Grid computing performance prediction based in historical information

1
Grid computing performance prediction based in
historical information
WP6 Resource Management and Scheduling Task 8
Performance Prediction

• Francesc Guim
• Computer Architecture Department
• Universitat Politècnica de Catalunya
• fguim_at_ac.upc.edu
• Ariel Goyeneche
• Centre for Parallel Computers
• University of Westminster
• goyenea_at_wmin.ac.uk
• Coregrid Integration Workshop, Pisa 28th-30th
  November 2005.
• http//www.coregrid.net

2
Performance in Grid Computing

• Grid Computing Coordinated resource sharing and
  problem solving in dynamic, multi-institutional
  virtual organizations (Sharing is not primarily
  file exchange, but rather access to computers,
  software, data, and other resources, as is
  required by a range of collaborative problem)
• Problems in mapping traditional performance
  analysis to Grid computing
• Machine specific
• Source code
• Extremely slow and memory intensive restrictions
• Non-dynamic analysis for supporting heterogeneous
  and not reliable environments
• Social aspects (Time, days, Holidays, etc)
• Intrusive and security aspects
• Proposed solution
• We suggest that the historical data related to
  applications, resources and users can provide an
  adequate amount of information for modeling and
  predicting Grid components behaviors.

3
Challenges to get through 1- Models built from
data semantically diverse

• Problems
• Internal structure and query interfaces of
  different VO and center are different.
• The provided data is semantically diverse and
  heterogeneous.
• Solutions
• Use a query-centric system for knowledge
  acquisition from distributed semantically
  heterogeneous data sources (that may employs
  ontologies (See INDUS)) to provide a collection
  of such data sources as though they were a
  collection of tables structured according to
  those rules supplied.
• Issues to tread
• Definition of such set of types, properties, and
  relationship types.

4
Challenges to get through 2- Data learning
models reflecting stochastic values

• Problems
• If the solution uses conventional point-valued
  performance parameters and prediction models the
  solution may be inaccurate since they can only
  represent one point in a range of possible
  behaviours .
• Solutions
• Try to characterize application and system
  through a set of possible values and their
  probabilities (stochastic values ) (See
  Performance prediction in production
  environments)
• Issues to tread
• Calculation of stochastic values corrected from
  comparison prediction performance vs. real
  measurements , is there any balance to find the
  best probability distribution for a given value?

5
Challenges to get through 3- Prediction
mechanisms

• Problems
• How to use the normalized and pondered data to
  predict performance behaviours?.
• Solutions
• Intuition and previous works (see ref in paper)
  indicates that similar applications are more
  likely to have similar run times than application
  that do not have nothing in common
• Issues to tread
• How do we define similar?
• Hints
• Define good templates for a particular work-loads
  (ANL)
• Transforming those templates into more
  sophisticated dynamic services and resource
  models
• Add potential social aspect .

6
The big picture
VO1
GS
GS
GS
GS
GS
GS
GS
GS
GS
GS
VO2
GS
GS
GS
GS
GS
GS
GS
VO3
GS
GS
GS
GS
GS
GS
GS
GS
7
First approach

• We focused on the three challenges explained
  before in the simplest scenario ? one VO
• We studied the workloads of the UPC system
• It was already unified (challenge 1)
• We tried to represent the workload variables with
  statistic descriptors and study their
  characteristics (challenge 2)
• We implement a simple historical predictors based
  on the previous work (challenge 3)
• Using statistical estimators
• Mean and standard deviation, Median and
  Interquartile difference and trimmed mean
• Using last value.

8
Our predictor example

• The input predictor variables are
• User, group, number of requested processes,
  submitted executable.
• The output prediction variables are
• Total time, user time, system time, memory used
• Predictors base their predictions on
• The historical information for the given user,
  executable and number of tasks

9
Result

• We obtained good results
• 35 of all the applications can be predicted
  90-100 of their executions.
• 30 of all the applications can be predicted
  40-60 of their executions.
• The rest of applications
• Do not have clear patterns neither relations.
• However predictors have to be still improved

10
Next steps

• Studying information coming from
• National Grid Service (NGS) UK
• UPC
• In order to solve problems related to
• Heterogeneous internal structure and semantically
  diverse
• Identifying relevant suitable workload variables
  to be used in prediction.
• New predictor based on previous information.

11
Questions?

• You can contact us on
• Francesc Guim, fguim_at_ac.upc.edu,
  http//francesc.guim.net
• Ariel Goyeneche, goyenea_at_wmin.ac.uk,
  http//www.cscs.wmin.ac.uk/goyenea/.