Yuan Yuan

1
Parallelisation of IBD computation for
determining genetic disease map

• Yuan Yuan
• CMSC 838 Presentation

2
Introduction

• Parallel Genehunter package
• In high level
• Provide dynamic allocation mechanism
• Reduce Space and CPU Time consumption

3
Reminder

• Genehunter
• Software package
• Construction of human genetic maps
• Generate inherited diseases maps
• Gene therapy
• IBD (Identity by Descent)
• One function in Genehunter
• Identify the identical locus transmitted by a
  common ancestor

4
Reminder

• Markers
• A number to quantify the genetic resemblance
• of two relatives affected by the same genetic
  disease
• Non-founder
• Individual whose parents are in the family
• Genetic disease maps
• Multipoint linkage analysis with many different
  markers
• Account for all the family information

5
Problems with Genehunter

• Space Requirement
• O(22n) n number of non-founder
• Time Requirement
• O(m22n) m-number of markers
• Genehunter is inefficient in terms of Space and
  Time usage !

6
Parallel Genehunter

• Goals
• Not alter the mathematical detail
• Solution
• Genehunter running for each family can be
  considered as an independent task
• Master-slave model
• Message Passing Interface

7
Algorithm
Slaves
Master
task1
task2
P0
task3
task4
8
Algorithm
Slaves
Master
WORK_REQUEST
P0
New tasks
9
Evaluation

• Experiment environment
• Network of workstation(NOW) 10 processors
  (distributed memory)
• HPC3500 8 processors (shared memory)
• SunFire6800 24 processors( shared memory)
• Parameter for test cases

10
Evaluation
11
Evaluation
12
Evaluation
13
Related Work

• Low-level approach
• Parallel Genehunter Implementation of a linkage
  analysis package for distributed memory
  architectures

14
General approaches

• Three model of parallelism
• High level based on family
• Low level based on different markers
• Combine low-level and high-level approach
• Select suitable model based on parameters
• C1 number of family
• C2 number of markers
• C3 structure of family

15
Observations

• Pros
• Useful Genehunter is a popular
• Simple strategy
• Cons
• Too simply to be efficient
• Heavy communication between processors
• Cannot scale to larger number of processors
• Workload unbalance with large family
• No concrete example how about to select different
  models
• List Implementing combine model as future work