Selfservice, with applications to distributed classifier construction

1
Self-service, with applications to distributed
classifier construction
Michael K. Reiter and Asad Samar
Service Implementations
Our approach

• The Self-Service approach
• Clients perform their own operations
• No resource increase required at server as
  clients increase
• Algorithms guarantee strong consistency
  properties
• Can recover from client failures, though at some
  cost

• Centralized services
• Provides strong consistency, but
• Scales poorly

• Cluster-based services
• Can provide strong consistency, but
• Requires adding resources to the cluster as
  number of clients increases

• Clients are arranged in a tree rooted at server
• Service state divided into small objects
• Each client retrieves the objects required for
  its operation
• And applies the operation locally

• Self-service
• Scales to a large client population (particularly
  compute scalability)
• Does not require adding resources to the server
  as clients increase
• Provides strong consistency guarantees
• Can recover from client failures
• but best suited for services where
• Clients are trustworthy (ongoing work) and churn
  is low
• Service state can be decomposed into small
  objects such that client operations typically
  involve only a few

Experiments
Properties Related Work
Algorithms

• Nodes arrange themselves in a logical distributed
  queue
• And take turns to apply their operations on the
  object

• Implemented algorithms and traffic modeling
  application
• Micro-benchmarks on PlanetLab
• Model building experiments (CPU intensive) on
  local cluster
• Compared with a centralized implementation of
  same service

• Self-Service achieves three properties
• Scalability
• Single clusters are only incrementally scalable
• Consistency
• Most P2P systems do not achieve strong
  consistency properties
• Multiple clusters (one per object) lack
  consistency for multi-object operations
• Failure recovery
• DHTs supporting atomic operations do not recover
  from failures
• Self-Service recovers from failures but may lose
  recent operations

b has requested from a, c from b and d from c. e
now requests from d. a completes operation and
transfers to b.

• If a node disconnects, any objects in its
  sub-tree are reconstituted from latest state seen
  at any connected node
• Connected nodes continue to make progress even if
  a node in the distributed queue gets disconnected

April 27, 2006