NAMD

1
NAMD

• September 18, 2003
• L.V. (Sanjay) Kale (kale_at_cs.uiuc.edu)
• http//www.ks.uiuc.edu/Research/namd/

2
NAMD Vision

• Make NAMD a widely used MD program
• For large molecular systems,
• Scaling from PCs, clusters, to large parallel
  machines
• For interactive molecular dynamics
• Goals
• High performance
• Ease of use
• Ease of modification (for us and advanced users)
• Incorporation of features needed by Scientists

3
Three Easy Goals for NAMD 3

• Easy to configure and run
• Help the user to avoid mistakes during setup
• Expect the machine to fail during the simulation
• Easy to extend and modify
• Maximize reuse of communication and control
  patterns
• Push parallel complexity down into Charm
  runtime
• Easy to publish first!
• New Linux clusters with high-latency gigabit
  ethernet
• New cellular machines with 10K-100K processors

4
NAMD 3 Design

• NAMD 3 will be a major rewrite of NAMD
• Incorporate lessons learned in the past years
• Use modern features of Charm
• Refactor software for modularity
• Restructure for supporting planned features
• Algorithms that scale to even larger machines

5
NAMD 3 Programmability

• Scientific Modules
• Forces, integration, steering, analysis
• Keep code with a common goal together
• Add new features without touching old code
• Parallel Decomposition Framework
• Support common scientific algorithm patterns
• Avoid duplicating services for each algorithm
• Start with NAMD 2 architecture (but not code)

6
MDAPI
New Science modules
Replica exchange
QM
Implicit Solvents
Polarizable Force Field
Bonds related Force calculation
Integration
Pair-wise Forces calculation
PME
NAMD Core
Charm modules
FFT
Fault Tolerance
Grid Scheduling
Collective communication
Load balancer
Core CHARM
Clusters
Lemieux
Teragrid

7
MDAPI Modular Interface

• Separate front end from modular engine
• Same program or over a network or grid
• Dynamic discovery of engine capabilities, no
  limitations imposed by interface
• Front ends NAMD 2, NAMD 3, Amber, CHARMM, VMD
• Engines NAMD 2, NAMD 3, MINDY

8
Terascale Biology and Resources
PSC LeMieux
TeraGrid
CRAY X1
NCSA Tungsten
ASCI Purple
Riken MDGRAPE
Red Storm Thors Hammer
9
Basis of NAMD Scalability

• Very active computer science collaboration
• UIUC Parallel Programming Lab, since 1992
• Charm system message driven objects
• Constant tuning for evolving parallel platforms
• Designed for parallel efficiency
• NAMD 1 discrete spatial decomposition, fast
  multipole
• NAMD 2 hybrid force-spatial decomposition, PME
• Dependency-driven execution, no barriers
• Measurement-based load balancing system

10
Modern Charm Design

• Virtualization Object-based Parallelization
• Object array - A collection of chares,
• with a single global name for the collection, and
• each member addressed by an index
• Mapping of element objects to processors handled
  by the system

Users view
A0
A1
A2
A3
A..
System view
A3
A0
11
Benefits of Modern Charm

• Adaptive load balancing
• Optimized communication
• Persistent Communication, immediate messages
• Optimized concurrent multicast/reduction
• Flexible, tuned, Parallel FFT libraries
• Automatic Checkpointing
• Ability to change the number of processors
• while running
• Scheduling on the grid
• Fault tolerance

12
NAMD 3 Fault Tolerance

• Fully automated restart
• Harder to incorporate into scripting interface
• Survive loss of a node
• Larger machines are less reliable

13
Highly Scalable Algorithms

• Generate fine-grained parallelism
• In non-bonded force evaluation
• Smaller patches, with 2-away communication
• Alternative parallel strategies for bonded
  foreces
• In PME
• Using pencil decomposition for FFTs where needed

14
Efficient Parallelization for IMD

• Characteristics
• Limited parallelism on small systems
• Real time response needed
• Fine grained parallelization
• Improve speedups on 4K-30K atom systems
• Time/step goal
• Currently 0.2s/step for BrH on single processor
  (P4 1.7GHz)
• Targeting on 0.003s/step on 64 processors of
  faster machine, that is 20picosecond/minute
• Flexible use of clusters
• Migrating jobs (shrink/expand)
• Better utilization when machine is idle

15
NAMD 3 Initial New Features

• Twin Goals
• Provide immediately useful functionality
• Verify completeness of framework design
• Scientific Modules
• Implicit solvent models (e.g, generalized Born)
• Replica exchange (e.g., 10 on 16 processors)
• Polarizable force fields (e.g., Drude model)
• Hybrid quantum/classical mechanics

16
More NAMD 3 Features

• Self-consistent polarizability with a
  (sequential) CPU penalty of less than 100.
• Fast nonperiodic (and periodic) electrostatics
  using multiple grid methods.
• A Langevin integrator that permits larger time
  steps (by being exact for constant forces).
• An integrator module that computes shadow energy.

17
Integration with CHARMM/Amber?

• Goal NAMD as parallel simulation engine for
  CHARMM/Amber
• Generate input files in CHARMM/Amber
• NAMD must read native file formats
• Run with NAMD on parallel computer
• Need to use equivalent algorithms
• Analyze simulation in CHARMM/Amber
• NAMD must generate native file formats