For large files, performance is determined primarily by the speed at which the server can move data to the network. As file size increases, the operating system overhead for user-mode servers accounts for less and less in the overall cost of processing requests. This is because processing latency is completely amortized for large files. For example, Apache lags behind the other Web servers for performance on small files, but is just as good as other user-mode Linux servers for large files. On Windows 2000, IIS performs as well as AFPA for files 128 kB and larger, while it is 3.27 times slower than AFPA for 64 byte files.
For user-mode Web servers, IIS was slower than Zeus for files smaller than 32 kB, but for larger files gained an advantage from having a zero copy send interface. The importance of a zero copy TCP send is further emphasized on the throughput graph. There is almost a threefold performance difference between Web servers using zero copy send interfaces and those using a one-copy send interface. TUX achieved three times the throughput of kHTTPd for large files. We also ran a modified version of AFPA on Linux that does not use the AFPA zero copy architecture. Its throughput on large files was half that of the zero copy version.