Scalability Factors

1
Scalability Factors

• Alexander Wolf
• John Mathon

2
Session Goals

• Dimensions of Scale
• Metrics for Dimensions
• Gaining Confidence in Scalability
• Things to Do, Things to Avoid

3
Numbers, Numbers, and More Numbers

• January 1998 Network Wizards Survey
• 29,669,611 hosts in DNS
• 5,331,640 of these reply to ping
• 204 active level 1 domains
• 781,602 level 2 domains
• 10,380,858 level 3 domains
• 1012 messages in stock market domain
• 1010 phone calls for average of 3 min each
• 20,000 events per second

4
Internet Scale vs. Scalability

• Can we characterize what we mean by Internet
  Scale?
• Can we define/build an ISEN service that scales
  (up or down)?

5
Examples of Scale

• News (but not really event notification)
• WWW (but essentially read only)
• DNS (but inherent structure)
• Telephony (but special purpose)
• Issue Internet vs. Wide Area

6
Metrics of Scale

• Number of objects of interests
• Number of observers
• Number of interested parties
• Data storage and queuing
• Number/freq/dist of events
• Number/freq/dist of notifications
• Behavior of parties and objects

7
Achieving Scalability

• What are the starting points for scale? How far
  can the system be stretched?
• No systems acknowledged to be truly (infinitely)
  scalable
• To how many orders of magnitude do we design a
  system to scale?
• It is difficult ahead of time to know the growth
  parameters

8
Dealing with Scale

• Where is the cost borne?
• What is the economic model?
• What are the requirements/scale of different
  classes of applications?

9
Implementation Discussion

• Tree vs. general graph structure of servers
• Is there a structure to subjects to optimize
  routing and filtering?
• Is content-based addressing a requirement?
• All systems may be scalable with sufficient
  expenditure of money

10
Things that Affect Scalability

• Granularity
• message size
• Reliability of notification
• Latency / Quality of Service
• Collateral transactions resulting from events

11
Elvin Inquest

• 20,000 events a second
• Content-based addressing at an Internet scale is
  feasible
• Mesh topology routing
• Not yet tested to Internet scale
• Subscriber is the next hop
• Not sure if linear with the number of leaves

12
Gaining Confidence in Scalability

• Lack of data to parameterize simulation models
• Assertion simulation wont be useful
• John Mathon will publish his guesses (based on
  TIBCO experience)