Enterprise QoS Reality Check

1
Enterprise QoS Reality Check

• Terry Gray
• Director, Networks Distributed Computing
• University of Washington

2
Context

• UW network
• 45,000 machines
• 500 GB/day
• 50 remote locations
• ATM-free zone
• Experts in breaking networks

3
Network Manager Goals

• Meet client expectations by matching capacity
  with demand via
• increasing capacity
• optimizing use of available capacity
• reducing demand
• Avoid becoming a victim of QoS myths...
• Re Admission Control, Reservations, and
  Bandwidth Guarantees

4
Admission Control

• Does not create capacity
• How many fast-busy signals success?
• Makes sense IFF
• BW cost gt BW Management cost
• Denial (busy signals) better than degradation
• Cost model precludes adding capacity
• A tactical solution to a strategic problem
• namely, matching capacity with demand.

5
Reservations

• Require sequestered bandwidth
• Require end-times
• Are useless for small chunk allocations
• Are generally not beloved by users
• Require manual coordination for large chunk
  allocations

6
Bandwidth Guarantees

• In a shared medium, there are no guarantees, only
  probabilities
• P(denial) or P(degradation)
• No free lunch
• If you dont invest in capacity, you will need to
  invest in more technology, but also people to
  develop and manage bandwidth policies

7
Bandwidth Management

• Different strategies needed for different
  congestion zones and timescales...
• Congestion zones
• Subnet
• Backbone
• WAN
• Timescales
• Per packet (Traffic Shaping, DiffServ)
• Per flow or session (Admission Control)
• Persistent (MPLS, DWDM)

8
Three Kinds of Traffic

• Preferred (usually Interactive)
• Best Effort
• Sacrificial

9
Provisioning

• Claim
• You must adequately provision for Preferred and
  Best-Effort traffic, or you will die at the hands
  of the few or the many
• Conclusion
• QoS is a labelling and feedback problem, not a
  signalling and admission control problem

10
Feedback Alternatives

• Admission controls (fast-busy signals)
• explicit short term feedback
• doesn't solve real problem.
• Usage pricing
• implicit long term feedback
• revenue stream for adding capacity
• Social pressure
• e.g. top user lists

11
One Approach

• Combo of provisioning simple diffserv
• Try laissez faire tagging by end-system
• Premium-port strategy if that fails
• Encourage resilient application design

12
QoS Worries

• Increasing network complexity
• Impact on network reliability
• Effects of Policy Jitter
• Beware the prophets of DEN
• WAN QoS accounting
• Building upgrades (cat 3 wireplant)
• Wireless

13
For more info

• http//staff.washington.edu/gray/papers
• gray_at_washington.edu