Capacity of Inter-Cloud Layer-2 Virtual Networking!

Transcription

1 Capacity of Inter-Cloud Layer-2 Virtual Networking! Yufeng Xin, Ilya Baldin, Chris Heermann, Anirban Mandal, and Paul Ruth!! Renci, University of North Carolina at Chapel Hill, NC, USA! (!

2 Overview! Introduction and motivation" Distributed Cloud IaaS : Economy of Scale" Applications: high-end, HPC" Inter-Cloud Virtual Networking : Multi-domain, wide-area" Inter-cloud layer-2 networking" Inter-domain VLAN connection " Point-to-point and multi-point connections" Capacity Model" Maximal Number of connections " Model: complete multipartite graph " Static and dynamic capacity" Conclusion" 2!

3 Virtual System Embedding Network topology workflows services etc. Multi-homed cloud hosts with network control Computed embedding

4 Virtual HPC, Condor, Workflow, etc! Workflow Dynamic Slice Condor Head Node (handles initial workflow staging) 1. Start workflow 2. Dynamically create compute nodes 4. Dynamically destroy compute nodes and provisioned netowork Time 3. Network intensive workflow staging Add compute nodes for parallel compute intensive step Dynamically provision compute nodes and network for workflow staging Free unneeded compute nodes after compute step 5. End workflow Montage(workflow( 4!

5 Virtual Networking (1)! Multiple VM interfaces" Management plane: Internet for reachability" Data plane: virtual system networking -> isolation, QoS" VM and data center networking" Layer 3 tunneling: GRE" Layer 2 emulation: VXLAN" Layer 2 VLAN" Wide area networks connecting distributed clouds: multi-domain network environment" IP tunneling: low performance" MPLS: complex and expensive" VLAN connections" Layer-1 optical path" 5!

6 Virtual Networking (2)! Mechanism" Label (tag) for communications channel isolation and identity : IP address, MPLS labels, vlan, lambda, " Bandwidth control: orthogonal to label control" Layer-2:" Cheap, QoS, everywhere" Carrier Ethernet" Dynamic circuits : PNNI, GMPLS, OSCARS, NSI, Stitching" Does it scale??" 6!

7 Laye-2 based Distributed Cloud: a rosy picture!

8 The reality : constraints! Label continuity: label locality vs global " Limited label space : 4096 vlans" Dynamic label path provisioning is not widely deployed : End-to-end automation is difficult" ESNet and I2 (OSCARs)" NSI (GLIF)" No multi-point connection" Presentation title goes here" 8!

9 The reality (2)!! Hybrid environment! Presentation title goes here" 9!

10 The reality : it is hard and not efficient! Challenge:" Static routing and tag assignment with tag continuity constraint is NP-Hard " Tag continuity causes low utilization" Provisioning process is painful and could be long" Solutions : dynamic stitching" Label translation" Label tunneling" Label exchange" End point location neutrality : virtual system" Presentation title goes here" 10!

11 Presentation title goes here" 11!

12 Static Capacity! Still vlan tags are scarce commodity in many networks : 10 vlans out of most Exogeni rack sites now " often the vlan tags are exhausted before the bandwidth is consumed " Inter-cloud network capacity (Static)" maximum number of concurrent inter-cloud connections in the system " Presentation title goes here" 12!

13 such that any two vertices not in the same group oice from each of edge in between, and any two vertices in the sam m to form an endanism. As VLAN do not have an edge in between. We first model Capacity graph cloud system model! like in Fig. 1 with n cloud sites as a ains in our intera novel Label Ex- multipartite graph such that each partite in the g resents the VLAN set offered by the regional netw ying extra VLAN location(s) Complete to fa- n-partite cloud site.. graph. The" i th site, C i, i {1...n}, connec wide area transit networks via m -point request, n cloud afltiple cloud sites " i pre-configured V site sites, ar site will share Cii, its regional cuits by its network regional R network i, M i pre-provioned R i,representedby set M i, M i = m i. We sort and index the site transit network. vlan, i { mn}, 1 connects m 2... to mthe n.then backbone sets may or may re normallynetworks" lap. We create a complete n partite graph G(V differnge points Backbone so that that V = n i=1 networks have unlimited m i, i.e., each VLAN from m i is re by a vertex v V,andthereexistsanedgee vlans" =(v on tag. x Edge e=(v x,vonly j ) E, if " v x M i,v y M j,i j, i, j, which rep a standalone Exslation capability feasible point-to-point connection between two site Since the wide area transit networks provide fu For a connection, " range dynamic circuit services without the constrain setupanend-to- " continuity, we can assume it is possible to create tion between any pair of thepresentation fixedtitle VLAN goes here" 13! tags arate VLAN tags rethetwoincomay, an end-to-end different cloud sites, and thus the edges in the tite graph G. We recall that there is no need to

