Infrastructure for OptIPuter Storage Systems Research

1
Infrastructure for OptIPuter Storage Systems
Research

• Simulators, Trace Files, and Models for
  Experimentation
• Summer 2003
• CSAG presentation
• Justin Burke and Andrew Chien

2
Goals of Infrastructure

• We need tools to allow us to simulate (or
  emulate) transactions over distributed file and
  storage systems.
• These tools need to complement actual
  experimental testbeds.
• We need to understand which are useful for
  various types of experimentation and proof of
  concept.

3
Overview of Storage Subsystem
4
Modeling Storage Systems

• A strong argument for detailed models An
  Introduction to Disk Drive Modeling.
• Complexities of disk subcomponents necessitate a
  detailed disk model.
• Establishes a base-line for metrics demerit.
  Widely used for model/simulator/emulator
  validation.

5
Demerit Values
From An Introduction to Disk Drive Modeling,
Ruemmler and Wilkes, 1994
6
Disk Characterization

• Difficult to obtain a useful simulation without
  accurate disk parameters.
• Manufacturers do not publish detailed
  characterizations.
• ? Have to obtain empirically.
• Unfortunately, standard characterization tools
  do not exist and may have to be custom-made.

7
Simulation Tool Evaluation

• Simulators require a lot of time and effort to
  build, debug, and validate.
• Ideal simulator includes programming interface,
  easy and flexible configuration, validated model,
  and support. )
• Availability and ongoing development
• Capabilities Trace Files, Ease of Use, Parallel
  Drive Configuration, Flexibility

8
Available Simulators

• Pantheon
• Disksim
• Kotz's Simulation Model of the HP 97560 Disk
  Drive and Yiming Hu's added support for multiple
  disks
• Nils Nieuwejaar's Disk Simulator
• STARFISH

http//www-csag.ucsd.edu/individual/jburke/Storage
/internal/simulators.txt
9
Personal Selection

• Disksim can meet most of our requirements fairly
  well flexible design, very accurate.
• Extensible
• It is also under active development.
• Frequently used in FAST papers.

10
Use of Trace Files

• When using a simulator, there are two main
  options for simulation synthetic workload and
  trace file.
• Synthetic workload good for broad evaluation.
  Traces good for targeted evaluation.
• Disk traces normally include
• start/stop timestamp, sector (or
• block) , request length

11
Trace File Availability

• Fantasy plentiful traces taken from variety of
  production environments using parallel/distributed
  file systems.
• Reality frequent use of synthetic workloads
  most popular traces are old (early 90s) very
  limited availability of parallel/distributed
  traces.

12
Trace Files in the Wild

• HP Cello 92, 96, 99, 02(?)
• HP TPC-C
• HP TPC-D
• HP Snake
• Storage Performance Council
• OpenMail email server
• Postmark
• Custom Traces

http//www-csag.ucsd.edu/individual/jburke/Storage
/internal/traces.txt
13
Models for Experimentation

• Data Server selection, high performance
• User data placement (for above)
• System Manager incremental benefit of resources
• System Manager fixed application deployment

14
Server selection

• Given an application computational and data
  access specification
• The set of resources which can satisfy these
• Select the set of resources which give best
  performance
• Select the set of resources which gives most
  robust performance (headroom, interference)
• How do these properties/choices vary based on
  resource configuration?
• Complement where should you place your replicas
  based on known access patterns? What about
  declustering?

15
Application Deployment

• Given an application computational and data
  access specification
• The set of resources which can satisfy these
• Existing set of utilizations and capacities
• Select the set of resources which give best
  performance
• Select the set of resources which gives most
  robust performance (headroom, interference)
• How do these properties/choices vary based on
  resource configuration? (and data placement)

16
Resource Deployment

• Given an existing workload, utilizations, and
  capacities
• Where to deploy additional resources (storage) to
  give maximum benefit to existing workload
  (increased performance)
• Where to deploy to give most robust performance
  (headroom, interference)
• Where to deploy to give greatest application
  choice freedom in future? Or maximum increase in
  capacity?
• May need some model of future workload (or may be
  same as current)

17
Application Examples SAR

• Earthscope SAR Application
• High speed data integration/visualization
• 32 gigabytes, delivered in less than 0.5 seconds
• Presumed to be sourced from MANY disks
  distributed throughout the OptIPuter network
• Which/How many disks? What are the critical
  performance factors?
• How are these affected by various sharing and
  security models?

18
Application BIRN

• High speed data integration/visualization of
  brain images
• 10 terabytes, delivered in seconds
• Comparative physiology and visualization
• Presumed to be sourced from distributed sources
  (control)
• Each source may consist of MANY disks distributed
  throughout the OptIPuter network

19
Requirements

• Large scale system modeling (high performance)
• On parallel resources?
• Flexible configuration
• Relation to real experimental platforms

20
Specific Hypothetical Experiments