next up previous
Next: performance measurements Up: Measured interference of network Previous: statement of problem


With most dial-up lines, often not faster than 14.4kbps, variations in interactive performance are extremely visible to users. Many optimizations can improve performance over slow links; examples include Jacobson's header compression [3] and the V.42bis compression [4] used in many modems. Modem compression can raise the effective end-to-end throughput of a 14.4kbps line to 20-30kbps. For example, after the rlogin application sends packets to the network layer, the host sends the packets out via a modem that compresses the data as it transmits it. The benefit of compression is particularly strong for data that exhibits considerable redundancy, e.g., ascii text data. Encrypting at the application layer, e.g., as performed in Kerberos services, makes it virtually impossible for the modem to compress

the data before transmitting it. This essentially limits the system to the raw 14.4kbps that the modem can offer. To confirm this hypothesis we verified that disabling compression on the modem yielded similarly sluggish performance for encrypted as well as unencrypted rlogin streams.

Note that one would likely not notice a performance difference on machines connected via an Ethernet or higher speed medium. Since both the end hosts and the network can perform fast enough without compression, even with competing traffic on the LAN, the optimization function of a modem is unnecessary. This effect also will not manifest itself for very small packets, e.g., the one or very few character packets sent from the client to the server side of a telnet session. Such packets are not ideally amenable to compression since most compression algorithms rely on repeated patterns in the data. The effect of the suboptimal ordering on the end user is most acute with interactive sessions, requiring large packets in at least one direction, over low speed lines.

k claffy
Sat Apr 29 09:10:26 PDT 1995