Feature #1282
open
Design duplicate suppression on multicast face
20%
Description
Design a Data packet duplicate suppression mechanism on multicast face.
ccnd's algorithm is:
- Multicast datagram face should delay every outgoing Data packet for a random duration (on the order of 10ms).
- During the delay of a packet, if an identical packet is received on the face, cancel the send.
but it causes packet reordering and cannot adapt to link condition.
This Task is to design a better algorithm for duplicate suppression.
Category is currently "Faces", but the design MAY propose to put part of this feature in Forwarding/Tables if necessary.
Files
Updated by Junxiao Shi over 10 years ago
20140220 conference call decides:
- Duplicate suppression applies to Data packets only, because delaying Interests causes too much end-to-end delay.
- Duplicate suppression with random delay will cause packet reorder. Congestion control scheme should not expect in-order delivery.
Updated by Davide Pesavento over 10 years ago
1/ I think this can be postponed until after the first release... it's just an optimization anyway.
2/ Duplicate suppression applies to multicast Ethernet faces too, so it should be factored out in a common implementation, if possible.
Updated by Junxiao Shi over 10 years ago
- Target version changed from v0.1 to v0.2
20140227 conference call agrees to this MAY be deferred to Version 2.
Updated by Junxiao Shi over 10 years ago
- Assignee deleted (
Giulio Grassi) - Target version deleted (
v0.2)
20140514 call decides:
ccnd's algorithm
- causes packet reordering, which is not desired
- has constant maximum delay duration; it may need to be adaptive to link conditions
We should evaluate and design the algorithm before implementation.
Updated by Junxiao Shi about 10 years ago
- Assignee set to Junxiao Shi
- Target version set to v0.3
20141114 conference call decides that Junxiao could start thinking about the design.
Implementation will be a separate Task after design is approved.
Updated by Junxiao Shi almost 10 years ago
- Subject changed from Duplicate suppression on multicast face to Design duplicate suppression on multicast face
- Description updated (diff)
- Status changed from New to In Progress
- Estimated time changed from 5.00 h to 3.00 h
I have some ideas and will make slides.
Updated by Junxiao Shi almost 10 years ago
- File mcast-dup-suppress_20150211.pptx mcast-dup-suppress_20150211.pptx added
- % Done changed from 0 to 10
- Estimated time changed from 3.00 h to 6.00 h
I have read ccnd code around this area. Attached mcast-dup-suppress_20150211.pptx is my understanding.
I don't understand why ccnd algorithm would cause packet reordering.
My initial idea turns out to be somewhat similar to what's already in ccnd.
I need to understand the cause of packet reordering before continuing with this design.
Updated by Junxiao Shi almost 10 years ago
- File mcast-dup-suppress_20150214.pptx mcast-dup-suppress_20150214.pptx added
- Status changed from In Progress to Feedback
- Assignee deleted (
Junxiao Shi) - % Done changed from 10 to 20
20150211 meeting with Beichuan clarifies that ccnd's packet reordering is observed in an experiment with an unspecified "old version" of ccnd (but I don't see any change in the duplicate suppression algorithm since ccnd v0.1.0), and there's no evidence for ccnd v0.7.2 algorithm to have packet reordering problem.
I feel that duplication suppression is not the most effective way to solve the "return Data packets from multiple upstreams waste bandwidth" problem, because all upstreams must still producer/fetch the Data in response to every Interest, although all but one upstream's work will be wasted. A better solution is multicast/unicast switch. Duplication suppression will still be useful after that's available, but is less important.
Attached slides show the ideas of both duplication suppression and multicast/unicast switch, and their comparison.
Given its ineffectiveness, I don't plan to further work on duplicate suppression design recently.
Updated by Junxiao Shi almost 10 years ago
- Target version deleted (
v0.3)
20150218 conference call agrees with note-8.
"Target version" is cleared. This issue can restart after multicast/unicast switch is designed.
Updated by Ernest McCracken about 6 years ago
- Assignee set to Ernest McCracken
Hi I have started working on this data reply suppression based on random delay wait. Our implementation was specific to multi access only. I will read up on previous work and update accordingly.
Updated by Ernest McCracken about 6 years ago
Some notes after reviewing the slides and my own implementation of data cache reply suppression.
- We found that data reply suppression can be accomplished solely in a forwarding strategy. However, a switch seems like it would require a naming schema.
- Link conditions for data reply suppression could also be tracked within the forwarding strategy via measurements table.
