Evaluation of a Novel TwoStep Server Selection Metric

1
Evaluation of a Novel Two-Step Server Selection
Metric

• Presented by
• Karthik Lakshminarayanan
• 11-26-2003

2
Problem statement

• Goal Client wants to download content from the
  best of k servers, i.e. minimize total time to
  transfer a document
• Issues to consider
• Cost of choosing the target server
• Lightweight mechanisms preferable
• Stability of ordering (over a period of time)
• More energy can be expended if stability is high
• Nature of content and corresponding workloads
• Frequency of downloads, and size of documents

3
Outline

• Problem statement
• Proposed algorithm
• Existing/possible approaches
• Methodology
• Results

4
Novel two-step server selection

• Pick k best servers out of the entire set by
  using pings (k 5)
• Retain the subset for a period of n days
• Choose servers from the subset of k servers
• Choose from this subset randomly
• Can choose from subset based on other metrics
• Call this Ping-twostep for convenience
• Main delay due to network delays, not server load

5
Selection metrics

• Dynamic metric (adapt to network condition)
• Ping
• Transfer of small files
• Ping-twostep
• Static metric (oblivious to network condition)
• Number of hops
• Number of AS hops
• Random

Summary Ping-twostep performs best!
6
Methodology

• Six client machines (USC, UNC, UCSC, Umass, UDel,
  Purdue)
• 193 servers in tucows.com mirror network
• Collected info continuously for 41 days
• Each run comprised
• 5 ICMP pings
• Traceroute
• Transfer times of files from 10KB 1MB
• More extensive set of servers than previous work

7
Comparison Ping metric

• RTT not always indication of transfer time
• Not surprising!
• Some oddities experienced with
• UNC
• Purdue

• Relative positions between ping 10k vary
  across nodes
• Do not care about the low end of the bw spectrum!

8
Comparison Small file transfers

• Improved with size of transfer
• Low correlation between time for small transfer
  vs. time for large transfers

9
Comparison Static selection

• Hop count
• Mostly equivalent to random selection when used
  to estimate transfer time
• Little correlation (restricted to USA and Canada)

10
Comparison Static selection

• Hop count
• Mostly equivalent to random selection when used
  to estimate transfer time
• Little correlation (restricted to USA and Canada)

11
Comparison Static selection

• Hop count
• Mostly equivalent to random selection when used
  to estimate transfer time
• Little correlation (restricted to USA and Canada)
• AS hop count
• Does not work well for them
• Global IP-Anycast (GIA) uses this
• Queried using BGP
• Small hop counts miss many servers, large hop
  counts would result in too much traffic

12
Stability of server ranking

• 70-98 of changes in rank are between zero and
  ten for top servers
• Average servers experience much higher change in
  rank

Rankings of top servers is stable
13
Stability of server transfer times

• Consider a subset of servers
• How many of them were ever
• at the top in the 41-day period
• Caveat they consider only
• the top server

• Consider different sizes of subsets of 193 hosts
• Number of top servers in an n-subset is a small
  fraction of the size of subset (lt10)
• Little overlap of top servers across clients

14
Ping-random

• Motivation revisited
• Ping technique
• Low overhead
• Good performance
• Top servers stable over time
• Choosing from the small subset
• Random provides load-balance
• Ping use ping again among that set
• Ping-best (for comparison)

15
Performance of Ping-Random

• Ping-ping gt Ping-random gt (10k, Ping)
• Ping-ping might not perform load-balance well

16
Effect of size of ping sets

• Influenced greatly by the size of ping sets
  chosen
• 40 of servers ever ranked first were in 20 of
  the pings

17
Effects of selection algorithms

• Load-balancing
• Different clients have different top servers
• Oscillations
• Respond to changing network conditions
• Fortunately, it is unlikely that many clients
  would be running tests at the same time
• No quantitative results!

18
Discussion

• How do we use this in practice?
• Useful for large file transfers
• What about small web transfers?
• GNP, Geoping approaches might work
• Set of servers is static?
• How can DHTs help in anycast?
• DOLR network for proximity
• Embed location information in Ids
• Use longest-prefix matching tricks (like i3)