MEASUREMENTS OF THE HOPCOUNT IN INTERNET Measurements of the Hopcount in Internet F. Begtaševiü and P. Van Mieghem End-to-end quality of service is likely to depend traceroute utility among 37 Internet sites. His on the hopcount, the number of traversed goal was to study the routing behavior routers. In this paper we present the results of including pathologies, stability and path the measurements of the hopcount from a source symmetry. His data showed that the paths for at Delft towards several destinations spread over most of the sites (almost two-thirds) stayed the three continents. Fitting the data with our same over longer period of time. About half of theoretical model illustrates that the number of routers based on our measurements is estimated the paths were asymmetric. at around 105. The changes in the hopcount over Vanhastel et al [7] published their results of time are investigated. It was found that though the measurements performed at the University there is virtually no short-term change in the of Gent in 1998, but the number of destination average hopcount the individual hopcounts do sites they used is rather small. Moreover, they change significantly. Rapid changes in the paths did not study the changes in the hopcount nor were observed to barely affect the hopcount. in that of the paths. PingER is a large end-to-end performance monitoring project [4], organized by the I. INTRODUCTION researchers from different institutes working In order to understand the current behavior of on High Energy Physics experiments the Internet better, the properties of qualifiers (HEPnet). PingER performs active measure- such as the hopcount between two arbitrary ments using ICMP echo requests sent at low nodes and the corresponding end-to-end QoS frequency to approximately 500 sites. It (delay, jitter and packet loss) may enhance our collects information on packet loss and RTT. insight. In this paper, we confine ourselves to Other projects on end-to-end performance the hopcount, apparently the simplest of the measurements are Surveyor and NLANR’s above mentioned qualifiers. (National Laboratory for Applied Network Knowledge of the hopcount may help to Research) Active Measurement Program unravel the topology and to dimension and (AMP). These programs are more community- propose a more efficient network infrastructure specific, which means that the number of sites than that of the current Internet. This participating in these measurements is limited knowledge is also beneficial to simulate more and that the measurements are done over the realistic network topologies. By combining the same set of sites. delay estimates and results of the hopcount The subject of this paper is to present the measurements, we expect to be able to specify measurements of hopcount for different how to guarantee the Quality of Service in regions, especially the distribution of the current Internet. hopcount for the designated regions and In recent years there were few attempts to get a possible differences between regions were better picture of the Internet topology. End-to- investigated. Furthermore, we have studied the end performance studies were rare. Most of the change in the average hopcount and the topology related measurements were based on changes of the paths to the sites. These the hop-limited probes with the aim to create a measurements were done using the traceroute router level map of Internet. Performance program. Although this tool has already been related measurements mainly involved used in most of the above mentioned studies, measurements of RTT’s between pairs of hosts our approach to the measurements in order to investigate how the RTT changes complements these results. The changes of the depending on the path, time of day, amount of average hopcount over different periods of traffic and other parameters. time were examined and the dynamics of these In 1995 Paxson [5] analyzed 40,000 end-to- changes was investigated. Also the changes of end path measurements, made by repeating the the paths to the sites were studied as well as Delft University of Technology, Information Technology and Systems, P.O. Box 5031, 2600 GA Delft, The Netherlands Corresponding author: [email protected] MEASUREMENTS OF THE HOPCOUNT IN INTERNET their influence on the results for the hopcount hopcount per continent and the probability measurements were studied. distribution function (pdf) of the hopcount (per In section II we start by presenting the set up continent and for the total sample). That of the measurement configuration and give an section also contains the study of the changes explanation of the procedure of gathering the in the hopcount and the paths during these addresses of the destination sites. In section III measurements. Finally there is a short we present the results of the measurements of discussion of our results. the short-term changes of the hopcount, SURFNET (Internet) Unix workstation gateway Dunet3 router gateway Dunet0 router BR1.Delft.surf.net Cisco router 7200 Cisco router 7513 Cisco router 12000 130.161.40.130 130.161.32.1 130.161.1.52 145.41.17.129 Figure 1: The measurement set-up While collecting addresses for our sets, we I. MEASUREMENT SET-UP AND DESTINATION encountered a number of problems. SITES The first problem was that many of the sites turned out to be unsuitable for the measurements due to the fact that many of the A. Set-up routers are programmed to discard ICMP For the measurements of the hopcount we have messages used by the traceroute program. used the standard UNIX tool traceroute [6] on The second problem was the geographical the dutetva (130.161.40.130) SUN-workstation distribution of the sites. We discovered that a at the Delft University of Technology. To certain number of the sites belonging to the reach the public Internet from dutetva packets European and Asian institutions (BBC for always make the same three hops as illustrated example) were actually situated in other parts in Figure 1. of the world (mostly in the United States). Two of these machines belong to the network Hence they must be excluded from the of the University of Technology (speed of their measurements. This was the case with a connection is 622 Mb/s), while the third limited number of sites. Despite the fact that (though physically in Delft) belongs to the some of the hostnames referred to a certain SURF-Net, the network of the Dutch country, some sites were actually situated in Universities (the TU-network is connected to another (e.g. sonera.be a name referring to the SURFNET via a 155 Mb/s link). Belgium is actually situated in Finland and many Chinese sites are placed in Hong Kong or Singapore). The problem of geographical B. Destination sites distribution was also solved after consulting We decided to concentrate on performing the the databases of the local Internet registries. hopcount measurements for three continents – After checking every address from our sets in Europe, Northern America and Asia-Pacific one of the databases, approximate (Far East, Australia and New Zealand). This geographical locations of sites were choice was made on the importance of these established. RIPE (Réseaux IP Européens) is regions from the point of view of a European one of the three regional Internet registries that user. The most time-consuming part of these exist today. It provides allocation of Internet tests was finding enough suitable destination resources in Europe (and surrounding addresses. While collecting addresses of countries). Its database contains a list of more suitable sites, preference to large companies, than ten million hosts in this area together with Internet Service Providers and universities was their IP numbers and addresses. Other regional given. If possible, universities were chosen registries are APNIC (Asia Pacific Network because they provided a better geographical Information Center) and ARIN (American distribution of the sites and a more responsive Registry for Internet Numbers) and their collaboration. databases were used during the creation of the sets of addresses for these regions. For MEASUREMENTS OF THE HOPCOUNT IN INTERNET European regions we tried a certain number of region, important countries (Australia, New hosts (up to twenty) per country in order to get Zealand, Japan, China) where represented with a better uniform spread of the average more sites than other countries. While hopcount in Europe. The choice of the collecting the addresses for North American countries was made with respect to size and we tried to include in our set at least four sites economical power. We mention that the set of from every state in the US and Canada with European addresses does not contain any site (naturally more important states like California in Netherlands. A similar approach was made and New York were represented with more while gathering addresses for the Pacific sites than states like Delaware or Idaho). 14,50 14,40 14,30 14,20 14,10 14,00 13,90 13,80 13,70 13,60 0:04 1:06 2:07 3:05 4:03 5:11 6:15 7:16 8:29 9:40 0:00 1:09 2:25 3:54 5:17 6:39 8:00 9:21 11:49 13:34 15:17 16:57 19:16 20:39 21:53 23:00 11:13 12:54 14:38 16:15 17:48 19:13 20:31 21:42 22:54 Figure 2 Average hopcount European set of addresses measured between 10th and 12th of April 2000 The third problem was caused by the fact that addresses distributed over the area. The one hostname can refer to the different resulting set of American sites contained 274 machines and that after some time the original addresses and was significantly larger than the hostname could be assigned to a different other two samples.