Check out the new USENIX Web site. next up previous
Next: Testbeds Up: 4.4 Performance Previous: 4.4 Performance


Latency and throughput are the most fundamental measures used to evaluate the performance of networks. The time for a small transfer is dominated by latency, while that for a large transfer is dominated by throughput. Interactivity, which is often dominated by small transfers, suffers if latencies are either high or highly variable [8]. Bulk transfers suffer if throughput is low. Our measurements were conducted on working days (Monday through Thursday) in the early morning to eliminate time-of-day effects.

Latency: To measure latency, we used the round-trip delay of an ICMP echo request/response pair (i.e., ping), taking samples over hour-long intervals. We computed the average, minimum, maximum and standard deviation of these measurements. Here, we report the average and standard deviation. Notice that this measure of latency is symmetric.

Throughput: To measure average throughput, we use the ttcp program. Ttcp is commonly used to test TCP and UDP performance in IP networks. Ttcp times the transmission and reception of data between two systems. We use a socket buffer size of 64 KBytes and transfer a total of 1 GB of data in each test. VNET's TCP connection also uses a socket buffer size of 64 KBytes. TCP socket buffer size can limit performance if it is less than the bandwidth-delay product of the network path, hence our larger-than-default buffers. All throughput measurements were performed in both directions.

next up previous
Next: Testbeds Up: 4.4 Performance Previous: 4.4 Performance
Ananth Sundararaj 2004-02-17