TeraII Traffic engineering in Terabit Network

1
TeraII Traffic engineering in Terabit Network

• Seminar
• May 21, 2003
• timoh_at_cc.jyu.fi
• public page http//tisu.it.jyu.fi/terabitti/

2
Contents

• Introduction (T. Hämäläinen)
• Adaptive scheduling mechanism for capacity
  optimization and revenue maximization (T.
  Hämäläinen)
• MPLS BGP/VPN (M. Laiho)
• SLA monitoring (J. Puttonen)
• Traffic classification and simulations with ns2
  (M. Ketola)
• IP-TVmulticast at terabit network (O. Alanen)
• Tasks for the second project year (discussion)
• IP-TV demos at the lab.

3
Major Achievements

• Test network
• Laboratory network has been increased to cover
  WTSs, Digitas and Yomi/Kesnets networks
• MPLS VPN and RPR (802.17) techniques can be used
  to find out potential traffic engineering
  solutions in multi AS environment
• Multicast multiAS configurations with MPLS VPN
  using BGP can be evaluated (digi-TV server for
  media distribution is in use)
• Revenue and QoS -aware load balancing and
  scheduling algorithms have been developed
• QoS Aware Traffic Flow Balancing
• Adaptive Weighted Scheduling Algorithm

4
Progress towards the plan

• 1.6.2002 31.5.2003
• test configurations with IP/MPLS multicast have
  been run
• Digi-TV server for IP multicasting has been
  installed
• MPLS VPN BGP routing configurations have taken
  quite a lot of time.
• Simulations with revenue and QoS aware scheduling
  under different kind of traffic scenarios has
  been run with ns2
• Published papers so far
• 1 master theses (T. Heikkilä)
• 2 journal papers
• 6 conference papers
• Budget running as planned

5
Intellectual contribution

• Know how related to
• traffic engineering configurations in multi AS
  environment has been increased
• multicast streaming and configurations
• Developed revenue and QoS aware scheduling
  algorithms has increased our knowledge of the
  fair capacity allocation
• Multicast connections seems to be problem for a
  certain network devices (performance fades)
• Know how related to network design for multicast
  traffic has been increased
• Management models for handling MPLS VPN has been
  proposed for WTSs network management product
  (www.wts.fi NetWrapper)

6
Broader Impact of the project

• Commercial
• Achievements of the project will be
  commercialised by WTS (portfolio of NetWrapper)
• Participants will get ideas and solutions for new
  networked services (We hope operators find
  mechanisms so that they can offer better service
  with better valuemax. their revenue)
• Scientific
• 2 journals and 6 conf. papers have been published
  
• Educational
• 11 PhD students and 5 M.Sc students work for
  TeraII
• Contribution for courses Networking Workshop,
  Advanced Computer Communications

7
Publications so far

• Journals
• J. Joutsensalo and T. Hamalainen Optimal
  Link Allocation and Revenue Maximization. Journal
  of Communications and Networks, Vol. 4, No. 2,
  June 2002, ISSN 1229-2370, pp 136-147.
• K. Kaario, T. Hämäläinen, P. Raatikainen
  Adaptive Parameter Setting for QoS Aware Load
  Balancing Algorithm. WSEAS Transactions on
  Communications, Issue 1, Vol. 1, 2002, ISSN
  1109-2742, pp. 144-149.
• Conference Proceedings
• M. Paakkonen, K. Kaario and T. Hamalainen CoS
  Aware Traffic Flow Balancing in MPLS Networks,
  Proc. of International Conference on
  Telecommunications, June 2002.
• T. Hämäläinen and J. Joutsensalo Network
  Channel Allocation and Revenue Maximization,
  Proc. of APOC 2002, October 2002.
• T. Hämäläinen and J. Joutsensalo Link
  Allocation and Revenue Optimization for Future
  Networks, Proc. of IEEE Globecom 2002, November
  2002
• J. Joutsensalo, T. Hämäläinen M. Pääkkönen, and
  A. Sayenko Adaptive Weighted Fair Scheduling
  Method for Channel Allocation, Proc. of IEEE
  ICC 2003, May 2003.
• A. Sayenko, J. Joutsensalo, T. Hämäläinen"An
  adaptive Approach to WFQ with the Revenue
  Criterion". To be published in Proc. of IEEE
  ISCC'2003, July 2003.
• K. Kaario, P. Raatikainen, T. Hämäläinen, and M.
  Pääkkönen "A Lightpath Allocation Scheme for WDM
  Networks with QoS". To be published in Proc. of
  IEEE ISCC'2003, July 2003.

