[1] A. Akella, B. Maggs, S. Seshan, A. Shaikh, R. Sitaraman, "A Measurement-Based Analysis of Multihoming", ACM SIGCOMM 2003. paper

Summary:  This paper is about advantages of Multihoming. The authors study the impact of multihoming on performance and reliability of networks and conclude that multihoming improves performance and reliability significantly if right set of providers is chosen.  

• Multihoming is defined as a customer network having more than one external link, either to a single ISP, or to different providers.

• Enterprise Perspective vs. Web Server Perspective

  • Enterprise Perspective: Routing the data being downloaded through appropriate ISP.

  • Web Server Perspective: Routing the data being provided (uploaded) through appropriate ISP.

• Motivation of the paper

  • Does smart routing improve performance?

  • Does choice of ISPs matter?

• Performance testing

  • Enterprise Perspective: 2-multihoming and K-multihoming

    • 2-multihoming set up (Data set A1): This data set contains statistics collected by 27 geographically distributed Akamai monitoring nodes. One or two of these monitoring nodes are located in major cities in the U.S., with multiple nodes in the same city attached to different upstream provider networks. Every 6 minutes, on average, these nodes download designated objects directly from Akamai customers. The turnaround time is noted.

    • K-multihoming set up (Data set H1): For this data set, instead of using monitoring nodes, they use multiple Akamai servers per city connected to different upstream provider networks. These servers periodically download designated objects directly from Akamai customers, and the average turnaround times for requests made by Akamai servers back to the origin content provider servers is noted.

    • The graphs show a varying degree of performance improvement, and there is no significant improvement after K=4 or 5. Knowing best ISP in advance is very important. 

  • Web server Perspective: 2-multihoming and K-multihoming

    • 2-multihoming and K-multihoming set up: There are K different Akamai servers that are connected to different ISPs. Every 6 min. each server downloads 50 KB object from the other Akamai servers. The turnaround time is noted.

    • The graphs shows incremental improvement as K increases up to K=4.

  • Reliability:

    • Reliability performance set up: They collected trace route data from 50 geographically diverse keynote system nodes to Akamai servers. There are 2 keynote nodes per city each singly homed to different provider. 20 Akamai servers from each city were used. These servers were connected to the top 20 ISPs in that city. This data set provides information about IP-level connectivity of the network and describes robustness to IP-level failures.

    • The results show that there is a significant difference in optimal, average, and worse cases of reliability. Hence choice of ISPs does matter.  

• Class discussion

  • A significant contribution.