Presentation is loading. Please wait.

Presentation is loading. Please wait.

Congestion Avoidance with Incremental Filter Aggregation in Content-Based Routing Networks Mingwen Chen 1, Songlin Hu 1, Vinod Muthusamy 2, Hans-Arno Jacobsen.

Similar presentations


Presentation on theme: "Congestion Avoidance with Incremental Filter Aggregation in Content-Based Routing Networks Mingwen Chen 1, Songlin Hu 1, Vinod Muthusamy 2, Hans-Arno Jacobsen."— Presentation transcript:

1 Congestion Avoidance with Incremental Filter Aggregation in Content-Based Routing Networks Mingwen Chen 1, Songlin Hu 1, Vinod Muthusamy 2, Hans-Arno Jacobsen 3 1 Chinese Academy of Sciences 2 IBM T.J. Watson Research Center, 3 University of Toronto July 2, 2015, ICDCS MIDDLEWARE SYSTEMS RESEARCH GROUP http://padres.msrg.org

2 Content-based routing S1S1 Subscriber EDCBA P1P1 Publisher S2S2 S3S3 FG

3 Subscription covering S1S1 Subscriber EDCBA P1P1 Publisher S2S2 S3S3 S1S1 S2S2 S3S3 FG Fewer subscription messages Smaller routing tables Faster matching

4 Problem with unsubscription Bursty traffic Network congestion High broker load Compare with aggregation in IP networks  Relatively static hierarchical addresses S1S1 Subscriber EDCBA P1P1 Publisher S2S2 S3S3 S1S1 S2S2 S3S3 FG

5 How bad is the problem? 5

6 Tradeoff: covering vs. filtering Suppose we retain covering subscription even after unsubscribe  Avoids bursty subscription traffic  May lead to false positives Amount depends on similarity between subscriptions Can we selectively retain portions of subscription tree? S1S1 Subscriber EDCBA P1P1 Publisher S2S2 S3S3 S1S1 S2S2 S3S3

7 Agenda Incremental filter aggregation algorithm  Record statistics of publications  Compute effective similarity among subscriptions  Selectively prune portions of subscription tree Quantitative evaluation  Compare with traditional covering algorithm  Incremental filter aggregation reduces congestion Also reduces routing table size and processing time

8 8 Incremental filter aggregation algorithm

9 Example: benefits of selectively pruning subscription tree It is possible to eliminate subscriptions triggered by unsubscription of S 1 and avoid false positives To tradeoff subscription bursts and false positives, we need to consider similarity among subscriptions S1S1 Subscriber EDCBA P1P1 Publisher S2S2 S3S3 S1S1 S2S2 S3S3

10 Subscription similarity ICDCS 2015 Congestion Avoidance 10 S: unsubscription R: other (intersecting) subs S: covered (triggered) subs |S ∩ ( S U R )| |S| Ø =

11 Selective pruning S 9 6 3 1 2 7 5 4 11 7 10 P P P P P P Ø(B 1 ) Ø(B 2 ) Ø(B 4 ) Remove subscription if  Similarity (Ø) is low  Number of triggered subs is low How to compute similarity?  Computing mergers of covered subs is expensive  Need to compute at every hop?

12 Subscription similarity ICDCS 2015 Congestion Avoidance 12 P(S): publications that match unsubscription P(R): publications that match other (intersecting) subs P(S): publications that match covered (triggered) subs |P(S) ∩ (P( S )UP( R ))| |P(S)| Ø = = |P(S)| |P*(S)|

13 Recording publication statistics Publications annotated with  Distance from SHB to nearest broker with another matching sub  Count of matching subscriptions Distance and count init to 0 S1S1 Subscriber EDCBA P1P1 Publisher S2S2 S3S3 S1S1 S2S2 S3S3 When publication matches multiple subscriptions  Increment count  Reset distance Otherwise  Increment distance D = 0 C = 0 D = 0 C = 1 D = 1 C = 1 D = 0 C = 2 D = 1 C = 3 FG

14 Example: publication statistics S1S1 Subscriber EDCBA P1P1 Publisher S2S2 S3S3 S1S1 S2S2 S3S3 FG Statistics at Broker A NotificationMatchesSubPublisherDistanceCount N1S1, S2, S3S1P113 N2S1 P141 N3S1, S3S1P132 N4S1, S2S1P112 N5S1, S2S1P112 N6S1, S2S1P112

