next up previous
Next: Name-Based Routing Performance Up: Implementation and Analysis Previous: Content-Layer Overhead

Improved Performance with INRP

The improved performance provided by INRP is illustrated by considering access times to content servers through Akamai versus our proposed content routing. An additional example shows the benefit of using INRP rather than contacting root name servers.

A conventional name lookup of a388.g.akamai.net from Stanford returns the addresses of two content servers which are located 6.6 ms round-trip-time away.

At the next level, the name servers for akamai.net are located throughout the Internet, with an average round-trip times ranging from 12 ms to 93 ms. Overall, this set of name servers has a mean response time of 65 ms and median of 83 ms, ignoring dropped requests.

Using INRP, the same request would go through about 5 content servers (at least one per intervening network), so we will estimate 3 ms extra round-trip time. The direct path to the content servers would then require approximately 10 ms for the name request. A similar example for a miss at the root name servers is carried out in Table 2 for www.cisco.com.

   table116
Table 2: Example name request round-trip times on cache miss (measured from Stanford) for a388.g.akamai.net and www.cisco.com.

As the latency measurements in Table 2 indicate, INRP reduces average request latency in these examples by 86 to 95 percent and also eliminates the variability in latency, providing more predictable performance.



Mark Geoffrey Gritter
Fri Jan 19 09:19:43 PST 2001