Combining Optimism and Pessimism in Accessing Replicas in Distributed Systems

1
Combining Optimism and Pessimism in Accessing
Replicas in Distributed Systems

• Joel M. Crichlow

2
INTRODUCTION

• The system contains a network of computing nodes.
  
• The data that can be accessed can have any level
  of replication across the nodes.
• Replication is employed in order to increase
  availability.
• Replication increases the need for effective
  control measures to preserve some level of mutual
  consistency.
• The scheme under discussion combines Optimism and
  Pessimism.

3
Introduction
Front End
x,y,z
x,y,z
x,y,z
x,y
x
4
Rationale

• Simplicity KIS (note only one S)
• Increasingly more powerful computers
• Faster networks
• Wireless?

5
Objectives

• Provide a high level of availability at a known
  penalty to the application
• Preserve data integrity
• Build a system that is conceptually simple

6
The Cost Bound

• Maximum Numerical Error
• Determines the extent to which a replica can act
  independently
• We are considering resources that can be counted
  and are of the same type

7
The Cost Bound

• A request for r resources can be satisfied if the
  reachable pool is controlled by a cost bound of C
  resources and r ? C.
• We can split the cost bound C among n replicas
  such that each replica i has local access to ci
  resources where

8
The Cost Bound

• Application characteristics can permit temporary
  over-allocation of resources (like over-booking),
  hence, if we let R be the no. of resources
  initially available, then a cost bound ci can be
  set at each replica i such that

9
The Cost Bound

• We propose a dynamic C value which changes in two
  ways
• In response to pessimistic global processing,
  and
• In response to optimistic local processing.

10
Dynamic C

• On pessimistic run
• Cnew is new cost bound
• Cinit is cost bound initially set
• RA is no. of resources allocated via requests
• R is no. of resources initially available.

11
Dynamic C

• On pessimistic run
• where
• Let the resources allocated at replica i after a
  pessimistic cycle be RAi then
• Then let

12
Dynamic C

• On optimistic run
• where

13
Testing

• There are published performance results for a
  prototype first coded in C and now coded in Java.
• In each cycle we used wi 1/n.
• Due to the expert work of Prof. Steve Hartley the
  system now runs on a Linux four-node platform in
  Rob.308.
• Thanks Steve!!!

14
Testing

• Three replicas process transactions that are
  generated by a transaction generator located at a
  remote computer.

15
Testing

• Each transaction is expressed as a parent and
  child transactions.
• The parent remains at the owner while the
  children are sent to all the replicas.
• Parents are processed pessimistically.
• Children are processed optimistically.

16
Testing

• Link to recent paper.

17
Conclusion

• A system has been designed to investigate the use
  of optimism and pessimism in accessing replicas
  in a distributed system. The viability of the
  system has been established and a Java test bed
  is being used to examine alternative features.