7
A Suboptimal Solution to Optimizing Cost and Performance in OSPs Presented by: Daniel Burgener Irene Haque

A Suboptimal Solution to Optimizing Cost and Performance in OSPs Presented by: Daniel Burgener Irene Haque

Embed Size (px)

Citation preview

Page 1: A Suboptimal Solution to Optimizing Cost and Performance in OSPs Presented by: Daniel Burgener Irene Haque

A Suboptimal Solution to Optimizing Cost and Performance

in OSPs

Presented by:Daniel Burgener

Irene Haque

Page 2: A Suboptimal Solution to Optimizing Cost and Performance in OSPs Presented by: Daniel Burgener Irene Haque

Time overhead Only look at RTT and cost, not bandwidth and

cost rate, which they leave for future work Optimizing these is necessary Adding each of these factors will increase

computational time demands significantly These also require more probing to test

They calculate at 500 different wRTT values in the paper, assuming similar increments for these factors: Total time = (time for cost+wRTT) x (bandwidths

studied) x (loss rates studied) Total time = 9s x 500 x 500 = 26 days They allot 1 hour for this calculation!

Remember that 9s is only for 20k prefixes and 2 routes

Page 3: A Suboptimal Solution to Optimizing Cost and Performance in OSPs Presented by: Daniel Burgener Irene Haque

Accuracy Issues

RTT will not remain accurate under changing bandwidth Since Microsoft wouldn't let them change routes

despite having four authors on the paper, these effects are untested

Throw out 98% of prefixes on Internet Limit themselves to top 10% of prefixes which have

90% of traffic Throw out 80% of these due to insufficient live IPs and

multi-location prefixes

Page 4: A Suboptimal Solution to Optimizing Cost and Performance in OSPs Presented by: Daniel Burgener Irene Haque

Speaking of multi-location prefixes...

They use 5ms as a cutoff threshold Where does this come from? Cited source

(“Geographic Locality of IP prefixes”) doesn't mention it

“some /24 prefixes span distances of more than 10,000 miles”

Other researchers have established the speed of data in the Internet as 2/3c

5ms * 2/3c = 0.005s * 3*10^8 m/s = 1000km Source: R. Percacci and A. Vespignani. Scale-free behavior of the

internet global performance. The European Physical Journal B - Condensed Matter, 2003.

Page 5: A Suboptimal Solution to Optimizing Cost and Performance in OSPs Presented by: Daniel Burgener Irene Haque

Speaking of multi-location prefixes...

1000km is roughly the distance from Chicago to Rochester, NY

Page 6: A Suboptimal Solution to Optimizing Cost and Performance in OSPs Presented by: Daniel Burgener Irene Haque

Furthermore,

This 5ms is NOT the distance between two locations

It is the difference in distance between them and the source

Page 7: A Suboptimal Solution to Optimizing Cost and Performance in OSPs Presented by: Daniel Burgener Irene Haque

Why does this matter?

The fact that multi-location prefixes aren't properly determined undermines the entire paper They are optimizing routes to one location and

assuming its the same for other nodes in that prefix The “experiment” is actually a simulation