Differentiated Service Routers in the Internet

Größe: px

Ab Seite anzeigen:

Download "Differentiated Service Routers in the Internet"

Hansl Kraus
vor 8 Jahren
Abrufe

1 Differentiated Service (DiffServ) Quality of Service Architectures in the Internet: Integrated Services (IntServ) Differentiated Services (DiffServ) DiffServ: Differrent packet treatment depending on packet marking in the IP-Header (DS-Field) Move complexity to the edge routers and keep processing in the core routers as simple as possible Packet marking and remarking performed at: edge routers (between subscribers), border routers (between provider domains) and applications.1 Proportional Differentiated Services: Two approaches to service differentiation: Absolute Differentiated Services: provides guarantees and statistical assurances to performance measures, e.g.: E2E delay Relative Differentiated Services: ordering of a number of traffic classes in terms of their packet forwarding behaviour Proportional Differentiated Services: further refinement of the relative Differentiated Services approach quantifies the spacing between the traffic classes Possible differentiation criteria: throughput delay jitter loss rate.2

(between subscribers), border routers (between provider domains) and applications.

2 Proportional Delay Differentiation: Definition of N traffic classes Differentiation criteria: queueing delay Definition of the ratios of mean delay of consecutive classes Ï Ï Æ Æ Class differentiation parameters (delay differentiation parameters: DDP) δ 1 > δ 2 >... > δ Ν so that higher classes are "better" than lower classes.3 Dynamic Scheduling Model Priority Function Priority q r (t) q r (t) q r (t) =(t t 0 ) b r 0 b 1 b 2... b r t 0 t 0 t Priority of a class-r packet increases linearly with the elapsed time in the queueing system at a rate b r.4

.. > δ Ν so that higher classes are "better" than lower classes.

3 Main Waiting Time: W r = W 0 1 ρ r 1 i=1 1 R i=r+1 ρ i W i ( 1 b i ρ i ( 1 b r b i b r ) ) Extended Priority Function: dependency on the current load ρ i dependency on the delay differentiation parameters (DDPs) δ i Õ Ö Øµ Ø Ø ¼ µ Ö ½ Ê Æ ½ Æ Ê µ ½ Ö Ê.5 Dynamic Priority Scheduling - Architecture.6

ρ i dependency on the delay differentiation parameters (DDPs) δ i Õ Ö Øµ Ø

4 Optimization Problem: R traffic classes Given DDPs: δ 1 > δ 2 >... > δ R Measured class utilizations: ρ 1 > ρ 2 >... > ρ R Ideally: Or: Ê ½ ½ Ê ½ ½ Ï Ï ½ Ï Ï ½ Æ Æ ½ ¾ ¼ Æ Æ ½ ¾ ÑÒ.7 Example of two Traffic Classes: Exact form to calculate the slopes of the priority functions: ½ ½ ½ ¾ ½ ½ ½ Æ ¾ µ Æ½ Feasability: ½ Æ ¾ Æ ½ Limit: As the load aproaches 100%, the ratio of the scheduler parameters: b 1 /b 2 tends to the inverse of the DDPs: δ 2 /δ 1.8

7 Example of two Traffic Classes: Exact form to calculate the slopes of the priority functions: ½ ½ ½ ¾ ½ ½ ½ Æ

5 Optimization using Genetic Algorithms Genetic algorithms(gas) are optimization and search procedures inspired by genetics and the process of natural selection Structure of a genetic algorithm Randomly generate a population of individuals (bit strings). Decode each individual and evaluate its fitness. Generate a new population partly by cloning, partly by combining and partly by mutating the fittest individuals. Repeat steps 2 and 3 until a stop condition holds..9 Example: 4 Traffic Classes, Required Delay Ratio: 2 Scheduler Prameters: b 1, b 2, b 3, b 4 Ò ± ½ ¾ ¼± ½º¼¼ º¾ ½ º½¼ ¾ ± ½º¼¼ º ¾¼º½ ¼ ¼± ½º¼¼ º½ ½º¼ ¾½º¾ ± ½º¼¼ ¾º º º ¼± ½º¼¼ ¾º¾ º¼½ ½º¾ ± ½º¼¼ ¾º½ º ½¼º¾ ½¼¼± ½º¼¼ ¾º¼¼ º¼¼ º¼¼.10

Generate a new population partly by cloning, partly by combining and partly by mutating the fittest individuals. Repeat steps 2 and 3 until a stop condition holds.

