Grid Computing

1
Grid Computing

• Advance Computer Architecture
• CSE 8383

2
Outline

• The Motivation Behind Grid
• Existing Grid Infrastructures
• Integral Components of a Grid System
• Open Issues

3
Back to the Future

• The complexity for minimum component costs has
  increased at a rate of roughly a factor of two
  per year
• circuit densities in chips would double every
  12-18 months.
• Gordon Moore, Electronics Magazine. April 19,
  1965

4
Processor Evolution

• From Processor Generation n to n1,
• Gate delay reduces by 1/?2 (basic frequency goes
  up by ?2)
• Number of transistors in a constant area (cost)
  goes up by 2
• Additional transistors enable an additional
  increase in performance by the factor of ?2
• Deeper pipeline, Offset pipeline penalties
  (branch prediction), Increased number and/or size
  of caches, Exploit Instruction Level Parallelism
  (ILP)
• Result 2x Performance at roughly the equal cost

5
Processor Evolution

• Over the past 37 years, Moores law has
  successfully predicted the exponential growth of
  component densities

Year Transistors
1970 2,300
1975 6,000
1980 29,000
1985 275,000
1990 5,500,000
2000 42,000,000
6
Processor Evolution

• Moores law held
• Mainly because of advances in micro-architecture
  that exploited huge growth in transistor/area and
  that overcame interconnect limitations
• Scarcity of new micro-architecture ideas
• Pipelining, branch prediction, IP, Caching, has
  reached a point of maturity
• Power / Current Issues
• Power density limitations
• Voltage and doping (less switching energy
  avaialable)

7
From single-CPU to Multi-Core

• Gate delay does not reduce much (basic frequency
  goes up a little)
• Number of transistors in a constant area (cost)
  goes up by 2
• Global Wiring is not stressed
• Result 2x performance at roughly equal cost

8
From single-CPU to Multi-Core

• On the order of 1000, classic pipelined CPUs
  can be fit in the area of a state-of-the-art
  high-end CPU.
• Can have the same performance as high end CPUs
  but with limited scope
• Result 1000x performance at roughly equal the
  cost
• If dual-core is 2x, quad-core is 4x, then why not?

9
Multi-Core The Pit Falls

• Amdahls Law
• A Portion of all parallel execution is serialized
• Communication dependence on uncompleted
  parallel execution
• Conflicts Shared Resources
• Coordination Synchronization with other
  executions
• Overhead For above mechanisms (semaphores,
  locks, etc)
• Speedup 1/(serial (1-serial)/CPUs)
• All parallel computations are uniform and take
  equal time

10
Costs

• The cost of manufacturing facilities doubles
  every generation. In the late 1980s,
  billion-dollar plants seemed like something a
  long way in the future. They seemed almost
  inconceivable. But now, Intel has two plants that
  will cost more than 2.5 billion. If we double it
  for a couple of generations, were looking at 10
  billion plants. I dont think theres any
  industry in the world that builds 10 billion
  plants, although oil refineries probably come
  close
• -Moores Law Repealed, Sort of, in Wired 1997
• In 1995, the cost for a wafer fabrication plant
  was US 1 billion and approximately 1 percent of
  the annual microprocessor market. By 2011, the
  cost will increase to US 70 billion per plant
  and approximately 13 percent of the annual
  microprocessor market

11
Multi-Core The Pit Falls

• Lack of single parallel system organization
• Client/ Server Architectures
• Data Parallel
• Pipelining
• Streaming
• One size does not fit all

12
Lessons from History - Summary

• Moores law will not continue to provide 2x
  performance every other year, using the usual
  techniques in CPU architecture and process
  technology
• It is time that software must play its part to
  provide performance
• Such architecture must support all kind of
  parallelism models
• Parallel programming must move to higher level of
  abstraction to become more pervasive

13
Parallelism through Software - Grid

• Grid infrastructure will provide us with the
  ability to dynamically link together resources as
  an ensemble to support the execution of
  large-scale, resource-intensive, and distributed
  applications.

14
Grid Infrastructure

• Catalysts for Grid Evolution
• Increase in Processing power / storage capacity /
  interconnection capacity
• Cost to own
• Off the shelf components

15
Grid Infrastructure

• Grid is inherently distributed, heterogeneous and
  dynamic
• as compared to cluster, whose presence is
  centralized, resource are homogenous and
  composition is static (including bandwidth).
• Due to high bandwidth connectivity, resource
  sharing is easy

16
Grid Infrastructure

• Resources can reside in multiple domains, and can
  be utilized as per need basis

17
Grid Components - Hardware

• Raw Resources
• Interconnection Networks
• Storage

18
Grid Components Software

• Security / Usage / Fairness Policies
• Scheduling and Resource Management
• Program Execution
• Data movement
• Program migration
• Programming Models

