This study analyzes the communication traffic on a medium-size Ethernet local area network that connects file servers to diskless workstations. Our measurements differ from those of other studies in two important aspects. First, the Ethernet traffic is much higher than the traffic measured in earlier studies. Second, the measurements are more accurate than any previous ones. For each packet, all the protocol information was extracted and recorded on tapes along with the packet arrival time. Less than one percent of the total number of packets was lost. The clock used to timestamp the packet records had a one-microsecond resolution.

In this paper we describe the measurement methodology, present the traffic statistics, and compare our traffic characteristics with those reported in a 1979 study by Shoch and Hupp. We detail the behavior of each of the protocols responsible for a significant proportion of the total traffic. Our study suggests that protocol behavior is sub-optimal. Traffic patterns are highly skewed: a single client workstation, under normal usage, can generate for and request from a file and paging server data amounting to more than 20 percent of the total raw Ethernet bandwidth. In addition, protocol timers cause numerous, unnecessary retransmissions.

The accurate tracing time has allowed us to study extensively the packet interarrival time and has provided strong grounds for the hypothesis that the arrival process is non-random. Based on projections from our analyses, we conjecture that network bandwidth may be a scarce resource in applications that involve high-performance virtual-memory diskless machines. We believe that a key prerequisite for the success of future diskless workstations will be the design of fast interactive communication protocols that function effectively under high network load.