1 / 16

Quantifying Overprovisioning vs. Class-of-Service: Informing the Net Neutrality Debate

Quantifying Overprovisioning vs. Class-of-Service: Informing the Net Neutrality Debate. Murat Yuksel (University of Nevada – Reno) yuksem@cse.unr.edu K. K. Ramakrishnan (AT&T Labs Research) kkrama@research.att.com Shiv Kalyanaraman (IBM Research India) shivkumar-k@in.ibm.com

harshman
Download Presentation

Quantifying Overprovisioning vs. Class-of-Service: Informing the Net Neutrality Debate

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Quantifying Overprovisioning vs. Class-of-Service: Informing the Net Neutrality Debate Murat Yuksel (University of Nevada – Reno) yuksem@cse.unr.edu K. K. Ramakrishnan (AT&T Labs Research) kkrama@research.att.com Shiv Kalyanaraman (IBM Research India) shivkumar-k@in.ibm.com Joseph D. Houle (AT&T) jdhoule@att.com Rita Sadhvani (Verizon Wireless) rita.sadhvani@verizonwireless.com

  2. rate 10000 0 0 40000 80000 time rate 10000 0 0 40000 80000 time Motivation: Thick (Over-provisioned) or Thin (Engineered) Pipes? • Media-rich applications require performance guarantees: • e.g.: VoIP requires <300ms round-trip delay, <1% loss • How to respond to these application needs? • CoS approach: provide priority (i.e. higher class) to premium traffic • Classless (best-effort) service approach: over-provision the capacity • Question:How much extra capacity does the classless service require to match the performance of the higher class (premium) service in the CoS approach? [Jim Roberts et al.] Thin: How to deal with bursts/overload? And meet premium SLAs… ! Thick: Cost of overprovisioning? Can this commodity model break even?

  3. Classless Link (neutral) BE N=? Two Approaches: CoS vs. Classless CoS Link (differentiated) Prem= gD Premium Scheduling (e.g. priority) D D BE=(1-g)D BE • GIVEN: D, D and a performance target (i.e. ttarget or ptarget) • FIND:The minimum N that gives the same performance as in the premium class of the CoS case? D

  4. REC: Required Extra Capacity REC = <required neutral link capacity> - <CoS link capacity> = N - D (rate) = 100(N/D – 1) (%) How to quantify REC?

  5. Analytical Link Model: Poisson traffic • Assume: • Poisson traffic, Exponential packet lengths for traffic in each class i.e. • Premium class traffic is Poisson with g D • Best-effort class traffic is Poisson with (1-g) D • The aggregate traffic for the neutral link is also Poisson with rate D Both the performance target and the REC can be expressed in terms of two key parameters: (i) ρ – utilization, (ii) g – proportion of premium traffic.

  6. More Bursty Traffic: MMPP • MMPP = Markov-Modulated Poisson Process • Easy to do the math… • Simplest MMPP is of two states. • MMPP traffic with mean D • Traffic w/ equivalent rate to the neutral case, but w/ more burstiness. 1 2 Higher r means more bursty traffic.

  7. Simulated Link Model: DelayMMPP/M/1 model If packet size is 1KB and the CoS link is D = 1Gb/s: 5,000packets of delay = 40.1ms a=0.5, r=4 Surface color shows the performance target. REC can be quite high even for very small g and medium utilizations.

  8. Simulated Link Model: DelayMMPP/M/1 model a=0.5, r=4 Can be drawn in multiple 2-d graphs REC increases as link utilization increases REC is large even for small proportion of premium traffic

  9. Simulated Link Model: LossMMPP/M/1/K model REC for the same performance target decreases as buffer size increases The graphs are generic for various buffer sizes. An example: For a 10Mb/s link carrying 1KB packets: K = ~15pkts  25ms buffer time K = ~60pkts  100ms buffer time

  10. Simulated Link Model: LRD Traffic LOSS – LRD/D/1/K DELAY – LRD/D/1 • Internet traffic : known to be LRD with Hurst parameter value between 0.7 and 0.9. • REC for Hurst=0.75 is significantly higher than our 2-state MMPP model results. • We also observed that REC increases as Hurst value increases towards 0.9. • Also looked at closed-loop traffic - many TCP flows - and observed similar trends. • We further looked at the case when Premium traffic is CBR and BE is TCP, and this increased REC further.

  11. Network Model • Steps to calculate network REC (NREC): • Step 1: Construct the routing matrix RFxL based on shortest path • Run Dijkstra’s algorithm on the topology matrices ANxN and WNxN • Step 2: Form the traffic vectorFx1 from TNxN • Step 3: Calculate the traffic load on each link: RT = Q • Step 4: Check the feasibility of the traffic load and routing • For any link • If link capacity is less than the traffic load (e.g. C < Q) then update T accordingly and go to Step 2. • Step 5: Calculate the required per-link REC (i.e. N - D) by using QI as the traffic rate D for Ith link, and the performance goal ptarget or ttarget. Used Rocketfuel topologies for ANxN and WNxN. Used gravity model for TNxN. Made a look-up to the simulated link model results.

  12. NREC: Two ways to calculate • Steps to calculate network REC (NREC) (cont’d): • Step 6: Calculate the NREC by averaging the per-link RECs from Step 5. • We calculated NRECs for the Rocketfuel topologies: • Used the MMPP link model (a=0.5 and r=4) or the LRD link model (H=0.75) – Much more conservative than real or TCP traffic • Assumed K=100ms buffer time • Only report Sprintlink, as the other topologies gave higher REC values total extra capacity needed on the whole network average extra capacity needed on each link

  13. NREC for Sprintlink: G2G Delay NREC can be much higher than 100% for a network operating with 60% utilization. 10ms queueing delay target for VoIP may require large REC values. Solid lines are NRECI and dashed lines are NRECA

  14. NREC for Sprintlink: G2G Loss NREC can be much higher than 1000% even for a network operating with 40% utilization. 0.1% loss target may require large REC values. Solid lines are NRECI and dashed lines are NRECA

  15. Summary • A framework to study REC for delay or loss being the performance target. • Link model • REC grows when: • traffic becomes more bursty • the utilization of the CoS link becomes higher • the performance target becomes tighter • the fraction g of the Premium class traffic becomes smaller • Closed-loop (e.g., TCP) or LRD traffic further increases REC • Network model: • For legacy g2g performance targets, REC ranges from 50% to over 100% as g reduces below 0.5 and the utilization goes up to 60%. • Future trends/work: • The performance targets will keep becoming tighter. REC is high perpetually – not just today, but in future also.. • The value of g is a crucial factor. Small g does not necessary favor a classless network. • Further research should estimate the actual costs of CoS and classless designs, as scheduling & management complexity need to be considered.

  16. THE END Thank you!

More Related