Service Composition: Breakout Session Summary

1
Service Composition Breakout Session Summary

• Randy Katz
• David Culler
• Summary Bhaskaran Raman

2
Discussion topics

• Background
• Emerging Internet service model
• Confederated vs. Overlay model
• Issues
• Granularity of composition
• Growth trends copious bandwidth
• Service economics

3
Background

• Emerging Internet service model
• Structure-less cloud ? Structured, with Internet
  data centers
• Portal service built out of pieces
• Portal ? Appln infra service ? Appln specific
  servers ? Overlay networks ? Physical topology
• Confederated vs. Overlay model
• At the network layer
• At the service layer

4
Granularity of services

• Heavy-weight pieces
• E.g., search engines
• Light-weight pieces
• E.g., transformation agents
• Yahoo portal is an example of heavy-weight
  composition
• Ninja paths light-weight composition

5
Composition beyond functionality

• Service composition flexibile
• Can a composed service be robust?
• Can we build in a negative feedback loop into the
  system?
• With an event queue based approach to
  composition?
• Can such a mechanism work across machines, across
  the network?
• How to compute reliability of a composed service?
• Who to blame when something breaks?
• SLAs for services complex
• Yahoo/Google have legal SLA document

6
Design for decomposition

• Design for decomposition how to design a service
  so that it can be built from pieces meaningfully
• End-2-End argument in service composition
• If we want to have admission control at the end,
  how to effect this when the bottleneck is in the
  middle?
• How to design protocols to allow for composition?
• Expose more information
• E.g., HTTP/1.0 has little support for caches,
  HTTP/1.1 has a lot of support

7
Internet Growth Trends

• Optical technology being deployed in a big way
• 1.6 Tb/s 24 Tb/s in the backbone
• Where to place services, and how many service
  instances?
• Does it matter, if backbone bandwidth is free and
  infinite?
• What about the people cost of managing
  distributed instances?
• In storage management, people cost is 3-4 times
  more than system cost
• Network access cost dropping can buy optical
  link for 2K a month

8
Service distribution vs. Centralization

• Might depend on granularity
• Hard to distribute search engine
• Easier to distribute Akamai servers soft-state
• People cost
• HP 50,000 node computer network
• 10K per server
• One person can manage 10 servers
• Service in the telephone network are distributed
• Might be related to the way the telephone network
  evolved, and its architecture

9
Service distribution vs. Centralization

• Latency/server-load might be a reason to
  distribute a service
• There are congested parts of the Internet today
• Trans-oceanic links
• Public-peering points
• Distribution
• Over-provisioning may be easier if there are many
  sites

10
Service Economics

• Composition economics meets engineering
• SLAs are an important part
• Trust relations
• Management tools for services
• What are the ones required for composed services
• Diagnose and isolate problems
• Composition by user (dynamic) vs. composition by
  service providers (less dynamic)