1. Efficient Packet Classification with Digest Caches Francis Chang Wu-chang Feng Wu-chi Feng Kang Li

2. Packet classification Essential in all network devices Routers, firewalls, NATs, Diffserv/QoS markers, etc. But, complexity increasing Number of rules Number of fields to classify Size of header (IPv6) Number of flows

3. Packet classification Performance-bound by memory Must store and access large headers and many rules quickly Lookup algorithms perform better when given more memory Classic space-time trade-off in performance Supporting line speeds requires a large amount of fast memory Fast memory expensive Large memory slow

4. Probabilistic Networking Goal of work Throw a wrench into space-time trade-off Examine a third axis: accuracy Reduce memory requirements by relaxing the accuracy of packet classification function What are the quantifiable benefits of sacrificing accuracy on the packet classification function?

5. What? Willingly make mistakes? Sure Packet errors and lack of reliability are a fact of life Masked by application layer or ignored Lots of packets are bad, some are undetectably bad [Stone00] TCP 1 in 1100 to 32000 TCP packets fail checksum 1 in 16 million to 10 billion TCP packets are UNDECTABLY bad UDP UDP packets are not required to have valid cksum Even if the cksum is bad, OS will give the packet to the application (Linux) Routing problems occur frequently Transient loops [Hengartner02] Outages

6. Several places to apply idea… Full classification Exact multi-dimensional solutions still too costly [Baboescu03] Inaccuracy may help Work in progress… Classification caches Space requirements grow linearly with number of flows and fields Use lossy recall in remembering previous classification decisions to reduce cache size Our current work..

7. Initial approach Bloom filter Approximate data structure to store flows matching a binary predicate Spell checkers Browser and web caches How it works L х N array of memory Addressed by L independent hash functions Each function addresses N buckets Storing new flows Set bits corresponding to the results of L hash functions on header Looking up flows Check bits corresponding to the results of L hash functions on header Collisions in filter cause inaccurate classifications Francis Chang, Wu-chang Feng, Kang Li, “Approximate Caches for Packet Classification”, in Proceedings of INFOCOM ’04, March 2004.

8. Bloom filter h L-1 h1h1 Flow insertion 1 1 1 Unknown flow 0 0 h0h0 0 1 2 N-1 N L virtual bins out of L*N actual bins

9. The value of making mistakes Initial results promising Small, high-performance caches with 1 in a billion error rate Storage capacity invariant to header size and fields Size of approximate cache determined by number of flows to store and desired accuracy Size of exact cache determined by number of flows to store and header size and fields IPv4-based connection identifier = 13 bytes IPv6-based connection identifier = 37 bytes

10. But… Some glaring disadvantages Large number of levels and memory lookups required Not amenable to most NP architectures Requires hardware support and parallel, bit-level memory addressing Aging properties Can not gracefully age cache No selective replacement policies possible (i.e. LRU) Must periodically expunge entire cache Results in large variance in full classifications required

11. New approach Digest caches Use a traditional cache architecture Store and use a digest of classification fields instead of full header(s)

12. Digest caches How it works Upon full classification of packet header fields (P) Calculate h 1 (P) and h 2 (P) Use h 1 (P) to select cache line Insert h 2 (P) and classification result into cache line Subsequent packets Calculate h 1 (P) and h 2 (P) Use h 1 (P) to select cache line Lookup h 2 (P) in cache line If match, follow cached result If no match, perform full classification Misclassification caused by hash-signature collisions Increases as the number of bits in digest decreases (c) Increases as the associativity of cache increases (d)

13. Digest caches Fixes all of the problems of Bloom filter caches Less memory accesses NP-friendly Does not require parallel, bit-addressable memory access Can alleviate need for associative hardware (more later) Gracefully ages Can smoothly remove old entries

14. Storage comparison between approaches

15. Evaluation Trace-driven simulation PCCS simulator (http://pccs.sourceforge.net)http://pccs.sourceforge.net Packet traces Bell Labs trace One hour trace at Bell Labs Research in May 2002 OGI trace One hour trace of OGI’s OC-3c link on July 26, 2002

16. Choosing associativity Experiment Fixed misclassification probability of 10 –9 Variable bit digest based on associativity Results similar to previous studies Small amount of associativity ideal for performance

17. Comparing approaches Digest cache 32-bit digests 4-way associativity Bloom filter cache Optimal, 30-level filter Exact caches IPv4 and IPv6 flow caches

18. Hit rates vs. cache size

19. Miss-rate variance vs. cache size

20. NP implementation IXP1200, L3Forwarder 4-way associative digest cache 803Mbs Bloom filter cache 1 level = 990 Mbs 4 levels = 652 Mbs

21. A final note to those who hate being wrong Can be used to accelerate exact caches Consider Exact cache where where associativity is emulated Entire cache line must be read sequentially to find match Digest cache acceleration Use smaller, digest cache stored in fastest (possibly associative) memory to mirror entries in exact cache Lookup in digest cache gives exact location of relevant entry in exact cache Good for implementing associative caches on NPs that do not have hardware support Speed-up analysis in paper

22. Questions?

23. Extra slides

24. Misclassification rates for digest caches 4-way associative digest cache

25. Cache misses over time