1. REDUNDANCY ELIMINATION AS A NETWORK-WIDE SERVICE Aditya Akella UW-Madison Shuchi Chawla Ashok Anand Chitra Muthukrishnan UW-Madison Srinivasan Seshan Vyas Sekar CMU Ram Ramjee MSR-India Scott Shenker UC-Berkeley

2. Growing traffic vs. network performance 2  Network traffic volumes growing rapidly  Annual growth: overall (45%), enterprise (50%), mobile (125%)*  Growing strain on installed capacity everywhere  Core (Asian ISPs – 80-90% core utilization), enterprise access, data center, cellular, wireless…  How to sustain robust network performance? * Interview with Cisco CEO, Aug 2007, Network world Enterprises Mobile users Home users Video Data centers Web content Other svcs (backup) ISP core Strain on installed link capacities

3. Enterprises Scale link capacities by suppressing duplicates 3  A key idea: suppress duplicates  Popular objects, partial content matches, backups, app headers  Effective capacity improves ~ 2X  Many approaches  Application-layer caches  Protocol-independent schemes Below app-layer WAN accelerators, de-duplication  Content distribution CDNs like Akamai, CORAL Bittorrent  Point solutions  apply to specific link, protocol, or app Mobile users Home users Video Data centers Web content Other svcs (backup) Wan Opt Dedup/ archival ISP HTTP cache ISP HTTP cache CDN

4. Universal need to scale capacities 4 Wan Opt Dedup/ archival ISP HTTP cache ISP HTTP cache Point solutions inadequate Architectural support to address universal need to scale capacities? Implications? Bittorrent ✗ Point solutions: Little or no benefit in the core ✗ Point solutions: Little or no benefit in the core ✗ Point solutions: Other links must re-implement specific RE mechanisms ✗ Point solutions: Other links must re-implement specific RE mechanisms ✗ Point solutions: Only benefit system/app attached

5. How? Ideas from WAN optimization 5 5 Cache WAN link Data center Enterprise  Network must examine byte streams, remove duplicates, reinsert  Building blocks from WAN optimizers: RE agnostic to application, ports or flow semantics  Upstream cache = content table + fingerprint index  Fingerprint index: content-based names for chunks of bytes in payload  Fingerprints computed for content, looked up to identify redundant byte- strings  Downstream cache: content table

6. Internet2 Packet cache at every router From WAN acceleration to router packet caches 6 Wisconsin Berkeley CMU Router upstream removes redundant bytes Router downstream reconstructs full packet Router upstream removes redundant bytes Router downstream reconstructs full packet (Hop-by-hop works for slow links Alternate approaches to scale to faster links…) (Hop-by-hop works for slow links Alternate approaches to scale to faster links…)

7. Implications overview: Performance and architectural benefits 7  Improved performance everywhere even if partially enabled  Generalizes point deployments and app-specific approaches Benefits all network end-points, applications  Crucially, benefits ISPs Improved switching capacity, responsiveness to sudden overload  Architectural benefits  Enables new protocols and apps Min-entropy routing, RE-aware traffic engineering (intra- and inter-domain) Anomaly detection, in-network filtering of unwanted traffic  Simplifies/improves apps: need not worry about using network efficiently Application control messages & headers can be verbose  better diagnostics Controlling duplicate transmission in app-layer multicast is a non-issue

8. Internet2 Implications example: Performance benefits 8 Network RE  12 pkts (ignoring tiny packets) Network RE  12 pkts (ignoring tiny packets) Without RE  18 pkts 33% lower Without RE  18 pkts 33% lower Wisconsin Berkeley CMU Generalizes point deployments Benefits ISPs: improve effective switching capacity 6  2 packets 3  2 packets

9. Wisconsin Internet2 Implications example: New protocols 9 RE + routing  10 pkts RE + routing  10 pkts Simple RE  12 pkts Berkeley CMU 9 ✓ Redundancy-based anomaly detectors ✓ Network-assisted spam filtering ✓ New content distribution mechanisms ✓ Redundancy-based anomaly detectors ✓ Network-assisted spam filtering ✓ New content distribution mechanisms ✓ Minimum-entropy routing ✓ New, flexible traffic engineering mechanisms ✓ Inter-domain protocols ✓ Minimum-entropy routing ✓ New, flexible traffic engineering mechanisms ✓ Inter-domain protocols

10. Network RE service: Quantitative results 10  Analysis of 12 enterprises: traffic 15-60% redundant [SIGMETRICS 09]  ~1GB of cache sufficient to identify redundancies  DRAM or PCM (PRAM) on routers  Network RE benefits both ISPs and end-networks [SIGCOMM 08]  Upto 15-50% better util, responsive TE, control inter-domain traffic impact  Centralized algorithm for min-entropy routing (using “redundancy profiles”) Reduces utilization by a further 10-25% in intra-domain case Inter-domain min-entropy routing: gains much more significant (50-80%)  Is network RE viable at high speeds? Not in its current form…  Compression is slow: limits hop-by-hop speed at each hop to 2.5Gbps Acceptable for access, wireless, cellular links, not for the core Also, wastes memory on multiple routers  limits effectiveness

11. SmartRE: Concerted network-wide RE 11  Toss out link-by-link view; treat RE as a network-wide problem per ISP [Current work]  Memory usage: each packet compressed/un-compressed once  Throughput: allow reconstruction multiple hops away from compression  Stand-alone reconstruction much faster when freed from dependence on compression immediately upstream  Reconstructor can reconstruct a lot more, from multiple different compression agents  Resource-awareness: carefully account for network and device resources, and traffic  Compression/reconstruction/caching locations decided based on memory capacity and memory operations  Also consider global TE objectives  Just 4% from ideal RE (no memory or processing constraints)

12. Summary and future directions 12  RE service to scale link capacities everywhere  Architectural niceties and performance benefits  High speed router RE seems feasible  Future directions  End-host participation  Role of different memory technologies – DRAM, flash and PCM  Theoretical issues – pricing and economics, routing policy, network design  Network coding as an alternative to compression