Live virtual machine migration based on future prediction of resource requirements in cloud datacenter

Project Guide:
Dr. G R Gangadharan
Institute for Development & Research in Banking
Technology, Hyderabad
1
Project Trainee:
Tapender Singh Yadav
B.Tech IIIrd Year,
Department of Computer Science & Engineering,
Indian Institute of Technology, Patna

Virtual Machine
 Software-based emulation
 Creation and Management done by Hypervisor (also
known as Virtual Machine Manager)
2
Hardware – “real physical machine”
Virtual Machine Manager (VMM)
Virtual Machine 1 Virtual Machine 2
Operating System 2Operating System 1
APP APP APP APP APP APP APP APP

VM Migration and its need
 Migrating VM from one host to another host is known
as Virtual Machine Migration
 Why we need VM Migration?
Dynamically changing workloads
Maintenance of Host Server
Server downtime due some fault
Ease in migrating OS + Applications from an outdated
host to new host.
3

Current Scenarios
 Lot of research has been done on:
 Workload balancing based on % CPU utilization
 Migration of VM(s) over LAN or WAN
 VM placement based on resource demand prediction
 But there are some factors where the work is not much
significant.
 Less focus on dynamic on-demand requests for
applications
 Less focus on providing better Quality of Service (QoS)
and minimizing the number of Service Level Agreement
(SLA) violations
4

Our Idea and Problem Statement
 There are “M” Virtual Machines (VMs) and their corresponding
resource usage based on the number of cores (processing elements)
used by them, in the past few days on each hour of the day. Based
on this historical data, we have to forecast the future resource
demand for number of cores required by all the “M” VM(s) in the
datacenter using the Trend Seasonality Model.
 Based on the forecasting result, we need to optimize the dynamic
allocation of VM to a best-fit host for migration from “N” available
hosts.
5

Forecasting Method
 Future Demand of the VM(s) are computed based on the
historical resource utilization of the VM(s)
 Trend Seasonality Model used
 Advantages of Trend Seasonality Model:
 Small size of dataset
 Efficient Prediction
7

Forecasting Method Continued…
 Trend is periodic change in the series which evolves slowly.
 Seasonality is the periodic recurrence of a pattern for each period
over the time.
 The “raw” historical data is composed of various components
such as seasonality, trend, irregularities and cyclic oscillations.
 In order to forecast the future resource demand efficiently, we
need to get rid of these components i.e., we need to decompose
(deseasonalize) the “raw” historical data first.
8

Forecasting Method – Deseasonalizing the
raw data
 Following steps are followed for deseasonalizing the raw
data:
1. Compute a centred 24-period moving average for all possible
periods for all given days.
2. Compute the ratio of actual resource demand in each period to
the centred moving average obtained in Step 1.
3. Average the above ratios for periods 1-24 for all given days.
4. Round-off of the averaged ratios from Step 3 to obtain a 24-
period seasonal index values.
5. Divide the actual resource usage by the seasonal indexes to get
the deseasonalized resource usage levels.
9

Forecasting Method – Trend Line Equation
 To find the ‘trend’ component from the raw data, we use trend
line equation.
 Trend Line equation is obtained by applying the Simple Linear
Regression (SLR) on the deseasonalized data and time variable
‘t’.
 Trend Line equation is of the form:
𝑌 = 𝐴 + 𝐵𝑋
where, A = vertical-intercept of the trend line
B = slope of the trend line
10

Forecasting Method – Final step
 Multiply the resource demand trend level from previous step
with the seasonal index for that period to include the seasonality
effect and get the final forecast of the resource demand.
11

Result of Prediction (Forecast)
12
Obtained MSE ≈ 1.3
Forecast
0
2
4
6
8
10
12
1 4 7 10 13 16 19 22 1 4 7 10 13 16 19 22 1 4 7 10 13 16 19 22 1 4 7 10 13 16 19 22 1 4 7 10 13 16 19 22
Day 1
Day 2
Day 3
Day 4
Day 5
Forecast
No. of CPU Cores (a)
Resource Utilization of VM 1 (in # of cores used)
No.ofCPUcores
Resource Utilization of VM 1 (in # of cores used)
No.ofCPUcores

Result – Prediction Step
13
Period Hour Seasonal Index
Trend
Component
Forecast
Day 5 1 0.11 5.40 1
2 0.25 5.40 1
3 0.24 5.40 1
4 0.37 5.40 2
5 0.43 5.40 2
6 0.24 5.40 1
7 0.49 5.39 3
8 1.10 5.39 6
9 1.11 5.39 6
10 0.99 5.39 5
11 1.48 5.39 8
12 1.38 5.39 7
13 1.34 5.39 7
14 1.91 5.39 10
15 1.52 5.39 8
16 1.55 5.39 8
17 2.13 5.39 11
18 1.35 5.39 7
19 0.95 5.38 5
20 1.71 5.38 9
21 1.72 5.38 9
22 1.32 5.38 7
23 0.37 5.38 2
24 0.27 5.38 1

Best optimized VM-host mapping and Live
VM Migration Algorithm
14

Overview of the algorithm
15
For each vm in the
datacenter, predict
the future demand
using trend
seasonality model
Based on the future
demands, create a list
of vm(s) which are
needed to be
migrated
Find the best-fit map
for the vm to host,
and start the
migration process

 Expected decrease in the power consumption due to switching off
of the un-utilized hosts
 Less server downtime during migration
 Less Migration Time
 More scalable and robust
 Best for small and large scale expanding businesses
16
Advantages of the proposed algorithm

Conclusions
 We have proposed a resource demand forecasting technique, a
key step in optimizing the VM-host mapping before the actual
Live VM migration could be triggered.
 Our prediction technique employs data mining and statistical
methods for forecasting (predicting) the future resource demand
of the VM(s).
 We proposed a Live VM migration algorithm, which based on
the future demands, will find the best host for the VM to meet
its future demands.
 Special care has also been taken in case of un-utilized hosts
which would be running unnecessarily and consuming power, to
switch off those hosts in order to save the power consumption
17

Future Work
 Future work on this problem includes experimental
implementation and testing of the proposed Live VM migration
policy using some of the known techniques and proposing an
optimized version of these techniques.
18

Live virtual machine migration based on future prediction of resource requirements in cloud datacenter

More Related Content

What's hot (20)

Viewers also liked (11)

Similar to Live virtual machine migration based on future prediction of resource requirements in cloud datacenter (20)

Recently uploaded (20)

Live virtual machine migration based on future prediction of resource requirements in cloud datacenter