ken@cs.cornell.edu (Ken Birman) (06/06/91)
Mike Greenberg of GE ATL sent mail asking, basically, what size of message gives the best performance in ISIS. In ISIS V3.0 there is no effective limit. ISIS itself will break things up in an efficient way. In fact, the highest average bandwidths are seen with the largest messages. However, this statement is only true for messages that can be sent in bypass mode (client-member or member-member). As messages get smaller, the UDP size limits begin to enter as a factor. In particular, since ISIS has to put headers on the messages, anything close to the UDP limit of 8k will be broken into two 4k messages. This fragmentation currently occurs around 7500 bytes. If your messages will not be going via bypass, or if they can be consumed incrementally (i.e. useful processing can be done as each chunk arrives) you would be best off sending less than the biggest possible messages, since otherwise things will be bursty -- you will compute, then ISIS will sit and transmit data for a while, then it will finally hand off this huge sucker to the LISP application which will suddenly crunch for a while. Very bursty behavior tends to be inefficient. Worse, in non-bypass communication (i.e. nearly all communication in V2.1) very large messages can cause ISIS itself to congest. This kicks in various heuristic flow control mechanisms and can actually make the system pretty sick if your destination process doesn't cooperate and drain data out of ISIS at a pretty hefty rate. To summarize: think in terms of splitting large objects into messages such that the application can do something useful on each incoming message. scheme will probably work best on V3.0 and should work under V2.1 too. However, where there is really no upper limit on message size in V3.0, the congestion problem puts an effective upper limit on message size and data rates in V2.1. The actual numbers at which these factors become significant will depend on the environment in which you use ISIS. In V2.1 make your messages about 4k long. -- Kenneth P. Birman E-mail: ken@cs.cornell.edu 4105 Upson Hall, Dept. of Computer Science TEL: 607 255-9199 (office) Cornell University Ithaca, NY 14853 (USA) FAX: 607 255-4428