> But it would be ironic if IPMP's requirements are better satisified by > non-IPMP single interface probe targets than by IPMP probe targets.
Are they better satisfied? In a similar environment without IPMP, the probe target simply becomes unreachable, and the situation never self-corrects. If that's in.mpathd's only target, then it will mark the interface failed, and it will remain that way until someone fixes the target or introduces another target onto the network. > If there is a failure in the other path, it can lead to misdiagnosis on > the sender side. In this case the misdiagnosis is not transient. It would only not be transient if there were no other targets to probe (otherwise, it would simply discard the target and find another). But AFAICT a failure in the path is no different from a failure of the probed node itself. In both cases, an incorrect diagnosis is made. > It could have been a total interface failure of C, not just a transmit > failure. > > in the example of my previous email. The result is still the same. In > this case B still sees responses from D. The logic of mpathd works by > looking at the time of last success of some response on B, versus the > time of first failure observed on A, and it would still misdiagnose the > same failure on A. I'm not sure I understand the case you're talking about, but it seems to me that any scenario that involves C failing is necessarily transient, so I'm not sure it deserves any special handling. > I think a response path that retraces the incoming path induces lesser > misdiagnosis at the other end. I can see that in some simple examples. > Do you see any thing one way or the other on this point ? I agree that retracing the incoming path reduces the risk that a single point of failure on the target (e.g., bad NIC that hasn't been flagged yet) will lead to a transient failure being diagnosed by in.mpathd. But it goes against the idea that IPMP should would equally well regardless of the targets' operating systems, and seems to only matter in pathological conditions where there are not enough targets to probe. So I'm really not a fan. -- meem
