Northwestern University

1
Access Patterns, Metadata, and Performance

• Alok Choudhary and Wei-Keng Liao
• Department of ECE, Northwestern University
• Collaboration with ANL

SDM kickoff meeting July 10-11, 2001
2
Virtuous Cycle
Simulation (Execute app, Generate data)
Problem setup (Mesh, domain Decomposition)
Manage, Visualize, Analyze
Measure Results, Learn, Archive
3
I/O Flow for Scientific Simulation

• Many scientific applications perform
  simulation/analysis
• 3 types of output Checkpoint, visualization,
  data for analysis
• Checkpoint keeps re-writing to the same files
• Visualization output
• 2 types of input new run or restart

4
Data Access Sequence Dependency

• Temporal dependency
• Access the same data set at different time stamp
• Spatial dependency
• Access different data sets at the same time stamp
• Resolution dependency
• Access the same data set at different resolution
• Sequence is useful for I/O performance
  improvement, eg. Pre-fetch, pre-stage, storage
  continuity

5
Spatial Data Access Patterns

• Parallel partition patterns
• Regular, irregular
• Static, dynamic during simulation
• Access sequence
• Spatial, temporal, resolution
• Access frequency
• Once only, multiple times (overwrite for restart)
• Access amount
• Large, medium, small chunks

6
Access Patterns for Visualization/Analysis

• Generated from real data during simulation or in
  post-simulation process
• Smaller size than real data
• Type conversion,
• eg. float ? unsign char
• Reduce/increase resolution
• Projection 3D to 2D
• 3 types of data generate and display sequence

7
Architecture
Simulation Data Analysis Visualization
User Applications
I/O func (best_I/O (for these param)) Hint
Query Input Metadata Hints, Directives Association
s
Data
OIDs parameters for I/O
Schedule, Prefetch, cache Hints (coll I/O)
Storage Systems (I/O Interface)
MDMS
Performance Input System metadata
Metadata access pattern, history
MPI-IO (Other interfaces..)
8
Approach

• Management meta data using OR-DBMS
• Collect and organize meta data in relation tables
• Design meta data query interface using SQL
• Access to HSS
• Obtain current storage layout, configuration
• Native I/O interfaces or MPI-IO
• I/O optimization
• Determine optimal I/O calls
• Overlap I/O with computation, communication, and
  I/O
• Pre-fetch, pre-stage, migrate, purge in HSS
• Sub-filing for large file, file container for
  small files

9
Objective and Goal

• Meta data management
• Collect historical meta data, process user
  provided meta data, update meta data w.r.t
  environment changes
• Efficient query for meta data
• High performance I/O
• Automatically determine optimal I/O calls from
  data access patterns
• Improve performance by prefetching, caching,
  layout, inter-object association, striping, etc.
• Support for Hierarchical Storage Systems (HSS)

10
Metadata

• Application Level
• Algorithms, compiling, execution environments
• Time stamps, parameters, result summary
• Programming Level
• Data types, structures, association of datasets,
  partition patterns
• Storage System Level
• File locations, file structure, I/O modes, host
  names, device types, path names, storage
  hierarchy
• Performance Level
• I/O bandwidth of HSS for local and remote access
• Data access sequence, frequency, other access
  hints
• Collective or non-collection I/O

11
Applications

• Asto3D -- study the highly turbulent convective
• layers of late-type star
• Write only
• regular partition on all data sets
• ENZO -- simulate the formation of a cluster of
• galaxies consisting of gas
  and stars
• Both read and write
• Both regular and irregular partition
• Adaptive Mesh Refinement dynamic load balancing
• Common feature
• Checkpoint / restart
• Post-simulation data analysis
• Visualizing the process of the computation in the
  form of a movie

12
Interface
13
Run Application
14
Dataset and Access Pattern Table
15
Data Analysis
16
Integrating Analysis
Simulation (Execute app, Generate data)
On-line analysis And mining
Problem setup (Mesh, domain Decomposition)
Manage, Visualize, Analyze
Measure Results, Learn, Archive
17
Visualize
18
Meta Data Representation in Database
19
Future Directions and Challenges

• H/W and S/W mainly driven by commercial
  applications (e.g., web, DB, Content Delivery
  etc.)
• H/W architectures (e.g., Infiniband)
• S/W architectures (e.g., ODB, ORDB, DW, mining
  tools)
• Challenge Can we adapt and enhance these to
  satisfy scientific computing file systems,
  storage, and data management requirements?
• E.g., Parallel file systems on Infiniband
  architectures so that uniform UI and access may
  be provided from different systems?
• Can we incorporate DM and analysis capabilities
  as well as efficient I/O techniques within DM
  systems
• File Systems and DM Challenges
• Can it be be customized with high performance?
• Can it learn from user access patterns?
• Can it optimize accesses automatically?
• Can it provide high-level interface which is
  uniform?