Trace Data Sample Clauses

Trace Data. The performance of algorithms of this type depends heavily on the traffic characteristics. It was essential to use real traffic characteristics and we achieved this by collecting traffic traces for driving our simulation. We chose the National Laboratory for Applied Network Research (NLANR) because it has a very large network and also a large measurement infrastructure already in place1. The NLANR site gathers and maintains raw packet header data from different points in their network on a daily basis. They gather data in several different formats. We decided to use the 24-byte FR format which included the timestamp (seconds and microseconds), source and destination IP and port, packet size, IP protocol and TCP flags. All the fields except the IP addresses had the original values from the packet header. For security purposes the IP address had to be made completely irreversible using a one-way hash function. This meant that multiple trace data sets could not be used simultaneously to generate a single profile. Before encryption of the IP address we had mapped it to its corresponding BGP address prefix. We refer to the header data set that was gathered at the SDSC (University of California San Diego) measurement points of the NLANR infrastructure. The 778MB trace data was gathered for about 3hrs and 40 minutes at the location on 7th June 1999. In all it had 1247 different source and destination domains whereas there were a total of 63625 address prefixes. Among the 1247 active domains, only 30 or so domains contributed more than 90% of the traffic. We looked at the traffic destined to some of highly active destination domains and we saw that fewer than 10 source domains were almost always responsible for over 90% of the traffic destined to each domain. Generating the traffic profile from the trace data required the mapping of the domains into a common ingress node. We relied on an approximate solution by assuming a value for the number of ingress nodes to the network and a value for the bandwidth available at each ingress link. Using the rules that • if traffic exists between two domains, then those domains must be having different ingress nodes to the network • if traffic does not exist between two domains, then these domains are very likely to belong to the same ingress • the total traffic coming from an ingress node should not exceed the assumed bandwidth for any interval of time • the domains were divided among the assumed 8 ingress nodes Thus we had as...