6 Mean Delays and Delay Ratios Ï½ Ï¾ Ï Ï Ï½ Ï¾ Ï¾ Ï Ï Ï ± Ñ Ñ Ñ Ñ ¼± º º¾ ¾ ¾º¾ ½º¾½¾ ¾º¼¼ ½º ¾º¼¼ ± º½ º¾ ¾º¾ ½º¾ ½ ¾º½ ½º ¾º¼¼ ¼± ½¼º º ¾º ½º ¾º¼¼ ¾º¼¼ ¾º¼¼ ± ½º¾¾¾ º½½½ º ½º ¾º¼¼ ¾º¼¼ ¾º¼¼ ¼± ¾½º ½¼º º ¾º ¾º¼¼ ¾º¼¼ ¾º¼¼ ± ¾º ¾½º ½¼º º ¾º¼¼ ¾º¼¼ ¾º¼¼ ½¼¼± ¾½ ½ ½¼ ¾ ¾º¼¼ ¾º¼¼ ¾º¼¼.11 Implementation Offline calculation of the scheduling parameters Dynamic adjustment during router runtime operation Adjustment: Look-up tables: pre-calculated parameter sets for intervals of the utilization range --> discretization error Use of simple interpolation functions Adjustment trigger: Periodical On special events: significant utilization change in one or more traffic classes.12

11 Implementation Offline calculation of the scheduling parameters Dynamic adjustment during router runtime operation Adjustment: Look-up tables: pre-calculated

7 Experiments with other Priority Functions Motivation Investigate the performance using other priority functions for lower utilizations (especially around 70%) Using initial priorities could better fulfill the proportional differentiation requirements, as the priority initially assigned to the arriving packets is higher than zero Priority Function with Initial Priorities Higher optimization error (partially due to the approximate formulae) Quadratic Priority Function No significant improvement in comparison to the linear priority function.13 Conclusion Conclusion Approach of Dynamic Priority Scheduling Applies equally well to heavy and moderate load conditions Offline determination of the scheduling parameters and table look-up or simple interpolation functions during runtime Use of genetic algorithms for optimization Ongoing work: Simulation to validate developed results Impact of infinite buffer approximation and exponential distribution Further consideration of new Internet traffic modelling methods (Self similar traffic, Pareto distribution) Optimization using simulation for nonexp. distribution.14

error (partially due to the approximate formulae) Quadratic Priority Function No significant improvement in comparison to the linear priority function.

Ähnliche Dokumente

ÃÔØÐ ¾ ÓÙÖÖ¹ÌÖÒ ÓÖÑ¹ËÔØÖÓ ÓÔ Ñ ÒÖÖÓØÒ ËÔØÖÐÖ ÓÙÖÖ¹ÌÖÒ ÓÖÑ Ìµ¹ÌÒ Ø ÑØ Ö ÒØÛÐÙÒ Ð ØÙÒ ĐÖ ÓÑÔÙØÖ ÙÒ ÒÐÐÖ ÐÓÖØÑÒ Ò Ò ÐØÞØÒ ¼ ÂÖÒ Ò Ö ÑØÒ ËÔØÖÓ ÓÔ ÒÞÙ ÐØÒ ÙÒ ÛÖ ÑØØÐÖÛÐ Ò Ø ÐÐÒ ÔØÖÐÒ ÖÒ ÚÖÛÒØº ËÓ ÒØ Ì¹ÌÒ Ò