Transcript ppt

EpiFast: A Fast Algorithm for
Large Scale Realistic Epidemic
Simulations on Distributed
Memory Systems
Keith R. Bisset, Jiangzhuo Chen, Xizhou Feng,
V.S. Anil Kumar, Madhav V. Marathe
23rd International Conference on Supercomputing (ICS'09)
June 11, 2009
Network Dynamics & Simulation Science Laboratory
Outline
•
•
•
•
Background
EpiFast Algorithm
Performance
Summary
Network Dynamics & Simulation Science Laboratory
Motivation
• Pandemic Flu of 1918 was
deadly
– Killed 2.5 - 5% of global population
– Many many more were sick
– Resulted in massive upheaval of
society
– Virtually no place on Earth was
spared
• More recently:
– SARS
– Avian influenza
– Swine flu
• Epidemic simulation problem
Network Dynamics & Simulation Science Laboratory
Network Dynamics & Simulation Science Laboratory
Components of Epidemic Simulation
Problem
• Population and contact network
• Infectious disease
• Interventions
Network Dynamics & Simulation Science Laboratory
Create a Synthetic Population
• Census data
– Individual demographics: age, gender…
– Household characteristics: size, income…
Network Dynamics & Simulation Science Laboratory
Generate Contact Network
• Locations: D&B data
• Activity surveys.
– Matched to individuals by demographics
– Matched to locations by activity type
• Generate social contact network
– People follow activity schedules
– Activities take them to locations
– At locations they interact with each other
Network Dynamics & Simulation Science Laboratory
Generate Contact Network
Network Dynamics & Simulation Science Laboratory
Social Contact Network
• All interactions in population captured
–
–
–
–
Duration of contact
Type of activity resulting in contact
Demographics of those contacted
Characteristics of locations
Network Dynamics & Simulation Science Laboratory
Social Contact Network
• Interactions provide
opportunity for
disease transmission
• All interactions in a
population can get
very complex
• Eg. Alabama has 4.3
million people and a
total of 291 million
interactions
Network Dynamics & Simulation Science Laboratory
Background: SEIR Disease Model
• Individuals move through states with different
characteristics
• Demographics
• Level of symptoms
• Level of infectiousness
• Response to treatments
Network Dynamics & Simulation Science Laboratory
Disease Spread in Contact Network
• Transmission depends on
–
–
–
–
Duration of contact
Type of contact
Characteristics of the infectious person
Characteristics of the susceptible person
Network Dynamics & Simulation Science Laboratory
Background: Interventions
• Different types of interventions help to mitigate
the epidemic
– Pharmaceutical: vaccination, antiviral
– Non-Pharmaceutical: social distancing, school closure,
work closure
• When, how, and to whom these are applied can
have different impact on the course of the
epidemic
Network Dynamics & Simulation Science Laboratory
Obstacles to Interventions
• Supply: many interventions are of a limited
supply thus only a fraction of the population may
be eligible for the intervention
• Compliance: not all individuals will be able or
willing to comply with the intervention
• Efficacy: not all interventions are fully effective
even if complied with
Network Dynamics & Simulation Science Laboratory
Vaccination
• Vaccination changes an individual’s role in the
transmission chain
– Lowers susceptibility to infection
– Lowers infectiousness if infected
• The degree these are lowered depends on the
efficacy of the vaccine
• Predicted efficacies and supply levels of
pandemic flu vaccines vary wildly
Network Dynamics & Simulation Science Laboratory
Antiviral
• Anti-viral treatment changes a individual’s role in
the transmission chain for the duration of their
treatment
– Lowers susceptibility to infection
– Lowers infectiousness if infected
• The efficacies of these treatments depends on:
– The kind of anti-viral administered
– When its administered
Network Dynamics & Simulation Science Laboratory
Social Distancing
• Generic Social Distancing reduces the
opportunities for transmission in the population
– Less contact at public places
• Either through closures or rules on occupancy
– Measures that might reduce transmission
• Masks, no hand shaking, frequent sterilization of common
surfaces
• The degree to which this occurs depends mainly
on compliance
Network Dynamics & Simulation Science Laboratory
School Closure
• School closure reduces opportunities for
transmission at schools
– School children are often involved in the early spread
of influenza epidemics
Network Dynamics & Simulation Science Laboratory
Work Closure
• Work closures eliminate the opportunities for
transmission within the workplace
– Workplaces close their doors
• The degree this will work will depend on the
compliance levels of businesses
Network Dynamics & Simulation Science Laboratory
Application of Interventions
The effectiveness of all interventions depend on when,
how, and to whom they are applied
• When is it triggered?
– An event triggers the implementation of the intervention
(day of simulation or % of a group is infected)
• How well is the plan executed?
– What proportion of the targeted population actually
received / complied with the intervention (levels of
compliance)
• Who was targeted?
– Supply limitations may require prioritization of groups for
different interventions
Network Dynamics & Simulation Science Laboratory
EpiFast Algorithm: Sequential
Network Dynamics & Simulation Science Laboratory
Parallelization
•
•
•
•
Data intensive & computation intensive.
Should scale on distributed memory systems.
Partition data (contact network).
Master-slave model.
Network Dynamics & Simulation Science Laboratory
Parallel EpiFast: Network Partitioning
E
A
D
B
C
Network Dynamics & Simulation Science Laboratory
Parallel EpiFast: Master-Slave Model
• Single master
processor:
communication
talk the talk
• Many slave
processors:
computation
work the work
Network Dynamics & Simulation Science Laboratory
EpiFast Algorithm: Parallel
Sequential:
Network Dynamics & Simulation Science Laboratory
Network Dynamics & Simulation Science Laboratory
EpiFast Performance: Running Time
• C++/MPI implementation, tested on commodity
clusters and SGI Altix systems.
• Los Angeles population: 16 million people.
• 180 days of epidemic duration.
• With and without interventions.
• 25 replicates for each configuration.
• Each replicate takes < 15 minutes.
Network Dynamics & Simulation Science Laboratory
EpiFast Performance: Running Time
Network Dynamics & Simulation Science Laboratory
EpiFast Performance: Strong Scaling
Network Dynamics & Simulation Science Laboratory
EpiFast Performance: Week Scaling
Population Population Size CPU Number
Running Time (seconds)
per simulation day
Miami
2.09
32
0.47
Boston
4.15
64
0.54
Chicago
9.05
128
0.54
Network Dynamics & Simulation Science Laboratory
Network Partitioning Revisited
• Our simple partitioning
method is scalable.
• Can be done online with
very little time: adjust
partitioning based on
available computing
resource to achieve load
balancing.
• Metis produces better
partitioning: slightly
improves communication
complexity, with a
significant overhead.
Network Dynamics & Simulation Science Laboratory
Summary
EpiFast:
• can handle realistic large scale populations;
• has many practical applications: evaluation of
various interventions, public health decision
support;
• runs extremely fast;
• is scalable: on both shared & distributed memory
systems.
• Is a novel HPC application: epidemic simulation.
Network Dynamics & Simulation Science Laboratory
Thanks!
Future Work
• Implement EpiFast with UPC.
• Port EpiFast to GPGPU or Cell based clusters.
Network Dynamics & Simulation Science Laboratory