14 Complete multipartite graph! m 4= =2( Presentation title goes here" 14!

15 Maximum Matching: set of pairwise vertex disjoint edges! Presentation title goes here" 15!

16 For a graph, we call a set of pairwise vertex disjoint edges a matching, and a maximum matching means no other matching has more edges. Clearly, a point-to-point connection between two sites would occupy two VLAN tags represented by the two corresponding vertices in G, whichwouldpre- vent a new connection from using any one of these two tags, Point-to-point connections! i.e., any edges adjacent to these two vertices in G become Theorem infeasible1 inthe the matching. maximum Therefore, number the maximumof number of inter-cloud connections is equivalent to the maximum inter-cloud graph. point- matching to-point of G. equals to the The size of the maximum matching in the complete n partite maximum matching in complete multipartite graph G(V,E) isgivenby[11]: graph. " n 1 M max = min{ i=1 m i, 1 2 Proof: Construction Algorithm" n m i } (1) The size would be equal to the first value if m n n 1 i=1 m i, and the second otherwise. For our example, the maximum matching of four is shown in Fig. 3(a), corresponding to the maximum number of inter-cloud connections that can co-exist in the system in Fig. 1. The formula is easily deduced from constructing the maximum matching, which can also give us the optimal strategy to provision the maximum number of connections simultaneously as follows: i=1 1. If m n n 1 i=1 m i,alwaysplaceoneendofaconnection in the site C n,andtheotherendinanyofthe other n 1siteswithavailableVLAN. 2. else, do following: by the exchange p be generally provis exchange point as Theorem 2. T point broadcasting K-dimensional ma Recall that a hy vertices V and a fa is called a hypered every hyperedge ha if the set of vertic M 1,M 2,...,M n a K partitions in exa system to support acompletek-unifo peredge represents maximum number can co-exist is equa ing in H, K-dimen a K-point hypered For the general NP-Hard [6, 15], b The maximum ma cedure described a Presentation title goes here" 16! us the optimal str of connections, M m

17 Multi-point connection! Theorem 2. The maximum number of inter-cloud K- point broadcasting connections is equivalent to the maximum K-dimensional matching in a complete multipartite hyper- graph. " A hypergraph H = (V,E) consists of a set of vertices V and a family E of subsets of V, where each e E is called a hyperedge. K-uniform if every hyperedge has exactly K vertices " K-point connection : complete K-uniform n-partite hypergraph " Proof : Construction! Presentation title goes here" 17!

18 Evaluation! ExoGeni(testbed:(14(rack(sites( Random(#valns(per(site:(maximum(tag(number:(10,(50,(100,(250,(500,(1000,( 2000(( m1! m2! m3! m4! m5! m6! m7! m8! m9! m10! m11! m12! m13! m14! 5 ( 5 ( 5 ( 6 ( 6 ( 6 ( 7 ( 7 ( 8 ( 9 ( 10 ( 10 ( 10 ( 10 ( 11 ( 13 ( 17 ( 27 ( 30 ( 30 ( 32 ( 35 ( 38 ( 42 ( 43 ( 44 ( 44 ( 47 ( 5 ( 5 ( 7 ( 18 ( 18 ( 18 ( 46 ( 49 ( 59 ( 65 ( 72 ( 72 ( 85 ( 87 ( 17 ( 62 ( 71 ( 106 ( 109 ( 139 ( 150 ( 159 ( 166 ( 181 ( 183 ( 196 ( 205 ( 244 ( 17 ( 56 ( 78 ( 100 ( 178 ( 193 ( 226 ( 228 ( 353 ( 357 ( 391 ( 403 ( 408 ( 496 ( 103 ( 131 ( 138 ( 143 ( 189 ( 244 ( 259 ( 300 ( 321 ( 321 ( 342 ( 729 ( 904 ( 972 ( 62 ( 268 ( 597 ( 658 ( 876 ( 952 ( 1143 ( 1161 ( 1191 ( 1230 ( 1259 ( 1300 ( 1372 ( 1392 ( Presentation title goes here" 18!

19 Result! Number of Connections k=2 k=3 k=4 k= Maximum Number of VLANs Per Site Presentation title goes here" 19!

20 Discussion! Point-to-point connection capacity scales well with number of sites and available tags per sites" Multi-point connection capacity scales much lower" Results can be useful for backbone network dimensioning design" " Presentation title goes here" 20!

21 Further discussion! Models and results can be generalized to other network layers" The graph model can be used to develop new topology embedding algorithms" Dynamic capacity: blocking performance" Maximum connections -> Erlang-B formula" Scheduling with small look-ahead window to archive low blocking performance and high system utilization " Presentation title goes here" 21!

22 Acknowledge! GENI, NSF SDCI, and DOE ASCR Support" Presentation title goes here" 22!