FIBRE CHANNEL PERFORMANCES WITH IBM EQUIPMENT

IV. Communication Benchmarks

• - Overhead: it is defined as the time from when an application sends a message to the time when it is ready to send another message. In practice, it corresponds to the measurement of the time taken for performing the write of a one byte message (the smallest message that can be written).
  

• - Communication latency: it is defined as the time from when an application (on the sender side) sends a message to the time when another application (on the receiver side) receives this message. In practice, we measure a round trip (RT) communication latency for a one byte message, and then divide it by two.
  

• - Throughput (or transmission rate): it is defined as the amount of data that can be sent within one second.
  

The program used to test out communication performances, when using TCP/IP protocols, was Netperf. On all machines, we configured the High-Speed TCP/IP Extension (rfc 1323) to increase the performance of TCP/IP over Fibre Channel. Other appropriate network options were modified to fully utilize the advantages of rfc 1323, like for instance the maximum socket buffer size (sb_max), the window size on both receiver and sender sides (tcp_sendspace, tcp_revcspace), ... [see appendix on how to configure network options].

The TCP_no_delay option (-D) was set on the sender side: it is necessary for relevant measurements because it disables the so-called Nagle mechanism for the buffering (it avoids buffering and grouping messages before sending). Sometimes, we got some troubles when using this option turned on and we were obliged to turn it off (only in case of sending large messages with large socket buffers).

For the light-weight protocol (Raw Socket), no program was available from the public domain; therefore we used and ran an IBM hand-made application, which was evolved jointly by Carlo Miron and Massimo Bernaschi.

Generated with CERN WebMaker