A Computation- and Network-Aware Energy Optimization model for - PowerPoint PPT Presentation

A Computation- and Network-Aware Energy Optimization model for Virtual Machine Allocation C. Canali, R. Lancellotti Department of Engineering “Enzo Ferrari”, University of Modena and Reggio Emilia M. Shojafar Italian National Council for Telecommunications (CNIT) CLOSER 2017, April 24-26, Porto 1

Motivation ● Energy consumption in Cloud – Typical problem of server consolidation, but... – Network-related energy is often neglected – VMs migration: additional energy consumption ● Challenges of future Cloud systems – Network softwarization: SDN → SDDC – Dynamic VMs behavior → VMs migrations CLOSER 2017, April 24-26, Porto 2

Reference Scenario ● C o m p l e x n e t w o r k t o p o l o g y ● I n t e r a c t i o n b e t w e e n n e t w o r k a n d a l l o c a t i o n m g r . CLOSER 2017, April 24-26, Porto 3

Model ● Multi-dimensional bin packing problem ● Use of dynamic programming: – Time divided in time slots – Start from placement at previous time slot – Define migrations of VMs – Turn ON/OFF servers ● G o a l s – Minimize energy consumption – No parameters to tune CLOSER 2017, April 24-26, Porto 4

Objective function ● E obj : 3 components (in most complete form) ● Energy for computation E C ● Energy for data transfer E D ● Energy for VMs migrations E M CLOSER 2017, April 24-26, Porto 5

Objective function ● Energy for computation E C ● Minimum energy consumption for servers turned ON ● Linear dependence from the CPU utilization of servers CLOSER 2017, April 24-26, Porto 6

Objective function ● Energy for data transfer E D ● Minimum energy for network interfaces in servers turned ON ● Linear dependence on data exchanged among servers ● Captures network topology through parameter E di1,di2 CLOSER 2017, April 24-26, Porto 7

Objective function ● Energy for VMs migrations E M ● Computational overhead for servers ● Data transfer of VM memory CLOSER 2017, April 24-26, Porto 8

Constraints ● Resource requests by VMs on a server must not exceed server capacity: – CPU – Memory – Network: no data exchange within the server ● VMs allocation only on servers turned ON CLOSER 2017, April 24-26, Porto 9

Constraints ● Each VM is placed one and only one server ● Consistency of VMs migrations ● Boolean nature of the problem CLOSER 2017, April 24-26, Porto 10

Considered Alternatives ● Our proposal – Migration Aware (MA) E obj =E C +E D +E M ● Existing solutions: – No Migration Aware (NMA) E obj =E C +E D – No Network Aware (NNA) E obj =E C CLOSER 2017, April 24-26, Porto 11

Simulation Setup ● Resource requests from real VMs – Default: 80 VMs ● Power consumption from datasheets ● Two network behavior scenarios: – Network 1: Random interaction – Network 2: Pareto law interaction (20% of destination IPs receive 80% of traffic) ● AMPL problem formulation – CPLEX 12 solver CLOSER 2017, April 24-26, Porto 12

Comparison Network 1 Network 2 CLOSER 2017, April 24-26, Porto 13

Impact of Migration Network 1 Network 2 CLOSER 2017, April 24-26, Porto 14

Results stability CLOSER 2017, April 24-26, Porto 15

Conclusions ● Challenges of VMs placement in cloud – Network becomes more important (SDDC) – More dynamic VMs behavior (migrations) ● Limitation of existing models ● → Migration-Aware model for VMs placement – No parameter tuning required ● Future work: – More focus on SDDC, model improvement – Heuristics for scalability CLOSER 2017, April 24-26, Porto 16

A Computation- and Network-Aware Energy Optimization model for Virtual Machine Allocation Contact: Riccardo Lancellotti riccardo.lancellotti@unimore.it CLOSER 2017, April 24-26, Porto 17