Krishna Gali

1
Optimizing Grid Computing Architecture using
Statistical Techniques for reducing Computation
Time

• Krishna Gali
• Senior Consultant
• Satyam Computer Services Ltd

2
Grid Computing

• A Grid is a network architecture that finds and
  exploits computing resources.
• It virtualizes distributed computing over a
  network of heterogeneous resources, with no
  geographical constraint.
• It based on the idea of parallel processing and
  sharing of resources.
• This technology takes advantage of the unsed
  compouting resources available in the network

3
Grid - Architecture
4
Working Principle
Large Job
Result
Grid Server
Enterprise network 1000s of PCs
5
Application

• Password cracking is a highly time consuming and
  compute intensive process
• Time taken to crack a password depends on
• Length of the password and
• Number of characters in the character set
• Number of possible passwords - Number of
  characters in the character set raised to the
  power represented by the password length

6
Initial Study
Descriptive statistics to measure and baseline
the Cycle Time for one of the Jobs
Outlier identification
7
Box plot

• Observations from Descriptive Statistics
• Data Non-Normal
• Use Individual Distribution Identification to
  study (P AD Values)
• Data follows Weibull Distribution
• Use Brushing Option to remove Outliers from the
  Box Plot
• Very Low number of Outliers
• Go to further Statistical Analysis

8
Fish Bone Diagram
9
Brainstorming
Using Brainstorming and Fish bone diagram,
we identified critical input parameters for output

• Input Parameters (X)
• Polling Interval
• Maximum Work-unit CPU Timeout
• Maximum Work-unit Clock Timeout
• Max CPU Utilization Threshold
• Max Physical RAM Utilization Threshold
• Minimum Disk Space Free
• Number of Nodes
• Concurrency
• Device preferences
• Device group targeting

• Output Parameter (Y)
• Cycle time

10
Steps in DOE

• Define Objective Cycle Time Reduction for PWD
  Cracking
• Select the Response (Y) Cycle Time in Seconds
• Select the factors (Xs) As shown in next sheet
• Choose the factor levels Two levels for all
  factors
• Select the Experimental Design L8 (Taguchi
  array)
• Run Experiment and Collect the data
• Analyze the data
• Conclusions
• Perform a confirmation run

11
Parameters of DOE
12
DOE - L8 Design
13
DOE1 - Results

• Concurrency
• Max. Disk Utilization
• Max Clock Unit
• Polling Interval

14
DOE 2 - Fixing Parameters
Based on DOE1, following parameters were fixed at
this level in DOE2
15
DOE 2
16
DOE2 - Results

• Work Unit CPU time out
• No. of nodes
• Max CPU utilization threshold
• Polling Interval

17
Confirmation Run
18
Achievement

• Baseline status It was taking 21 min 16 sec
  (1276 sec) on 1 node and 14 min 26 sec (866 sec)
  on 10 nodes to run Password Cracking Application
  on Grid.
• Current Status After DOE implementation, the
  time reduced to 4 min 4 sec (244 sec.) with 10
  nodes.
• Improvement With 10 nodes, the improvement is
  75

19
For Queries Please Mail to Krishna_Gali_at_satyam.c
om Amit_S_at_satyam.com