Migration of groups of virtual machines in distributed data centers to reduce cost
1. Migration of groups of virtual
machines in distributed data
centers to reduce cost
Sabidur Rahman
Netlab Friday Group Meeting
Feb 17, 2017
http://www.linkedin.com/in/kmsabidurrahman/
krahman@ucdavis.edu
2. Paper review
“Energy-aware migration of groups of virtual
machines in distributed data centers”
by
Rodrigo A. C. da Silvaa and Nelson L. S. da Fonseca
from
Institute of Computing
State University of Campinas, Brazil
published in
Global Communications Conference (GLOBECOM), 2016.
3. Paper review
Introduction:
Select groups of virtual machines (VMs) to be migrated
Select VM groups with network proximity in order to increase potential
number of equipment to be switched off
VMs are migrated only if it results in energy savings
Consolidate workload to take advantage of underutilized servers
Switch off physical resources to gain energy savings
Novelty:
“We consider workload migration by choosing groups of VMs rather than the
entire workload of a data center. Moreover, we analyze the effects of the
data center network topology on energy consumption, when choosing the
virtual machines to be migrated.”
da Silva, Rodrigo AC, and Nelson LS da Fonseca. "Energy-Aware Migration of Groups of Virtual Machines in Distributed Data Centers."
Global Communications Conference (GLOBECOM), 2016 IEEE. IEEE, 2016.
5. Migration algorithm
Migration decisions involve two steps:
Selection (SEL) algorithm: selection of potential sets of VMs in a data center
to be migrated. SEL runs in source DCs. Output of the SEL algorithm is used
by NEG algorithm.
Negotiation (NEG) algorithm: negotiation of migration of these potential sets
with other data centers. NEG runs in destination DCs (potential host DCs)
9. Performance evaluation
• Topology-aware threshold (TT): considers topology correlation when
migration
• Random Threshold (RT): migrates random VM, no correlation
• TT and TR policies always choose a fixed fraction (10%)of
the workload of the data center
• Algorithm is run 8 hours interval, to minimize large transfers across
backbone network
13. Energy consumption model
Three components:
Servers: Idle power 70% of full load power. Linearly grows with
load.
Switches: Chassis, line cards and ports.
ri = Potential transmission rate.
Cooling infrastructure: Derived from PUE.
15. Traffic model
• Group size: medium and large
• Traffic intensity: low, medium, high
V. Paxson, “Fast, approximate synthesis of fractional gaussian noise for generating self-similar network traffic,”
SIGCOMM Comput. Commun. Rev., vol. 27, no. 5, pp. 5–18, Oct. 1997