Observations on What is ``Close''

As described previously, the current version of SARA relies on the user to choose the ``closest'' data server. Typically, many users assume that transfer times increase with geographical distance. That is, the farther away a server is on a map, the longer data transfers will take. However, it is becoming increasingly true that network characteristics between two sites depend less on actual distance and more on the media which connect them. Our Simple SARA experiments led to several interesting observations about what is ``close'' on the network:

• ``Close'' network sites may have little correlation with ``close'' geographical sites.
  Fast networks, like the vBNS, make it difficult to determine ``closeness'' based solely on geographical distance. Sites that may be separated by thousands of miles may be ``close'' in the network sense, whereas sites separated by only a few hundred miles may be ``far''.

  Consider experiments run between UCSD, the University of Washington (which was on the general Internet), Georgia Tech (which was on the vBNS), and the University of Illinois (also on the vBNS). During the time frame shown in Figure 4, data files from both Georgia Tech and UIUC enjoy more bandwidth than data files from UW (on average 36.9% and 42.4% respectively) and thus, arrive at UCSD sooner. In this experiment, the AppLeS agent, with the help of NWS forecasts, is able to detect that the UW server, although geographically the closest to UCSD, is not the site that yields the shortest network transfer time.

  Figure 4: Experiments between UCSD, UW, GeorgiaTech, and UIUC on 8/23/98.

  

• Dynamic factors influence what is ``close''.
  In the Internet, network links may be added, removed, and changed on a daily, and often more frequent basis. In addition, the great majority of network links are used by multiple users at any given time. All of these factors contribute to wide fluctuations in the performance of the network. As this happens, the relative network ``distance'' between sites can change dramatically.

  Although both sitar.cs.uiuc.edu and lolland.cc.gatech.edu are supposedly connected to the vBNS, we found that Georgia Tech's vBNS connection was subject to intermittent failures during the time frame of our experiments. Figure 4 shows that bandwidth from both of these two machines is roughly the same. However, at trial 17, we see that bandwidth to the Georgia Tech server is significantly increased, and we conjecture that this is due to the vBNS link being re-established.⁸ The NWS measurements and forecasts correctly detect the change allowing the AppLeS agent to profitably switch servers.

  Also, consider the experiments performed in the morning on 9/21/98 between CalTech, the University of Utah, and UCSD (Figure 5). During this time frame, the general Internet (to CalTech) provided better performance until trial 26 (approximately 9 AM), at which time the webcast of the Grand Jury testimony of President Clinton commenced. At this point, bandwidth on the general Internet began to fluctuate greatly, and the vBNS connection to Utah provided superior performance.

  Figure 5: Experiments on 9/21/98 between UCSD, CalTech and University of Utah.

  

• What is ``close'' may not be obvious to the user.
  Whereas a user can fairly easily determine the geographical distance between two points, network ``closeness'' is often an opaque concept to the user. Because of considerable performance variation, and the difficulty in determining the difference between ``noise'' and ``trends'' in network performance, it is not straightforward for users to determine which network link will provide the best performance for their application [14].