Uncertainty in LargeScale Computing

1
Uncertainty inLarge-Scale Computing

• Travis Desell

2
Overview

• Motivation
• Uncertainty?
• ARM (Autonomic Runtime Manager)
• Active Harmony

3
Extensible TeraScale Facility (ETF)
RPI
4
iVDGLInternational Virtual Data Grid Laboratory
www.ivdgl.org
5
CERN Worlds Largest Computing Grid
http//goc02.grid-support.ac.uk/googlemaps/lcg.htm
l
www.cern.ch
6
PlanetLab (November 06)
718 Nodes at 315 Sites
7
Map of Rensselaer Grid Clusters
Nanotech
Multiscale
Bioscience Cluster
CS /WCL
Multipurpose Cluster
CS
CCNI Cluster
8
Areas of Uncertainty

• Application
• Non-determinism
• Environment
• Dynamic resource availability
• Unknown target environments
• OS/Hardware effects
• Faults

• Interaction
• Competition
• Distribution
• Unknown performance on different architectures

9
Overview of ARM

• Autonomic Runtime Manager (ARM)
• What is it for?
• What is it?
• Reconfiguration Methodology
• What can it do?
• Graph Partitioning
• Natural Region Partitioning

10
Adaptive Distributed Applications

• Non-deterministic
• Application divided into computational regions
• Regional computational complexity can change
  dramatically in both space and time
• Example Wildfire Simulation

11
ARM (Autonomic Runtime Manager)
12
When to reconfigure?

• Imbalance Ratio (IR)
• Reconfigure if IR gt threshold

13
Natural Region Partitioning

• Calculate Processor Computational Load (PCL)
• Processor Load Ratio (PLR) is normalization of
  PAF
• Processor Allocation Factor (PAF) is a ratio of
  how fast cells are computed
• Application Computational Workload (ACW) is the
  total workload of the application

14
Graph Partitioning

• Wildfire simulation domain represented as graph
  G(V,E)
• Vertices are cells
• Edges connect neighboring cells
• Verticies weighted by computational complexity
• Burning gt Unburned
• Graphs partitioned to create cuts with similar
  vertex weight, and inter-cut edges

15
Results
16
Results (Continued)
17
Overview of Active Harmony

• Active Harmony
• What is it?
• Reconfiguration Methodology
• Case Studies
• PETSc
• POP
• GS2

18
Active Harmony

• Previously used for runtime performance tuning in
  dynamic environments.
• Developers specify tunable parameters which can
  modify applications during runtime.
• This work focuses on using Active Harmony
  off-line as opposed to on-line.

19
Active Harmony Architecture
20
Reconfiguration Methodology

• Representative short runs used to provide
  measurements of application performance for
  different tuned parameters.
• Uses an algorithm based on the Neader-Mead
  simplex method.
• Parameters are treated as independent dimensions.
• Modified because parameters are non-continuous.
• Iteratively tunes application performance by
  repeatedly modifying parameters and converging
  and an optimum.

21
Case Study - PETSc

• PETSc Portable Extensible Toolkit for
  Scientific Computation
• Suite of data structures and routines for
  scalable (parallel) solution of scientific
  applications based on partial differential
  equations
• Uses MPI
• Used Active Harmony on two PETSc examples
• SLES linear equation solver, uses matrix
  decomposition
• 2-D driven cavity problem

22
Results - SLES

• Ran 50x50 matrix decomposition with 4 homogeneous
  and 4 heterogeneous processors (see above).
• Ran 21,025 x 21,025 matrix decomposition and 32
  processors. Resulted in a 18 performance
  speedup after tuning.
• Also ran 90,601 x 90601 matrix decomposition.
  Full search space O(10100). After 120
  iterations resulted in a 15-20 performance
  improvement.

23
Results 2-DCP

• Uses PETSc SNES non-linear equation solver.
• Tunable parameter is how many grid points
  distributed among nodes (default is equal
  distribution).
• Above shows results from a small problem, 2500
  grid points and 4 nodes.
• Ran with 40,000 grid points on 32 processors,
  speedup was 11.5 over default partitioning.
  Search space was O(1036).

24
Case Study - POP

• POP Parallel Ocean Program
• Developed at Los Alamos National Labs.
• Used by the Community Climate System Model (CCSM)
  as the ocean component.
• Solves three-dimensional primitive equations for
  fluid motions on a sphere.

25
Results - POP

• Problem size 3600x2400 grid points divided into
  480 blocks (processors).
• Default configuration is 180x100 sized block.
• Run on 16-way SMP nodes.
• No single block size is best for all topologies
• Execution time reduced up to 15 by tuning block
  size.
• Tuning other parameters (20 performance related
  with 2-4 possible values) on a 8 node 4 processor
  cluster resulted in 12.7 speedup after 12
  iterations, and best speedup of 16.7 after 27
  iterations.

• First set of x-axis labels is processing nodes
• Second set of x-axis labels is best block size

26
Case Study GS2

• Physics application used to study low-frequency
  turbulence in magnetized plasma.
• Simulation involves billions of mesh points.
• Primary tuning parameter was data layout.

27
Results GS2

• Results from NERSC Seaborg (8 16-processor
  nodes).
• Active Harmony reduced execution time from 55.06s
  to 16.25s (3.4x speedup) without collision mode
  and from 71.08s to 31.55s ( 2.3x speedup) in
  collision mode.
• Used benchmarked runs of 10 iterations (as
  opposed to typical runs of 1000 iterations).

28
Results GS2 (2)

• Search space of GS2 is O(105). To test Active
  Harmony, systematic sampling was used to evaluate
  O(104) configurations (see left).
• Best performance sampled was 125.8s, only 2 of
  the sampled configurations resulted in lt 200s.
• Active Harmonys result was in the top 5 of
  sampled configurations.

• Also tuned for other performance related
  parameters in addition to data layout. Total
  speedup was 5.1x improvement.