On the Performance Benefits of Multihoming Route Control
Mon, 03/17/2008 - 18:19 by Damien Saucez
The paper "On the Performance Benefits of Multihoming Route Control" written by A Akella et al. in IEEE/ACM Transactions on Networking February 2008 estimates the possible gain of performances while correctly route traffic in multihomed enterprise environments.
The paper focus on two metrics: the RTT (called turnaround time) and the throughput of http-based downloads.
The main result of the paper is that the selection of the best ISP among different choices can dramatically increase performances for a small set of destinations even if the overall performances are only modestly improved.
The main interest of the paper is the introduction of different ideas. First, two new metrics are introduced to estimate the benefit in term of RTT and throughput (section II B.). The principle is simple: present a function that is the sum of the best possible value of the metric and the value of the metric for each possible choice and normalize it. The two metrics are really readable when plotted.
In addition to the new metrics, the authors highlight the most important questions for multihoming path selection techniques:
- "How should end-networks monitor the performances of the ISP links? Which destination to monitor and how?"
- "How should the end network estimate the future performance of an ISP to a destination?"
- "How should the end-network direct traffic to use the chosen ISP links for a destination?"
- "How to accommodate temporary shifts in popularity (e.g., flash crowds)?"
- "Active/Passive/Hybrid measurements?"
- "Which destination monitor?"
To evaluate the proposition, the authors have implemented a web-proxy in conjunction with NAT that estimates the performances of the most important destination. The implementation uses NAT to move traffic to the correct link.
Figure 9 introduces a nice ISP monitoring scheme for the web-proxy to estimate the performances of the different ISPs for the most relevant destinations.
The destinations to monitor during a period T are selected according to two different schemes. (i) FrequencyCounts maintains the frequency of the destination in the request. This technique avoid estimating destination occasionally encountered but is really bad for ephemeral events like flash crowds. With (ii) SlidingWindow, measurements are performed only for destinations that appear within a sliding window. The technique is interesting for sudden events but may forget some frequently destinations if they not appear within the period of the window.
This paper is related to our IDIPS researched (http://inl.info.ucl.ac.be/idips).