CQoS Update

CQoS Update

Li Li, Boyana Norris, Lois Curfman McInnesArgonne National Laboratory

Kevin HuckUniversity of Oregon

2

Outline

Interfaces and components for– Performance database management – CQoS parameters

PerfExplorer and ongoing work on CQoS analysis components

3

Analysis Infrastructure Performance monitoring, problem/solution characterization, and performance model building

PerformanceDatabases

(historical & runtime)

Interactive Analysis and Model Building

SubstitutionAssertionDatabase

SubstitutionAssertionDatabase

Scientist can analyze data interactively

Scientist can provide decisions on substitution and reparameterization

Instrumented Component

Application Cases

Instrumented Component

Application Cases

Control Infrastructure Interpretation and execution of control laws to modify an application’s behavior

Control System(parameter changes andcomponent substitution)

Control System(parameter changes andcomponent substitution)

CQoS-Enabled Component Application

CQoS-Enabled Component Application

Component AComponent A

Component BComponent B

Component CComponent C

ComponentSubstitution Set

ComponentSubstitution Set

4

Outline

Motivation Introduction to components for high-performance computing and

computational quality of service (CQoS) architecture Database component design Application examples Ongoing and future work

5

CQoS Database Component Design

Designed SIDL interfaces for CQoS database management

Implemented prototype database management components

– Description and software:

http://wiki.mcs.anl.gov/cqos/index.php/CQoS_database_components_version_0.0.0

– Based on PerfDMF performance data format and PERI metadata formats

– Comparator interface and corresponding component for searching and matching parameter sets

6

CQoS Database Component Design

AdaptiveHeuristicAdaptiveHeuristic

Perf. ComparatorPerf. ComparatorPerf. data: compare/match

Perf. DatabasePerf. DatabasePerf. data: query/store

… …

… …

: component

: component connection

Fig.1. Connect database and comparator components to adaptive heuristics component. There can be multiple database and comparator components that deal with different data types.

Metadata: query/store

Metadata: compare/match Meta-ComparatorMeta-Comparator

Meta-DatabaseMeta-Database

7

CQoS Performance and Metadata

Performance (general)– Historical performance data from different instances of the same application

or related applications:• Obtained through source instrumentation, e.g., TAU (U. Oregon)• Binary instrumentation, e.g., HPCToolkit (Rice U.)

Ideally, for each application execution, the metadata should provide enough information to be able to reproduce a particular application instance. Examples:– Input data (reduced representations)

• Matrix properties, condition number

– Algorithmic parameters• Convergence tolerance, CFL number, maximum number of iterations

– System parameters• Compilers, hardware

– Domain-specific• Provided by scientist/algorithm developer

8

Outline

Motivation Introduction to components for high-performance computing and

computational quality of service (CQoS) architecture Database component design Application examples Ongoing and future work

9

Database Component Application – Example 1: 2D Driven Cavity Flow1

1 T. S. Coffey, C.T. Kelley, and D.E. Keyes. Pseudo-transient continuation and differential algebraic equations. SIAM J. Sci. Comp, 25:553–569, 2003.

Linear solver: GMRES(30), vary only fill level of ILU preconditioner Adaptive heuristic based on:

– Matrix properties (which change during runtime) computed with Anamod (Eijkhout, http://sourceforge.net/projects/salsa/)

10

How Database Components Work?

During runtime, the driver (e.g., linear solver proxy component) evaluates important matrix properties, and matches the properties to historical data in MetaDB through PropertyComparator interfaces.

Linear solver performance data is retrieved and compared given the current matrix properties. This is accomplished by the PerfComparator component.

The linear solver parameters resulting in the best performance, in this case fill level of ILU preconditioner, is returned back to the driver.

The driver adapts accordingly to continue execution.

11

Example 2: Parallel Mesh Partitioning in Combustion Simulations1

J. Ray et al. (Sandia) have developed a CCA toolkit for flame simulations using structured adaptive mesh partitioning (SAMR). No single partitioner is optimal; thus, CQoS support for choosing an efficient meta-partitioner and an appropriate configuration for a given mesh is desirable.

Meta-partitioner related information includes:– Algorithm (i.e. partitioner) settings

• E.g., actual_levels, good_enough, smoothing, maxNRLoadImbalance

– Problem (mesh) characterization• E.g., number of levels, amount of refined area per level

– Performance metrics• E.g., synchronization cost statistics, data migration cost statistics

1J. Steensland and J. Ray, "A Partitioner-Centric Model for SAMR Partitioning Trade-Off Optimization : Part I," International Journal of High Performance Computing Applications, 2005, 19(4):409-422.

12

Meta-Partitioner Example1

1Johan Steensland and Jaideep Ray, "A Partitioner-Centric Model for SAMR Partitioning Trade-Off Optimization: Part I", Proceedings of the 4th Annual Symposium of the Los Alamos Computer Science Institute (LACSI04). 2004.

Database Components for SAMR Partitioner

13

14

How Database Components Work?

The CharacterizationComparator component matches current AMR grid characterization against historical data in MetaDB to find and extract the most similar previously encountered state

For the returned state, we choose an appropriate rule that matches the state to an optimal partitioner– How are the rules constructed?

• The performance of various partitioners are compared for a given grid characterization• The performance comparison takes into account the offset among

different metrics • A rule is created to map the grid state to a best-performed

partitioner setting• These are accomplished through the PerformanceComparator and

AlgorithmComparator components The rule maps the current grid state to an optimal partitioner The main driver adapts to the new partitioner to continue

simulation.

15

Ongoing and Future Work (Incomplete List)

Validate current algorithm/solver selection strategies with application experiments

Incorporate more offline performance analysis capabilities (machine learning, statistical analysis, etc.)

Introduce a lightweight runtime database to avoid overhead of accessing SQL databases (should only access database in the beginning and after the end of the main computation)

Apply to more problem domains, implementing extensions as necessary

Integration of ongoing efforts in – Performance tools: common interfaces and data representaion (leverage

PERI tools, PerfExplorer, TAU performance interfaces, and other efforts)

– Numerical components: emerging common interfaces (e.g., TOPS solver interfaces) increase choice of solution method automated composition and adaptation strategies

16

Acknowledgements to Collaborators

Victor Eijkhout, the University of Texas at Austin Jaideep Ray, Sandia National Laboratory Henrik Johansson, Uppsala University, Department of Information

Technology, Sweden

17

Thank you!