Skip to content

Latest commit

 

History

History
130 lines (101 loc) · 5.64 KB

notes.md

File metadata and controls

130 lines (101 loc) · 5.64 KB
Raw

LASER Motivation and Priorities

Goals

  • Support elimination and eradication programs through answering questions related to spatial transmission of disease.
    • Measles
    • Polio
    • Malaria
  • Enable IDM researchers to lead the way with cutting edge capabilities.
    • Implement new features and capabilities previously unavailable.
  • Enable new scenarios for existing models
    • Increased performance – explore larger parameter space
    • Increased performance – explore larger population scenarios on existing hardware
  • Enable IDM researchers to model large scale scenarios.
    • GPU based compute
    • Multi-core/multi-threaded compute
    • Cluster based compute
  • Enable external partners and users to model relevant scenarios on readily available hardware.
    • Maximize performance on commonly available hardware (laptops <= 10 years old).
    • Performance on laptops w/Nvidia GPUs.
    • Consider performance on Apple Silicon devices.
  • Design for quick prototyping and extensibility.
    • Enable internal investigation into new features and capabilities
    • Enable timely development of necessary features
    • Enable external partners and users to extend and customize tools for their scenarios

Technology

  • Python w/NumPy
    • Numba as appropriate
    • PyCUDA as appropriate and possible
    • C++ or Rust for specific, high performance capabilties
  • Agent based communities
    • Homogeneously mixed
    • Agent based communities with cohorts
    • Agent based communities with explicit networks
    • Stochastic, compartmental communities
  • Flexible spatial connectivity component
    • “scoops” of contagion, i.e. “anonymous” contagion, e.g. measles
    • Traceable/trackable agents and infection, e.g. malaria vector or parasite genetics
    • Layered connectivity – local, regional, etc.

Processes

  • Documentation
  • Scenarios and examples
  • Test suite

Other

  • multiple independent species, e.g. humans, mosquitos, dogs, ...
  • multiple independent transmission networks, intra-mural and inter-state
  • event based modeling, e.g. draw for non-disease death date
  • state machine based modeling
  • toolset vs. builtin - e.g. is building network a tool in the toolset or a function in the model setup?
  • relationship w/calibration

Edward on Model Architecture and Light Agent Spatial Models

https://wiki.idmod.org/pages/viewpage.action?pageId=195791084

The spatial dimension is critical in modeling transmission from endemic conditions through to the last cases before elimination. In some cases, it's also important to explicitly model arbitrary details of individuals, their behavior, and the combinations of interventions they have ("heavy agent"). In other cases, questions are more about meta-population dynamics and village-scale variation in connectivity, vaccination, surveillance, etc. and more simplistic representations of individual state ("light agent") can improve performance, configurability, interpretability. We expect to scope + explore some of the design considerations within the Architecture Function, and we would expect the Advanced Spatial Modeling cluster to drive a lot of the problem-space assessment as well as the engineering work.

  • usability (API, syntax, portability, installation+configuration)

  • performance (algorithms, approximations)

  • generalizability (code design)

  • scientific validity (context relevance, dynamical system behavior)

  • calibration as a 1st class consideration

Technicalities

Minimal Payload bytes
immune status (T/F) 1
infectious timer (< 256 days) 1
Σ 2
Realistic Payload bytes
age of DoB (115 years = 41,975 days) 2
susceptibility (would use float16 on GPU) 4
accessibility 1
SES/nutrition status 1
exposed timer (< 256 days) 1
infectious timer (< 256 days) 1
age at infection 2
community ID (up to 65K communities) 2
Σ 14
Population Minimal Payload
(2 bytes)		 Minimally Realistic
Payload (14 bytes)		 CPAS* pass/sec† notes
1M (106) 2MB 14MB 2,000 23,000/~3,300
10M (107) 20MB 140MB 200 2,300/~330 megacity
100M (108) 200MB 1.4GB 20 230/~33
200M (2x108) 400MB 2.8GB 10 115/~16 Nigeria (231M)
Pakistan (242M)
1G (109) 2GB 14GB 2 23/~3.3

*CPAS - cycles/agent/second (2GHz CPU). SIMD instructions could be a multiplier on the number of instructions/cycle, e.g. AVX on Intel with 256-bit registers can operate on 8 float32s simultaneously. Apple Silicon supports ARM NEON with 128-bit registers.

†How many passes over memory/second based on memory bandwidth (Lenovo ~46GB/s, Apple Silicon up to 800GB/s).

PRNG State Considerations

PRNG state appears to use between 16 and 268 bytes depending on the algorithm. CPU based modeling can probably allocate state for each CPU code in use (vanishingly small memory use). GPU based modeling will probably need to group agents to use draws from a shared PRNG to amortize the memory usage across several agents. E.g., 16 agents using the same PRNG with 16 bytes of state only incur +1 byte of state/agent.

Algorithm State Size (bytes)
XOROSHIRO128+ 16
other XOROSHIRO variations 16+
XORWOW 44
MRG32k3a 44
Sobel32 140
Sobel64 268

Cohorts

  • How many properties are tracked?
  • How do cohorts progress over time?
  • How do cohorts interact with other entities/agents?
  • How do agents exit the cohort?

Other

Compose Rather than Configure

Choose the necessary components for the model and for reporting rather than enabling/disabling features from a large feature set.