High Performance Computing for Disease Surveillance

1
High Performance Computing for Disease
Surveillance

• David Bauer
• Brandon Higgs
• Mojdeh Mohtashemi

The MITRE Corporation 7525 Colshire Drive McLean,
VA 22102
2
Goal Apply HPC to ScanStat

• Improve the performance of ScanStat algorithm by
  distributing algorithm across multiple processors

Expected Number of cases is defined for (s, t),
where s is a spatial cluster, and t is the time
span
then for all cases of outbreak N
3
HPC Amdahls Law

• Amdahl quantifies the amount of improvement in
  runtime (SpeedUp) that can be achieved using
  multiple processors

where F is the percentage of an algorithm that is
sequential (i.e., cannot be parallelized)
4
ScanStat Details

• Data source San Francisco Dept of Public Health,
  Tuberculosis Program
• Spatial blocks may be census tracts or individual
  addresses
• 76 CT form 441 centroids (rectangular)
• 392 individuals form 4,234 centroids
  (rectangular)
• Time window 4-72 weeks spanning a 10 year period
• Based on case counts

5
Parallel ScanStat census tracts

• Analysis of ScanStat workload suggests most
  centroids complete quickly
• A few centroids require almost 2 hours to compute
• 441 centroids
• 10 year time span
• Single CPU 28 hrs

6
Parallel ScanStat census tracts

• SpeedUp from parallelization limited by longest
  running centroid
• 440 centroids computed in 0.5 hr
• Longest centroid bound running time
• (7k seconds)

7
Parallel ScanStat individual addresses

• Analysis of ScanStat workload suggests most
  individual addresses also complete quickly
• A few addresses require over 3 hours to compute
• 4,234 centroids
• 10 year timespan
• Single CPU 74 hrs

8
Parallel ScanStat individual addresses

• SpeedUp from parallelization limited by longest
  running centroid
• 4,233 centroids completed in less than 0.5 hr
• Longest centroid bound running time
• (11.8k seconds)

9
Future Work

• Parallelize ScanStat algorithm in the time domain
• Move ScanStat to distributed computing
  environment to
• Reduce HPC environment costs
• Potentially increase computing power
• Apply to a decision support problem with
  real-time constraints
• Interested in finding collaboration partners