B99705021 資管三 李奕德 http://ppt.cc/41rH Introduction Background Virtual machine placement Algorithm Algorithm evaluation Result Discussion and future work Scalability issue Aim to solve different problem - Dcell, Bcube, PortLand, VL2…… No thinking of traffic issue - high traffic from end to end 1. 2. 3. three character of all traffic average pairwise traffic rate & end-to-end cost has low correlation Uneven between VMs Stays almost the same Traffic-aware placement may be beneficial Traffic-aware VM Placement Problem (TVMPP) given: traffic matrix , cost matrix Goal: minimize cost Cost can be: Total switch used/Compute Time An algorithm that solve the NP-hard problem Architecture difference NP: by nondeterministic algorithms in polynomial time nondeterministic -Every “guess by hunch” is right at least as hard as the hardest problems in NP Introduction Background Virtual machine placement Algorithm Algorithm evaluation Result Discussion and future work Data set I : IBM Global Services’ data warehouse About 17000 virtual machines Data set II: Server cluster About Hundreds of virtual machines round-trip latency measurement at 68 VM Uneven between VMs 80% of VM’s traffic < 800kb/sec 4% of VM’s traffic > 8mb/sec Stays almost the same Low correlation between average pairwise traffic rate & end-to-end cost Correlation : -0.32 Old style VL2 Portland Bcube Introduction Background Virtual machine placement Algorithm Algorithm evaluation Result Discussion and future work n VM to assign n slot for VM static and single-path routing Cost and traffic matrix from historical data Cost D C e g i , j 1,...,n ij i j i 1,...,n i i is equivalent of finding min tr DX T C T X eX T g T X Dummy VM is assigned when no. slot > no. VM Quadratic Assignment Problem (NP-hard) Impossible to find optimality when size > 15 TVMPP is a special case of QAP reduction from Balanced Minimum K-cut Problem (BMKP) BMKP: extended problem from the Minimum Bisection Problem (MBP) BMKP & MBP are NP-hard Introduction Background Virtual machine placement Algorithm Algorithm evaluation Result Discussion and future work approximation algorithm Cluster-and-Cut Divide VM into VM cluster Divide slot into slot cluster Put VM cluster into slot cluster A smaller problem Feasible when size is sufficient small Complexity determine by SlotClustering and VMMinKcut Slotclustering: O(nk) VMMinKcut: O(n4) Total complexity = O(n4) Introduction Background Virtual machine placement Algorithm Algorithm evaluation Result Discussion and future work Cluster and cut VS. other benchmark algorithms Local Optimal Pairwise Interchange (LOPI) Simulated Annealing (SA) hybrid traffic model Gravity model compute the GLB for each settings Introduction Background Virtual machine placement Algorithm Algorithm evaluation Result Discussion and future work Cost matrix Compare with random assign Traffic is assumed to be in normal distribution Variance is change to show difference Different architecture & variance affect result View as VM cluster GLB prediction GLB prediction VS. optimal solution Thing that brings better performance: - bigger variance - smaller cluster (less VM in a group) - Architecture difference (generally) Bcube > tree > fat-tree > VL2 Good scenario: multiple service in a data center Bad scenario: single service / map-reduce Introduction Background Virtual machine placement Algorithm Algorithm evaluation Result Discussion and future work Dynamic VM placement Other VM placement with different goal Thank you for your attention