Is IP going to take over the world (of communications)? Pablo Molinero-Fernandez, Nick McKeown Stanford University Hui Zhang Turin Networks, Carnegie Mellon University Alan Mislove, Ansley Post COMP
Background The Internet is one of the most successful communications platforms Seen exponential growth in the past decade Almost all Internet traffic is over Internet Protocol (IP) Designed in 1970s through DARPA funding IP’s great success due to Reachability Heterogeneity
Alan Mislove, Ansley Post COMP Background (cont.) Success has lead to the assumption that IP will become the sole communication platform Voice-over-IP systems will replace phone network TV, Movies will be disseminated using Internet Related assumption is that packet-switching (IP) routers will become the only type of switching device
Alan Mislove, Ansley Post COMP Motivation IP is technically able to support all types of applications Request-reply (web traffic) Real-time (telephony) Despite its strengths, not necessarily the best solution Goal: Question previous assumptions that IP will “take over the world (of communications)” Evaluate what would happen if we started over
Alan Mislove, Ansley Post COMP IP Folklore There are many widely held assumptions (“sacred cows”) about IP that must be reevaluated The current dominance of IP for communications The efficiency of IP The robustness of IP The simplicity of IP IP’s suitability for real-time applications
Alan Mislove, Ansley Post COMP IP Communications Dominance It is widely (and incorrectly) believed that IP already dominates global communication ISP markets have revenues of $13B Other communication markets total over $300B For data and telephony applications alone, IP routers total $4B, while circuit-based router total $32B Internet reaches 59% of US, phone 94%, TV 98%
Alan Mislove, Ansley Post COMP IP’s Efficiency IP makes efficient use of scarce bandwidth Very good for wireless channels, satellite links, etc… But is bandwidth actually scarce? Average Internet link utilization is 3%-20% LAN usage is much lower, about 1% Long-distance phone utilization is 33% Networks are highly overprovisioned to provide a consistent user experience Low packet delay is the goal
Alan Mislove, Ansley Post COMP IP’s Efficiency (cont.) Many reasons given for overprovisioning Internet traffic is asymmetric and bursty Difficult to predict traffic growth on a link Economical to add large increments of capacity However, there are “less talked-about” reasons Under congestion, IP performs badly Control traffic transmitted in-band Results in black holes, loops, etc… Much easier to keep utilization low
Alan Mislove, Ansley Post COMP IP’s Efficiency (cont.) In practice, user experiences the same delay in packet-switched or circuit-switched network Average user’s work (65%) is request-response Web traffic File sharing For these types of workloads, circuit-switching provides same user response time
Alan Mislove, Ansley Post COMP IP’s Robustness Internet was designed to withstand catastrophic event, but Median Internet downtime is 471 minutes/year Median phone downtime is 5 minutes/year BGP convergence is slow (3-15 minutes) SONET/SDH switches to a backup path in 50ms Nothing inherently unreliable about circuit-switching
Alan Mislove, Ansley Post COMP IP’s Simplicity Beginning principle is that complexity should be at the endpoints Increasingly, IP routers have become sophisticated Multicast Quality of Service VPN Configuring IP routers can be very difficult Single misconfigured IP router can cause instability for a large portion of the network
Alan Mislove, Ansley Post COMP IP’s Simplicity (cont.) Circuit-switched routers have 3 million lines of code IP routers have about 8 million IP routers have 300 million gates, 1 CPU, 300 MB of buffer space Circuit routers have 25% of the gates and no CPU Circuit-switched routers sell for 1/2 - 1/12 the price Circuit switching is compatible with optical technology
Alan Mislove, Ansley Post COMP IP’s Real-Time Support Widely held assumption that IP will support real-time applications This assumption relies on overprovisioning of the network Or quality-of-service in the network that has yet to be implemented
Alan Mislove, Ansley Post COMP What if we started over? Hybrid solution would be most appropriate Uses packet switching at the edges Circuit-switching at the core and with applications with QoS demands Tightly integrate these two parts
Alan Mislove, Ansley Post COMP Conclusion IP does some things good, but not everything Good for scarce-bandwidth situations Wireless, undersea cables, satellite links Inappropriate for real-time applications Voice traffic, telephony If we redesigned the Internet, not all routers would be packet-switching Core routers and real-time application data would be circuit-switched
Alan Mislove, Ansley Post COMP Questions? Mike O’Dell, former Senior VP, UUNet: “[to have a voice-over-IP network service one has to] create the most expensive data service to run an application for which people are willing to pay less money every day”