- Multicast/Unicast switch guaranties a single data reply. Data reply suppression does not.
- More than one reply may be beneficial in high loss link conditions.
- Our data cache suppression works on content store hits. Therefore it works for upstream NFD's that have the data cached but are not the original data providers.
- In response to "all upstreams must still producer/fetch the Data in response to every Interest, although all but one upstream's work will be wasted." for a multicast/unicast switch there is an additional network overhead of responding to the initial interest where in data reply suppression the overhead is a randomized delay. If link conditions can be monitored the data reply suppression delay could result in less overall delay than two rounds of interests being sent.
- Could a multicast/unicast switch be implemented at an application library level? This would avoid altering NFD code base.
Updated by Ernest McCracken about 6 years ago
- Status changed from New to In Progress
Another issue with a multicast/unicast switch is the multicast group will have a decreased caching efficiency due to data packets not being forwarded back to the multicast face but instead using the unicast return tunnel back to the sender. Multicast groups especially on broadcast mediums would lose cache efficiency significantly.
If we are able to monitor RTTs for multicast faces we can create a random delay based on average RTT plus the additional accepted delay of the data suppression algorithm. This way the delay responds to changes in link conditions.
I think with the caching and the simplicity in its implementation(a forwarding strategy) data reply suppression is the best route.
A multicast/unicast switch still seems it would be useful especially in the multicastStrategy class where the multicasting is name based.
Updated by Ernest McCracken over 5 years ago
- Assignee deleted (
Ernest McCracken)
Removing myself from this for the time being since I have not had any time to work on the multicast/unicast switch itself.
Updated by Davide Pesavento over 4 years ago
- Tracker changed from Task to Feature
- Status changed from In Progress to Feedback
- Assignee deleted (
Ernest McCracken)
Updated by Saurab Dulal over 4 years ago
Updated by Junxiao Shi over 4 years ago
2020-04-09 NFD call discussed this issue.
Davide claims that you cannot suppress both Interest and Data due to hidden terminal problem.
In the topology below, A and B can hear each other, and B,C,D can hear each other.
- Both A and C send an Interest. B hears both Interests.
- D replies with Data. B hears the Data.
- If B suppresses Data forwarding, A would not be able to retrieve the Data.
A B C
D
Lan explains that Alvy has a solution to the hidden terminal problem: suppression probability.
By tuning the probability, it is possible to adjust how many duplicates you get, without breaking communication above.
Lixia suggests that we can start with one-hop WiFi ad hoc scenario.
However, it is unclear who is using NDN WiFi ad hoc, other than in simulations and completely fabricated research scenarios. In particular, Android does not support WiFi ad hoc at all.
Lan mentions vehicular use case (DSRC), but Davide believes it's no longer relevant in industry.
Lixia believes that industry's ignorance on ad hoc is because the lack of mature solution.
Alex mentions WiFi mesh that use peer-to-peer communication.
It allows WiFi to have mostly point-to-point communication without access point, but the protocols support multicast as well.
Davide explains that it uses HWMP protocol to establish a routing table, but this solution can only work with mostly static network and will break down if there's too-fast mobility.
Alex also mentions sensor networks including Zigbee that are fundamentally ad hoc.
No further discussion occurred on this scenario.
It's still being debated whether this feature is part of face or part of forwarding/strategy.
Lixia insists that this feature should belong in the face, because this is a L2 feature and suppression should occur per face.
Both Alvy and Alex have experimental implementation that have this feature in the strategy, because the strategy is already remembering relevant information in the tables.
No agreement was reached in this point.
Updated by Davide Pesavento over 4 years ago
Junxiao Shi wrote:
Davide claims that you cannot suppress both Interest and Data due to hidden terminal problem.
It's actually worse than that. The scenario in the example breaks down even if you only have Data transmission suppression. And you have a similar problem with only Interest suppression.
(also note that "hidden terminal" is a term I made up, it is not related to the hidden terminal problem in 802.11)
Lan explains that Alvy has a solution to the hidden terminal problem: suppression probability.
By tuning the probability, it is possible to adjust how many duplicates you get, without breaking communication above.
Yes but I'm not sure if I'd call that a "solution", it's not clear to me how to effectively set this probability. If it's too low, some nodes may need several retransmissions before obtaining the content; if it's too high, then you still have a large number of duplicate Data packets in the network.
Maybe Data suppression should be disabled (or probability of suppressing should be greatly reduced) if a node receives the same Interest more than once in a short period of time?