15 Interpretation of statistics S1S1 EDCBA P1P1 S2S2 S3S3 S1S1 S2S2 S3S3 FG Statistics at Broker A NotifMatchSubPubrDistCnt N1S1, S2, S3S1P113 N2S1 P141 N3S1, S3S1P132 N4S1, S2S1P112 N5S1, S2S1P112 N6S1, S2S1P112 Portions of sub’s tree … ... >= D hops away can be kept without false positives  … < D hops will incur false positives if kept For given D, broker can calculate false positive rate  D = 1  2/6 false positives  D = 3  1/6 false postiives

16 When to perform unsubscription Unsubscription of covered subscription S 2 does not trigger a subscription burst  Always perform unsubscription immediately Unsubscription of covering subscription S 1 can trigger a subscription burst  Compute similarity between subscriptions to determine if unsub of S 1 should be forwarded another hop S1S1 S2S2 S3S3

17 Selective subscription tree pruning S1S1 EDCBA P1P1 S2S2 S3S3 FG Statistics at Broker A NotifMatchSubPubrDistCnt N1S1, S2, S3S1P113 N2S1 P141 N3S1, S3S1P132 N4S1, S2S1P112 N5S1, S2S1P112 N6S1, S2S1P112 Include following along with unsub message for each (publisher, distance) pair  Publications and false positives Count total pubs that match sub  Add the publication counts in the list Count false positives if sub is not removed  If the list entry’s host broker is further than D hops, all pubs are false positives Included with unsub of S 1 PubrDistPubsFalseExamined by Broker Similarity: (pubs - false) / pubs P1066A0/6 P1162B4/6 P1262C4/6 P1361D5/6 P1461E5/6 If similarity threshold = 4/6  Prune up to B If similarity threshold = 5/6  Prune up to D

18 Result of selective pruning S P P P P P P Certain subtrees of the original subscription tree are preserved Can compute similarity in a distributed manner  With little overhead (up to 24 MB in experiments) False positive rate decreases (similarity increases) downstream  See paper for proof  Justifies preserving contiguous subtrees

19 19 Evaluations

20 Experimental setup Implemented in PADRES pub/sub system 49 brokers 80000 subs distributed across edges 20 unsubs Algorithms  Active covering  Lazy covering  Subscription packing  Incremental unsubscription Metrics  Broker input queue length  Broker routing table size  Publication processing time

21 Why bother with covering? ICDCS 2015 Congestion Avoidance 21 Active coveringLazy coveringNo covering Sub matching0.811.6 Sub covering12270 Unsub1.400 Pub3.06.27.9 Average message processing time in milliseconds Covering is worthwhile Covering dominates the subscription processing pipeline

22 Publications experience severe delays (large and long-lived) Queue lengths follow similar trends (see paper) 22 avg delay: 1000 s avg delay: 0.4 s

23 Increased covering causes the excessive queue lengths and publication delays ICDCS 2015 Congestion Avoidance 23 700 covered subs1400 covered subs2100 covered subs

24 Unexpected incidental benefit: incremental algorithm also reduces routing table size Scenario: mobile subscribers Workload from Siena paper [1]  200k subs  Some subs cover ~2000 subs 24 [1] A. Carzaniga and A. L. Wolf, “Forwarding in a content-based network,” in Proceedings of ACM SIGCOMM, 2003.

25 Conclusions Common filter aggregation techniques expose a vulnerability in content-based routing networks  Removal of a covering subscription can trigger burst of subscriptions and cause congestion Proposed solution selectively maintains portions of a subscription tree  Avoid triggered congestion and control false positives  Compute “effective” similarity among subscriptions for a given workload Evaluations show congestion is virtually eliminated with incremental aggregation algorithm  Also, proposed algorithm reduces routing table size and publication processing time Future work  Incrementally prune subscription tree over time  Consider more dynamic environments with mobile publishers and subscribers, and changing workloads  More precisely control tradeoff between subscription traffic congestion and false positive publication traffic

26 26 Congestion Avoidance with Incremental Filter Aggregation in Content-Based Routing Networks http:// padres.msrg.toronto.edu Q&A


Download ppt "Congestion Avoidance with Incremental Filter Aggregation in Content-Based Routing Networks Mingwen Chen 1, Songlin Hu 1, Vinod Muthusamy 2, Hans-Arno Jacobsen."

Similar presentations


Ads by Google