NAMD2: Greater Scalability for Parallel Molecular Dynamics

1
NAMD2 Greater Scalability for Parallel
Molecular Dynamics

• Scott Callaghan
• CSCI 653
• Spring 2006

2
Overview

• Why?
• Applications
• Current models
• Algorithm
• Pieces
• Computations
• Implementation
• Coding
• Load balancing
• Performance

3
Why yet another approach?

• Biomolecular systems
• 104 105 atoms
• 106 timesteps
• Bonded and non-bonded computation
• Modular

4
Current Methods

• Force Decomposition
• Not scalable
• Spatial Decomposition
• Scalable
• High overhead for low N/P
• Combine both!
• Scalable
• Keep N/P high while having high P

5
Algorithm Components

• Patches
• Compute Objects
• Proxies
• Sequencers

6
Patches

• Spatial decomposition components
• Responsible for
• Checking and migrating atoms (4-8 timesteps)
• Forces on those atoms

7
Compute Objects

• Force decomposition components
• Responsible for
• Bonding and non-bonding forces
• Relaying forces to the respective patch
• Modular

8
Algorithm so far
This seems good
but what about this?
P0
P0
P1
Patch (0, 0, 0)
Patch (0, 0, 1)
Patch (0, 0, 0)
Patch (0, 0, 1)
?
Compute non-bonding
Compute non-bonding
9
Proxies

• Act as go-betweens
• Responsible for
• Getting atoms from home patch
• Giving atoms to compute node
• Getting forces from compute node
• Giving forces to home patch
• Better than direct communication

10
Sequencers

• Manages the home patch
• Responsible for
• Sending compute requests for a subset
• Updating velocities
• Notifying home patch of position updates
• Calculates energies
• Changes to logic go here
• Basic pseudocode on pg. 293

11
Lets pause for a summary
P0
P1

• Sequencer updates
• Notifies home patch
• Migrate atoms
• Populate proxies
• Notify compute objects
• Calculate forces
• Update forces
• Update velocities
• Calculate energies
• Repeat!

Sequencer
Patch (0, 0, 0)
Patch (0, 0, 1)
Proxy
Compute non-bonding
12
Non-bonded interactions

• 3 types
• Self
• Pair
• Exclusion
• Ignored for 1-2 and 1-3 atoms
• Modified for 1-4
• Pair-list of atoms within cutoff distance
• Compute all and exclude later

13
Self and Pair algorithm

• For all pairs in this and neighboring patches
• If the atoms are inside the cutoff
• If the atoms are inside the exclusion cutoff
• If the atoms are 1-4 excluded
• calculate modified interaction
• else if the atoms are not excluded
• calculate normal interaction
• else
• calculate normal interaction

14
Exclusion

• For all pairs of excluded atoms
• If the atoms are beyond the exclusion cutoff
• calculate and subtract normal interaction
• If the atoms are 1-4 excluded
• calculate modified interaction

15
Why this complex approach?

• Greater force decomposition
• Expensive to check exclusion
• Dont add/subtract L-J potential for
• Add cutoffs to provide some speedup
• Exclusion
• Hydrogen group
• Perform cutoffs together

16
Bonded interactions

• 5 compute objects (bond interactions) per node
• Computed by proxies on min(x, y, z) patch
• Minimizes communication 7 down, 7x up
• Can use same proxies for non-bonded
• Uses lists of atoms in tuples

17
Bonded algorithm

• For all home or upstream (proxy) patches
• For all atoms in this patch
• For all tuples with this atom as the first atom
• look up the patches with atoms in this tuple
• calculate the common downstream patch
• If the downstream patch is you
• add this tuple to the list to calculate

18
Implementation

• Objects which send messages
• Uses Converse
• Charm (execute and communicate)
• Other strange packages
• Type hierarchy for easy modification
• Encapsulation

19
Load Balancing

• First on startup
• Distribute patches
• 1 bond force compute object per processor
• Non-bonding
• Self on the same processor as the patch
• Patch-pair on upstream

20
Load Balancing, cont.

• After n steps
• Calculate t for non-migratable objects
• Determine average load
• Iterate through migratable objects, long to short
• Put on proc so new load
• Examine 2-proxy, then 1, then 0 (self objects)
• Assign and update loads
• Do second pass with smaller overload
• Repeat second pass after m more steps

21
Performance

• 80 efficiency on P128
• 10 communication overhead
• 10 idle time
• Independent comparisons can be found at
  http//www.ks.uiuc.edu/Research/namd/performance.h
  tml
• Best for very large processor counts

22
Questions?