edw@ius2.cs.cmu.edu.UUCP (05/29/87)
Hi, I'm looking for some information about local area networks, particularly about PRONET, others are of interest too. I work with a group on an application that needs distrubuted computing. The current system configuration we are using consists of 2 sun 3/280, 1 sun 3/75, all connected by Ethernet, and 1 warp super computer - memory mapped into the address space of one of the sun 3/280s. The area of interest for me is the communications between the suns. The current communications consists of Ethernet with TCP/IP protocol running underneath the BSD Unix stream sockets. The communications consist of a large number of small messages of a sizes ranging from 32 to 200 bytes. The network is relatively unloaded - collision monitoring indicates 0 collisions almost always. The message transfer must also be reliable. If a module in the application requests some information then it better get it hence datagrams won't cut it. Profiling system preformance indicates that ~70% of the data transfer time is spent in the Unix system calls sendmsg and recv basically the network primatives, ~15% in ioctl, sigblock and sigsetmask all used to impliment semaphore calls (asynchronous processing is supported in the application) and ~10% of the time is spent in packing and unpacking data buffers (~5% miscellaneous stuff). The effective data transfer rate I see is less than 8k at the sendmsg and recv data transfer level. According to other people, the data transfer rate for TCP/IP should be about 1/10 the band width of Ethernet or 1/10*1.25bytes ~= 100k. The bad preformance here is attributed to the programming paradigm I'm forced to work in which is the large number of small messages. I am interested in how other networks and protocols preform under simular conditions. The criteria the communications must satisfy can be characterized as - needing to be reliable, running on a relatively unload network, take advantage of the short distances apart these machines are, handle a message load that consists of a large number of small messages, and finally it must be fast at least faster than TCP/IP and Ethernet. Does anyone have any figures on some networks or any other sugguestion? Any information would be much appreciated. -- Eddie Wyatt e-mail: edw@ius2.cs.cmu.edu