Network Resource Management Jason Gaedtke Chief Scientist W

Network Resource Management Jason Gaedtke Chief Scientist W 3 C Video on the Web Workshop December 2007�

Topics • Abstract § Compelling network neutrality arguments notwithstanding, not all IP traffic exhibits uniform distribution requirements (e. g. , bandwidth, latency, jitter and TTL). § Further, automated P 2 P file-sharing agents exploit TCP congestion control algorithms to gain a disproportionate share of network resources. § Some measures should be explored to address this natural, shared-network heterogeneity. • • • Heterogeneous Applications and Network Requirements Web Video Distribution Trends A Resource Consumption Example (Briscoe Draft) Potential Management Strategies References and Collaborative Activities January 12, 2006 © Cable Television Laboratories, Inc. 2007. All Rights Reserved. Proprietary/Confidential 2

Heterogeneous Apps and Network Reqs • Real-Time Apps: dependent upon low-latency delivery, exhibit highly-variable bandwidth requirements, few simultaneous connections § § Online gaming Vo. IP and video chat IM and Presence Streaming video • Interactive Services: tolerant of modest delivery delays, modest bandwidth, few simultaneous connections § E-mail § Web browsing § Progressive download • Content Distribution: automated, many simultaneous connections, greedy – will consume available bandwidth § P 2 P file-sharing § File/mail/news/Web servers January 12, 2006 © Cable Television Laboratories, Inc. 2007. All Rights Reserved. Proprietary/Confidential 3

Web Video Distribution (Background) • Web Servers/Farms § Simple client/server architecture § Commodity servers, scaled horizontally – Capacity – Redundancy/Availability • Content Distribution Networks (CDNs) § Specialized client/server architecture with aggressive caching § Geographical distribution and load-balancing • P 2 P Networks § Decentralized, distributed and self-organizing § “Super-nodes” avoid n 2 link scaling and search § Participants contribute bandwidth, storage and processing • Hybrid CDN/P 2 P Networks § Benefits of P 2 P resource sharing; <10% distro costs § Seed and “long-tail” content sourced via CDN caches January 12, 2006 © Cable Television Laboratories, Inc. 2007. All Rights Reserved. Proprietary/Confidential 4

A Resource Consumption Example (Briscoe, draft-briscoe-tsvwg-relax-fairness) • 10 Mbps, shared access network, 100 subscribers § 80 subscribers primarily interactive Web/e-mail: – 10% concurrency, 2 TCP connections each – 9. 9 kbps average during congestion – 7. 1 MB per day (16 -hours active) § 20 automated P 2 P file-sharing clients: – 100% concurrency, 100 TCP connections each – 496 kbps average during congestion – 3. 6 GB per day – 500: 1 volume skew TCP congestion control treats each flow equally; greedy apps spawn many connections January 12, 2006 © Cable Television Laboratories, Inc. 2007. All Rights Reserved. Proprietary/Confidential 5

Resource Consumption (4 x Capacity) • 40 Mbps shared, 100 subscribers § 80 interactive Web/e-mail: – 4% active (due to more responsive apps) – 40 kbps (vs. ~10 kbps) during congestion – 11 MB (vs. ~7 MB) per day § 20 automated P 2 P file-sharing: – 2 Mbps (vs. ~500 kbps) during congestion – 14 GB (vs. ~3. 5 GB) per day As expected, a 4 x increase in network capacity yields a 4 x increase in average, per-flow rates under congestion; only exacerbates skew (>1250: 1) January 12, 2006 © Cable Television Laboratories, Inc. 2007. All Rights Reserved. Proprietary/Confidential 6

Resource Consumption (Capacity + Churn) • 40 Mbps shared, 60 subscribers: § 50 interactive Web/e-mail (30 churn): – 2. 5% active – 80 kbps (vs. 40 kbps) during congestion – 14 MB (vs. 11 MB) per day § 10 automated P 2 P file-sharing (10 churn): – 4 Mbps (vs. 2 Mbps) during congestion – 29 GB (vs. 14 GB) per day Trends: fewer subscribers, greater network capacity/cost, >2000: 1 consumption skew; ergo, rational operators will not add capacity January 12, 2006 © Cable Television Laboratories, Inc. 2007. All Rights Reserved. Proprietary/Confidential 7

Resource Consumption Summary • TCP congestion control treats all flows equally • Automated P 2 P agents are (very) greedy § 100+ simultaneous connections § 100% concurrency • These aggressive algorithms will absorb an increasing amount of added capacity, thus degrading cost/benefit for other users • Light, interactive users subsidize P 2 P distribution • New economic/technical management strategies should be explored January 12, 2006 © Cable Television Laboratories, Inc. 2007. All Rights Reserved. Proprietary/Confidential 8

Potential Management Strategies • • Upstream/downstream rate limiting Aggregate capacity limiting (tiering) Application-specific throttling (via DPI) Differentiated/priority service classes Reservation-based resource management Explicit protocol-level feedback/heuristics Variable/metered pricing strategies Bandwidth/resource trading schemes and virtual economies • Others? January 12, 2006 © Cable Television Laboratories, Inc. 2007. All Rights Reserved. Proprietary/Confidential 9

References and Collaborative Activities • IETF Transport Area § RFC 2309: Recommendations on Queue Management and Congestion Avoidance § RFC 2581: TCP Congestion Control § RFC 2914: Congestion Control Principles § draft-briscoe-tsvwg-relax-fairness • • DCIA P 4 P Working Group Cable. Labs Packet. Cable Multimedia Qo. S DSL Forum TR 58/59 Harvard SEAS and Tribler. org § Bandwidth Virtual Economy January 12, 2006 © Cable Television Laboratories, Inc. 2007. All Rights Reserved. Proprietary/Confidential 10