In a previous study we considered the extent to which multihoming can be leveraged by enterprises and Internet data centers to improve network performance [4]. This measurement-based study revealed the maximum benefits attainable in a variety of scenarios, but provided only limited guidance as to how to extract those improvements. In this paper, we perform an experimental evaluation of a number of practical techniques for using multihoming to improve performance.
In a study closely related to ours, the authors conduct a few trace-driven experiments to evaluate several design options using a commercial multihoming device [9,17]. The evaluation focuses on the ability of several algorithms to balance load over multiple broadband-class links to provide service similar to a single higher-bandwidth link. The authors find that the effectiveness of hash-based link selection (i.e., hashing on packet header fields) in balancing load is comparable to load-based selection. In addition, their results show that managing load at a connection-level granularity is only slightly less effective than per-packet load balancing. They also show that using knowledge of the asymmetric nature of some applications (e.g., Web connections) can be useful in improving traffic balance, although it requires additional application-specific information.
A number of vendors have recently developed dedicated networking appliances [15,7,14] or software stacks [21,16] for optimizing the use of multihomed connectivity in enterprises settings where BGP is not used. Most of these products use techniques similar to those we evaluate in our study, though their focus is geared more toward balancing load and managing bandwidth costs across multiple ISP links, rather than optimizing performance. All of these use NAT-based control of inbound traffic and DNS to influence links used by external client-initiated connections. They also ensure, by tracking sessions or using policy-based routing, that the same ISP link is used in both directions.
Another class of products and services are targeted at settings where BGP is employed, for examples large data centers or campuses [18,1]. These products mainly focus on outbound control of routes and, as such, are more suited for content providers which primarily source data. Details of the algorithms used by any of the above commercial products to monitor link performance or availability are generally proprietary, and little information is available on specific mechanisms or parameter settings. Here, we review the general approaches taken in enterprise route control products.
Most commercial products employ both ICMP ping and TCP active probes to continuously monitor the health of upstream links, enabling rapid response to failure. In some cases, hybrid passive and active monitoring is used to track link performance. For example, when a connection to a previously unseen destination is initiated from an enterprise client, active probes across the candidate links sample performance to the destination. Connections to known destinations, on the other hand, are monitored passively to update performance samples. Another approach is to use active probing for monitoring link availability, and passive monitoring for performance sampling. Some products also allow static rules to dictate which link to use to reach known destinations networks.
Finally, some products use ``race''-based performance measurements, in which SYN packets sent by enterprise clients to initiate connections are replicated by the route control device on all upstream ISPs (using source NAT). The link on which the corresponding SYN-ACK arrives from the server is used for the remainder of the connection. The route control device sends RST packets along the slower paths so that the server can terminate the in-progress connection establishment state. The choice of best link is cached for some time so that subsequent connections that arrive within a short time period need not trigger a new race unless a link failure is detected.