Explicit Application- Network Cross-layer Optimisation by dfsiopmhy6

VIEWS: 16 PAGES: 8

									Prognet06                                                                                    14/12/2006




                                      Explicit Application-
                                      Network Cross-layer
                                      Optimisation
                                      Dimitrios Pezaros

                                      Computing Department
                                      Lancaster University
                                      [dp@comp.lancs.ac.uk]


             12 Dec 2006                         Prognet’06




             Overlay Networks and the Underlay
             Internet
                 Increasingly popular to deploy customisable services at
                 the application layer
                      File-sharing, object location, application-level multicast, network-
                      embedded storage, etc.
                      Implicitly or explicitly take control over routing
                      Create highly variant & unpredictable traffic dynamics

                 Multiple overlays over the underlay Internet
                      Can destabilise ISPs’ TE and load-balancing policies at the
                      underlay
                      Cost-ineffective for overlays to independently probe the underlay
                      to optimise their operation
                      Fast reaction to underlay events ignoring background traffic


             12 Dec 2006




D. Pezaros                                                                                           1
Prognet06                                                                                      14/12/2006




             Cross-layer Issues
                 Topology congruence not preserved
                      Black-box view of the underlay
                      Path selection based on an overlay cost minimisation function
                 Underlay path measurement load
                      E.g. ping storms: 1 GB/day of ping traffic on planetlab
                 ISP Traffic Matrices (TM)s
                      Entry duplication: overlay routing through intermediate access PoPs
                      Entry shift: alteration of ISP exit point for inter-domain traffic
                      ISPs assume stable traffic demands over short timescales
                 Destabilisation of routing and Traffic Engineering (TE) policies
                      Enforce routing through high-(IP-)cost links
                 Traffic oscillation
                      Two closed-loop systems react independently to the same set of events
                      Multiple overlays synchronising reaction to underlay events, e.g. link
                      failures
                 Inter-domain decoupling violation
                      Network state in one domain can influence behaviour in another

             12 Dec 2006




             Proposed solutions: Underlay evolution
             vs. Explicit Cross-layer interaction
               Enrich underlay service to alleviate the need for overlays
                    Purposes and services offered by overlays are diverse & global


               Information exchange between overlay and underlay
                    Underlay exposes topology and performance information. Overlay
                    exposes structure. However…
                    Global topology information exposure may not scale
                    Makes ISPs’ policies even more vulnerable to opportunistic
                    overlays
                    Conflict of business interests between stakeholders
                           Cross-layer conflict: between ISP and overlay applications
                           Intra-layer conflict: among ISPs


             12 Dec 2006




D. Pezaros                                                                                             2
Prognet06                                                                                     14/12/2006




             ISP Hints: Explicit Cross-Layer
             Interaction
                 Explicit cross-layer interaction without exchange of
                 structural information
                      Deploy services at the ISP level to take explicit input from the
                      application and return hints about this input
                      Hints aim at improving network resource utilisation
                           E.g. minimise intensity and length of bursty requests; minimise
                           duplicate inter-domain load
                      Applications should also have incentives to use ISP hints
                           Improved overlay performance and user satisfaction
                 It is up to the respective layers
                      What inputs mean, how to use them, how to form hints
                      Each layer retains its own control to:
                           Set policies, optimise given performance metrics

             12 Dec 2006




             ISP Hints: Examples
                 Distance hint
                      Input: a set of destination IP addresses;
                      Output hint: a distance metric hint for each
                           ISP chooses what the metric is (hop count, AS path length, RTTs,
                           etc.) and its resolution
                 Similarity hint
                      Input: a set of destination IP addresses;
                      Output hint: clustered addresses based on a “similarity” metric
                           E.g. cluster all destination seen through the same egress link
                 Hints can be composed by:
                      A set of Input/output formats; some semantics description
                      An ISP domain-local implementation
                 Much of the information needed is already readily
                 available to the ISPs
             12 Dec 2006




D. Pezaros                                                                                            3
Prognet06                                                                                       14/12/2006




             Applications Using ISP Hints
                 Client-server
                      Use similarity hint to cluster requests based on the egress links
                      response traffic is routed through
                      Employ intra-cluster FCFS
                           Favouring inter-similarity cluster concurrency increases load
                           balancing
                      Better handling of flash crowds
                      Increased effective network capacity

                 Peer-to-peer file-sharing overlay
                      Employ similarity and distance hints simultaneously
                           Cluster requests per-egress link
                           Prefer minimum AS-hop distance providers among alternatives
                      Improved overlay/underlay topology congruence
                      Spread popular content faster to physically diverse segments of
                      the Internet

             12 Dec 2006




             Client-Server Performance using
             ISP Hints




                100 clients                        200 clients                    600 clients




             12 Dec 2006




D. Pezaros                                                                                              4
Prognet06                                                                                                                                                                                                                             14/12/2006




             P2P File Sharing Overlay
                                          Designed a bittorrent-like overlay to examine
                                          performance improvements of employing ISP hints
                                                  For the underlay
                                                  And the overlay
                                          Examined how a chunk (1MB) of data is distributed to
                                          the overlay nodes
                                                  During flash-crowd and steady-state phases
                                          Three different algorithms compared and contrasted
                                                  Simple overlay: no cross-layer interaction
                                                  ISP hints using domain-local knowledge
                                                  ISP hints using overlay-global knowledge


