is avoided only by imposing hard upper and lower limits.

•Also, the oscillation in individual runs are rather significant

36

System identification

Choose a model, e.g.,

Or,

… and apply (Recursive)

Least Squares.

Then, optimal point can

be found analyticallyWhich model(s) to choose in each case, remains an open issue!

37

Summary Remarks

- Component-based adaptivity solutions- Enable a wide range of adaptive solutions- Decentralized adaptations- Services are self-monitoring- In line with the IBM autonomic toolbox

- Control theoretical approaches- seem very promising for self-optimization

when a performance metric must be optimized on the fly

- In the presence of noise …- … and absence of a model

38

Thank you!

Questions?