Skip to content

LASER vs EMOD

Jump to bottom
Jonathan Bloedow edited this page Sep 6, 2024 · 20 revisions

Purpose

We were asked to do a little "apples-to-apples", bare metal comparison of LASER and EMOD.

Design

  • Scenario: Measles in England & Wales​
  • Population: ~500k agents at the start of simulation
  • Nodes: 950-1450​
  • Duration: 10 Years
  • Sampling: Full in time and people
  • Platform: COMPS​
  • Parellelization: Multicore​

Design Details

Scenario.

Both EMOD and LASER had working versions of Measles in England and Wales in the 1950s and these are highly spatial, so it made sense to use this for the exercise.

Population.

The population was chosen to be as big as possible while still running in about an hour.

Nodes.

These ought to be the same. Investigating why EMOD has more than LASER. Note also that in the existing EMOD example, we were pruning out nodes with populations before a certain threshold. This was because with our 5% sampling rate, a population of 2000 people, say, would come to 100 agents, which really doesn't perform well when our susceptible population is mostly under 5, or about 5-10 agents. This same math comes into effect if we sample at 100% but scale down the population in absolute terms.

Duration.

Could easily be 20 years, but 10 years proves the same point.

Sampling.

In order to get the best apples-to-apples comparison, we did away with all the resampling/downsampling optimizations that a modeler would exploit in EMOD or LASER. In EMOD, we adaptively downsample immune individuals. In LASER, we use EULAs. So while this creates a 'fairer' comparison in some sense, it also means the numbers don't really represent how one would really use either system.

Platform.

It was easy enough to run both on COMPS to get the best numbers on common, high powered hardware.

Features

  • SEIR
  • Migration
  • No interventions
  • Seasonality
  • Maternal Immunity

Discrepancies

  • EMOD has disease deaths
  • Endemicity is robust to population scaling in EMOD but seems to require recalibration in LASER prototype.
  • EMOD scenario has 1450 nodes while LASER has 940. Investigating.

Results

COMPS:

800k: EMOD runs in about 66 minutes (with too many births) while LASER runs in about 35 seconds.

370k: EMOD runs in 25 minutes, while LASER runs in about 25 seconds.

Open Items

  • Scaling population up and down in EMOD doesn't seem to be playing nicely with how fertility was being done in the example with the result that the CBR seems to be doubling (too many babies) in the "800k population" version.
  • Going with 375k -- 1/10 the the original -- seems to preserve the fertility CBR of ~19 without further investigation.
Clone this wiki locally