8
Bottlenecks

• Technical
• Cisco specific IOS compability problems
• some features supported only in a few IOS version
• Educational
• Getting familiar with ns2 simulator environment
  takes some time
• debugging of the code
• checking the code
• Getting familiar with MPLS VPN/BGP routing
  configurations takes also some time

9
Second Project Year

• 1.6.2003-31.5.2004 Things to do
• autenthication and charging methods for multicast
  services
• evaluation of the different MPLS VPN solutions in
  multiAS environment
• evaluation of the different multicast solutions
  at the point of network operator and service
  provider view
• ???
• Publication plan
• Master Thesises
• IP-TV services (auth.charg.) at terabit network
  (R. Riikkola, 12/03)
• Multicast with MPLS (O. Alanen, 5/04)
• XXXXX (N.N)
• Journals and Conference proceedings related to
  following topics
• Combine load balancing algorithm with adaptive
  scheduling
• Adaptive network parameter tuning in multinode
  case
• Continue with traffic classification and pricing
  models
• 2 journal manuscripts and 4 confernce papers will
  be submitted during the second project year.
• Final project report
• Summary of the results

10
Adaptive Weighted Scheduling Algorithm

• Let d0 be the minimum processing time of the
  scheduler
• The number of service classes is denoted by m.
• The processing time is Nid0/wi, , where
• w_i(t)w_i, i1,,m are weights allocated for
  each class, and N_i(t)N_i is a number of
  customers in the ith queue.
• The constraint for the weights are
• wi gt 0 and
• .
• If some weight is wi 1, then the other weights
  are wj 0, j ? i, and class i is served by
  minimum processing time d0, if Ni 1.

11
Adaptive Weighted Scheduling Technique
12
Adaptive Weighted Scheduling Technique
13
Adaptive Weighted Scheduling Technique
Fig1. Three linear pricing functions. Horizontal
axis delay vertical axis price.
14
Adaptive Weighted Scheduling Technique
15
Simulations and results

• In this simulation case, three service classes (m
  3) have pricing functions
• ki has only the vertical shifting effect on the
  revenue (and thus it should be positive in
  practical application largest for
  highest-priority class, and smallest for
  lowest-priority class).

 
16
Simulations and results

• At the beginning of the simulations
  w_i(0)1/m1/3, i.e. the processing delays are
  d_0/w_i(0)md_03.
• To illustrate the performance of the algorithm,
  two other weighted scheduling methods with fixed
  parameters are also used.
• In the first one (denoted by WFQ1), the fixed
  weights are w_10.44, w_20.32, and w_30.24.
• In the second one (denoted by WFQ2), the fixed
  weights are w_10.50, w_20.30, and w_30.20
• Next figures show the revenue and delays as a
  function of time for adaptive scheduling
  algorithm and for fixed ones.

 
17
Revenue as a function of time (fixed weights case
1)
Revenue as a function of time (adaptive)
 
Revenue as a function of time (fixed weights case
2)
18
Delays as a function of time (fixed weights case
1)
Delays as a function of time (adaptive)
 
Delays as a function of time (fixed weights case
2)
19
 
Evolution of the weights as a function of time
with used pricing scenario
20
Simulations and results
The value of the weight of the gold class (blue
pricing curve ) decreases due to the small
arrival rate. Thus the gold class buffer is
almost full. Note, the corresponding delay
caused by almost full buffer is sufficiently
small to keep the revenue paid by gold class
customers high, (can be checked by comparing
delay curve to pricing curve). Same observations
hold for silver and bronze class
weights. Comparing fixed weights cases one can
see that the revenue obtained by adaptive
algorithm is clearly largest, and quite clearly
close to maximum at all the time.
 
21
Result Analysis

• The algorithm is deterministic and
  non-parametric, ie. it uses only the information
  about the number of customers, not about
  statistics of arrival rates or duration
  distributions.
• The algorithm tends to have oscillating behavior,
  as shown in the figure representing the evolution
  of the weights.
• However, there are potentially lots of more
  effective fixed point rules, which can be derived
  from the solution of the first order derivative
  of the Lagrangian-based revenue criterion.
• The algorithm seems to converge linearly in the
  logarithmic scale.
• Also, Call Admission Control (CAC) mechanism can
  be used in the context of the algorithm. It is
  based on the hypothesis testing. This is one of
  our future research topics.

22
Conclusions

• Quality of service
• Not inherent in the IP networks
• Requires congestion management or avoidance
• Reservations are useful as well
• We introduced an adaptive scheduling algorithm,
  which was derived from revenue target function.
• The experiments demonstrated the revenue
  maximization ability of the algorithm, while
  still allocating delays in a fair way.
• In the future, revenue criterion is also used as
  an admission control mechanism.
• In admission control, call is accepted/rejected
  by hypothesis test, where revenue
  increase/decrease is estimated, when call is
  observed.
• This kind of an approach removes the need of
  fitting pricing functions with arrival rates.
• Multi-node case will also be investigated in the
  future. It is important to develop such a
  distributed approximation, which does not suffer
  the curse of dimensionality and computational
  complexity of the optimal global approach
  (signalling