19
Grid Components Stack
20
Grid Workflow
21
Grid Workflow
22
Grid Resource Management

• Resource Discovery
• Resource Allocation
• Resource Maintenance
• Administrative Hierarchy
• Communication Services
• Information Services
• Naming Services
• Distributed File System and Caching
• Security and Authorization
• System Status and Fault Tolerance

23
Grid Resource Management - Globus

• Coined the term Virtual Organization
• Globus was envisioned as a
• system for sharing computational resources
• Resource manager, which is capable of collecting
  resources and matching them with potential
  requests
• For scheduling Globus relies on third party
  schedulers e.g. Condor (-G) and AppLes

24
Grid Resource Management - Globus

• Provides a framework for
• Discovering resources
• Resource Information Base
• Co-Allocation
• Monitoring Resources and Online Control
• Service Level Agreements (SLAs)

25
Grid Resource Management - Globus
RSL specialization
RSL
Application
Information Service
Queries
Info
Ground RSL
Simple ground RSL
Local resource managers
GRAM
GRAM
GRAM
LSF
EASY-LL
NQE
26
Grid Resource Management - Globus

• Advance Reservation Mechanism
• Given dynamic nature of Grid, mechanism ensuring
  (future) availability of resource is necessary
• Service Level Agreements (SLA)
• Resource Service Level Agreements (RSLAs)
• Task Service Level Agreements (TSLAs)
• Binding Service Level Agreements (BSLAs)

27
Grid Scheduling

• How can we execute a set of tasks T, on a set of
  processors P subject to some set of optimizing
  criteria C
• Distributed Vs Parallel
• True test is the support for autonomy of the
  individual node (not communication!)
• Design autonomy, communication autonomy,
  execution autonomy, and administrative autonomy
• By this test, hypercube is parallel machine
• Network of Workstations is distributed

28
Grid Scheduling

• Distributed Architectures
• Network of Workstations (Commodity Grid
  Computing)
• Grid dedicated machines
• Grid of Clusters

29
Grid Scheduling

• Efficient application performance and efficient
  system performance are not necessarily the same
• It may not be possible to obtain optimal
  performance for multiple applications
  simultaneously
• Load balancing may not provide the optimal
  scheduling policy
• Application and system environment must be
  modeled in some detail in order to determine a
  performance-efficient schedule

30
Grid Scheduling

• Authorization Filtering
• Application Definition
• Min. Requirement Filtering

• Advanced Reservation
• Job Submission
• Preparation Tasks
• Monitoring Progress
• Job Completion
• Clean-up Tasks

• Information gathering
• System Selection

31
Grid Scheduling - Condor

• Modern processing environments that consist
  of large collections of workstations
  interconnected by high capacity network raise the
  following challenging question can we satisfy
  the needs of users who need extra capacity
  without lowering the quality of service
  experienced by the owners of under utilized
  workstations? . . . The Condor scheduling system
  is our answer to this question.
• Michael Litzkow, Miron Livny, and Matt
  Mutka, Condor A Hunter of Idle Workstations,
  IEEE 8th Intl. Conf. on Dist. Comp. Sys., June
  1988.

32
Grid Scheduling Condor

• Classads

33
Grid Scheduling Condor

• Matching / Match making / Task Allocation

34
Grid Scheduling Condor

• Schedulers
• Master Worker

35
Grid Scheduling Condor

• Schedulers
• DAGMan

36
Grid Scheduling Condor

• Prepare job to run un-attended Batch processing
• Select the condor run time environment (universe)
  Serial Job, Parallel Jobs, Grid and
  Meta-scheduler
• Create a submit description file
• Submit the job

37
Grid Scheduling Condor

• Preemptive Resume Scheduling
• Take advantage of resources that may only be
  available occasionally
• Handling of job priority
• Fair sharing
• Checkpoint
• Preempt
• run elsewhere

38
Open Issues

• How to study grid ?
• Actual Implementation
• Simulation
• Simulation Emulation
• Managing Workstations / Workers
• Gangalia
• Availability based models
• Failure based models

39
Open Issues

• Resource Managers
• Maintaining state of the resources
• Push based models
• Pull based models
• Sharing Fairness
• Scheduling
• Core of Grid
• Parallel programs come in many flavors
• Parameter-sweep applications (massively parallel)
• Dependency oriented graphs

40
Open Issues

• Application profiling
• Taking out the guess work from predicting the
  application performance on a grid workstation
• Data Staging
• How to minimize data transfer ?
• How to transfer data in minimum amount of time
• Network flow problem
• esp. in data parallel applications

41
Applications

• Geometry-Grid Generation
• Scientific Visualization
• Computational Structural Mechanics
• Computational Electromagnetics
• Computational Fluid Dynamics
• Computational Ocean Modeling
• Computational Chemistry
• Computational Astrophysics
• Computational Biology

42
Acknowledgement

• Introduction part of the presentation is taken
  from Mike Johnsons talk on Jan. 20th 2006 at the
  meeting of IEEE Dallas Chapter