User-agent: Mozilla Thunderbird 1.0.2 (X11/20050323)
Jari Arkko wrote:
Can you explain why the first model wouldn't be able to use negative
feedback for faster failure detection? It seems like it could use a
trigger from TCP to send a probe earlier than it would otherwise
do it.
I see two potential reasons for not using any negative advice in the
NUD-based model.
1. If the probing by the shim (which is suppressible by positive advice)
have been designed so that a failure can be detected in N seconds, then
the shim would satisfy a requirement of detecting any failure in, at
worst, N seconds. Is it really useful to sometimes be better, or is
worst case the thing that matters? (I guess this is more a question than
a reason.)
2. When you have multiple ULP "connections" using the same shim context,
you have some more complexity to handle when some ULP advice might be
negative and some positive at the same time (for the same context). If
you place more weight in the negative advice, then you will react more
strongly to random packet loss causing one connection to think there is
a problem, etc.
Only using positive advice avoids this complexity, because it implicitly
assumes that if anybody ULP finds the locator pair working, then the
probes can be suppressed.