Nonce list in PIT entry "infinitely" grows for long-lived Interests
In cases of "Long-lived" interests (interests that are being constantly refreshed, such as ChronoSync status update and some other NLSR updates) set of nonces is never reduces.
Whenever an interest is received, we record its nonce in PIT entry first and only then decide what to do. If PIT entry expires nonce list suppose to go away, but this is not the case with long-lived Interests, where an application will refresh PIT entry almost immediately.
This causing nonce list to include basically all possible nonce combinations for some PIT entries.
Just a small extract from the modified log to confirm the issue:
1405621927.020631 DEBUG: [Forwarder] onIncomingInterest face=18 interest=/ndn/NLSR/sync/5575d5392185c8d5a3f261a2f2a516dfc529f14824471e6ea67e08854a23d599 1405621927.023368 DEBUG: [PitEntry] addNonce 1075754550 (total 179 entries)
My conclusion is that in these cases, Nonce list instead of being used to prevent/detect duplicates, actually harm communication. May be there should be a timer attached to each nonce value so it is not kept "forever", or at least not for the lifetime of the PIT entry. But I'm not sure what could be, at least for the short-term, solution for this problem.
#2 Updated by Beichuan Zhang about 5 years ago
Does the long list of nonces cause any problem right now? performance? correctness?
A quick and dirty fix is to put a limit on the number of nonces kept in the PIT entry.
If nonce is only used for loop detection, then only the nonces of outgoing interests need to be recorded, which should be much less than that of incoming interests. A small number, say 10, may be enough.
If there's a need to keep nonces of incoming interests, would one per incoming face be sufficient?
#3 Updated by Alex Afanasyev about 5 years ago
yes. The problem, I suspect, is right now. After an hour, NLSR interests start being rejected and bad things start to happen.
I asked Hoque to make a test with completely disable loop detection to see if this is the core of the problem or something else.
#5 Updated by Alex Afanasyev about 5 years ago
From the test Hoque did, it seems that the core problem is not this one, though we should fix this one anyways.
I have found another problem with rib self-registrations (#1772) that probably is real reason. Will report more as I have results from new tests.
#6 Updated by Junxiao Shi about 5 years ago
Suppose an Interest is expressed once a second, the probability of having a duplicate Nonce after one month is 0.06%.
Thus I don't think this is a critical problem in terms of correctness.
The growing Nonce list takes memory, which is a problem.
Nonce list for one Interest, expressed once per second, needs 10MB memory after a month.
I agree with the solution in note-2.
The limit could be 256 Nonces (1KB memory).