             12 Dec 2006




             Simultaneous Internet-wide application
             throughput and link stress improvement
                                                  Individual Transfer Throughput                                            Mean Link Stress                                                  Maximum Link Stress
                                              Min-AS-Hop Distance Provider Selection                              Min-AS-Hop Distance Provider Selection                              Min-AS-Hop Distance Provider Selection
                                                                                                            1.0




                                                                                                                                                                                2.5




                                                           AS-Local Region-Aw are Algorithm                                    AS-Local Region-Aw are Algorithm                                    AS-Local Region-Aw are Algorithm
                                       1400




                                                           Region-Aw are Overlay                                               Region-Aw are Overlay                                               Region-Aw are Overlay
                                                           Simple Overlay                                                      Simple Overlay                                                      Simple Overlay
                                       1200




                                                                                                            0.8




                                                                                                                                                                                2.0
                                       1000
              Mean Throughput (KB/s)




                                                                                                            0.6




                                                                                                                                                                                1.5
                                                                                              Link Stress




                                                                                                                                                                  Link Stress
                                       800
                                       600




                                                                                                            0.4




                                                                                                                                                                                1.0
                                       400




                                                                                                            0.2




                                                                                                                                                                                0.5
                                       200




                                                                                                            0.0




                                                                                                                                                                                0.0
                                       0




                                                  2               3                 4                                 2              3                  4                                 2               3                 4

                                                         Access Edge Degree                                                  Access Edge Degree                                                  Access Edge Degree




             12 Dec 2006




D. Pezaros                                                                                                                                                                                                                                    5
Prognet06                                                                  14/12/2006




             Going Deeper…
                                                    E

                                B




                 A                      D               F

             Leaf AS                            G


                                C                              I
                                                    H



             12 Dec 2006




             Going Deeper: Flat Second-hop
             Clustering
                                    X requests


                                            A



                                                            5X/8
                     3X/8
                            B                           D



                  D         E       F               B   E      C   G   F




             12 Dec 2006




D. Pezaros                                                                         6
Prognet06                                                                                                                                                                                                                          14/12/2006




                                         Going Deeper: Hierarchical
                                         Second-hop Clustering
                                                                                            X requests


                                                                                                                 A
                                                                                 X/2                                            X/2


                                                                                                                                                           5X/10
                                                     3X/6
                                                                      B                                                                       D



                                                 D                    E                     F                              B                  E                     C                       G                   F




                                         12 Dec 2006




                                         How much does “deeper” knowledge
                                         give back?
                                                 Individual Transfer Throughput                                            Mean Link Stress                                                 Maximum Link Stress
                                             Min-AS-Hop Distance Provider Selection                              Min-AS-Hop Distance Provider Selection                             Min-AS-Hop Distance Provider Selection
                                                                                                           1.0




                                                                                                                                                                              3.0
                                      1400




                                                         AS-Local Region-Aw are Algorithm                                    AS-Local Region-Aw are Algorithm                                   AS-Local Region-Aw are Algorithm
                                                         Flat Clustering                                                     Flat Clustering                                                    Flat Clustering
                                                         AS-Local Region-Aw are Algorithm                                    AS-Local Region-Aw are Algorithm                                   AS-Local Region-Aw are Algorithm
                                                         Hierarchical Clustering                                             Hierarchical Clustering                                            Hierarchical Clustering
                                                         Region-Aw are Overlay                                               Region-Aw are Overlay                                              Region-Aw are Overlay
                                                                                                                                                                              2.5




                                                         Flat Clustering                                                     Flat Clustering                                                    Flat Clustering
                                      1200




                                                         Region-Aw are Overlay                                               Region-Aw are Overlay                                              Region-Aw are Overlay
                                                                                                           0.8




                                                         Hierarchical Clustering                                             Hierarchical Clustering                                            Hierarchical Clustering
                                                         Simple Overlay                                                      Simple Overlay                                                     Simple Overlay
                                      1000




                                                                                                                                                                              2.0
             Mean Throughput (KB/s)




                                                                                                           0.6
                                                                                             Link Stress




                                                                                                                                                                Link Stress
                                      800




                                                                                                                                                                              1.5
                                      600




                                                                                                           0.4




                                                                                                                                                                              1.0
                                      400




                                                                                                           0.2




                                                                                                                                                                              0.5
                                      200




                                                                                                           0.0




                                                                                                                                                                              0.0
                                      0




                                                 2              3                 4                                  2              3                 4                                 2              3                 4

                                                        Access Edge Degree                                                  Access Edge Degree                                                 Access Edge Degree




                                         12 Dec 2006




D. Pezaros                                                                                                                                                                                                                                 7
Prognet06                                                                    14/12/2006




             Conclusions
                 Synergistic cross-layer operation can improve perceived
                 performance at both layers

                 Explicit cross-layer interaction
                      With no explicit layer-specific information exchange


                 Even when deployed using readily available information
                 at the ISP level it can make a difference

                 Enforcement
                      May not be necessary
                      Can be seen as a value-added service

             12 Dec 2006




D. Pezaros                                                                           8

								
To top