ken@gvax.cs.cornell.edu (Ken Birman) (02/02/90)
This seems like a popular proposal; I've now received 8 notes from
people who think it sounds really useful.
As a short-term hack I propose the following (for V2.0):
* I assume that applications always run on UNIX systems.
* Connections are by the ISIS UDP transport protocol (fast and
has less buffering overhead than TCP, so you can have more of them)
* No transparent rebinding if the mother machine fails, but your
application can catch this event and rebind "by hand".
As noted before, the bypass protocols will run directly from process
to process, without indirecting via protos. So, you can get the full
benefit of it in this setting. The payoff will be highest if a small
number of processes do all the multicasting and the other group members
just receive.
This should be easy to add and I will have it in the system for
the March beta release. I hadn't appreciated how many systems have
a "server" with "clients" structure, with ISIS useful mostly to build
reliable servers and clients running mostly standalone workstations.
Are there other mechanisms that people would like to see for this setting?
Lucky I was planning to rewrite the manual: at this rate, the architecture
of V2.0 will differ drastically from V1.xxx (upward compatible, though!)