GlobalStore: A PRACTICAL FAULTTOLERANT STORE ABSTRACATION

1
GlobalStore A PRACTICAL FAULT-TOLERANT STORE
ABSTRACATION

• Catholic University of Louvain Department of
  Computing Science and Engineering
• Mostafa Al-Metwally
• E-Mail metwally_at_info.ucl.ac.be
• http//www.info.ucl.ac.be/metwally
• Belgium

2
Objectives

• Writing efficient, concurrent, open and
  fault-tolerant distributed applications.
• How to simplify distributed programming?
• How to insulate applications from network and
  failure complexities?
• Solution Shared object store based on
  replication.
• Hiding network latency
• Hiding site and network failures
• Hiding replication management

3
GlobalStore A PRACTICAL FAULT-TOLERANT STORE
ABSTRACATION

• GlobalStore is a logical set of objects which is
  implemented as a replicated set of objects in the
  way that gives the illusion of a single store of
  object.
• GlobalStore is shared, transparent, and coherent
  (as if centralized)
• -Object can reference other store objects.
• -Interface with external world through
  notification
• Any site can update the store by initiating a
  transaction (transactional)
• -Without waiting for network(optimistic,
  speculative)
• Sites (Users) can come and go at any time (open)

4
GlobalStore Logical View
5
GlobalStore A PRACTICAL FAULT-TOLERANT STORE
ABSTRACATION

• As long as at least one site survives, the whole
  store survives (fault tolerant)
• -Replicated object state, no single point of
  failure
• Fault tolerance based only on process redundancy
  (lightweight)
• -Completely detached from file system or other
  persistent storage

6
GlobalStore Architecture View
7
GlobalStore components

• Server (only one)
• Maintain object lock table (transaction
  serialization)
• Help in connection disconnection (virtual
  Membership)
• Is not point of failure new server is created if
  old one fails

• Local Store Client(LSC)
• Maintain local object store per site
• Provide simple API with
• - object creation/deletion
• - transactions
• - notifications
• - connect/disconnect/migration

8
Two concurrent accesses to O
9
Speculative work due to network delay
10
Internal actions in case of failure (informal)

• Client failure (LSC)
• Server side
• Unlock any object(s) for this LSC
• Remove LSC from the membership
• LSC side
• remove this LSC from membership

• Server failure
• LSC side
• Select the minimum id to be the leader
• Send to the leader all locked object(s) and their
  last state(s)
• If you are the leader
• then
• -collect all locked object(s) and last states
  from all LSC(s) -create new server
  -send the server reference to all
  -send last state(s) to all

11
Conclusion and future work

• This talk gives a high level overview of a new
  abstraction for fault-tolerant distributed
  applications.
• Future work
• Solving the problem of network partitioning.
• Provides different concurrency protocols for
  different application domains