APPENDIX A.1 Heterogeneity aware dynamic VM

153
APPENDIX
A.1 Heterogeneity aware dynamic VM placement and consolidation
controls
Simulation runs results (Table A.1) shows the influence of datacenter
rack or server level temperature coverage controls (Section 3.2.9.1) on energy
consumption. Using server threshold temperature (Temperaturethreshold ) as
350 °K at both (Server and Rack) levels gives the same energy consumption.
Simulation run results (Table A.2) at a reduced server threshold temperature
Temperaturethreshold of 325 °K shows higher energy consumption using rack
than the server level. Moreover, the server level temperature coverage run
(Table A.2) at Temperaturethreshold 325 °K shows a higher SLAV value than
the similar server level temperature coverage run with Temperaturethreshold
350 °K.
Hence, Server level temperature coverage control using
Temperaturethreshold at 350 °K has been finalized for the experimentation
runs.
Table. A.1. Rack vs. server level temperature coverage
(Temperaturethreshold at 350 °K)
Temperature
coverage
approach
Server level
temperature
control
Rack level
temperature
control
Datacenter
Energy
(kWh)
πŽπ€π’π‹π€π•
(%)
πŽπ•πŒππƒ
(x
0.0001)
SLAV
(x
0.0001)
ESV
(x
0.001)
Number of
migrations
30.67
19.72
5.57
1.1
3.43
10809
30.67
19.58
5.60
1.1
3.43
11082
154
Table. A.2. Rack vs. server level temperature coverage
(Temperaturethreshold at 325 °K)
Temperature
coverage
approach
Server level
temperature
control
Rack level
temperature
control
Datacenter
Energy
(kWh)
πŽπ•πŒππƒ
πŽπ€π’π‹π€π• ( x
(%) 0.0001)
SLAV
(x
0.0001)
ESV
(x
0.001)
Number of
migrations
30.39
19.32
22.37
4.3
13.09
39169.66
56.98
12.15
1.80
21.8
1.23
2711.66
Simulation run results (Table A.3) shows the influence of VM migrations
and oscillation effect controls (Section 3.2.9.2). Observe that OASLAV metric
value obtained with VM migrations and oscillation control (VM would be
considered for migration only if the last migration time is greater than 30
seconds) is higher than the one without VM migrations and oscillation
control. This means that higher number of VMs within the server will exhibit
performance impacts due to resource constraints. Average number of VMs
impacted with OASLAV in case of VM migrations and oscillation control is 99
as opposed to 72 when not considering VM migrations and oscillation
controls. Also, using VM migrations and oscillation control reduces the
number of migrations and hence the performance degradation (OVMPD) due
to VM migration is lesser than the run without VM migrations and oscillation
control. But here, the performance overhead due to VM migration is known
and can be addressed easily and hence is lesser of an impact. Based on the
simulation results, impact due to host resource constraint (above 20%) on VM
performance is unknown and hence VM migrations and oscillation controls is
not selected.
Table. A.3. VM migration oscillation controls
Scenario
With VM migration oscillation control
Without VM migration oscillation
control
Datacenter
Energy
πŽπ€π’π‹π€π• πŽπ•πŒππƒ
(kWh)
(%)
(x 0.00001)
28.72
55.88
4.62
30.67
19.58
56.04
155
Simulation run result (Table. A.4) shows the influence of recent host
historical transaction utilization data counts in computing server utilization and
energy consumption (Section 3.2.9.3). With host history transaction count of 2
gives better energy consumption result. Hence, host history transaction count
as 2 is selected in the simulation run.
Table. A.4. Host history transaction count controls
Host history
transaction count
control Scenario
1
2
5
10
20
Datacenter Energy (kWh)
30.55
30.67
30.70
30.74
30.78
SLAV
(x
0.0001)
2.6
1.1
1.1
1.0
1.0
A.2 Server DPM sleep state power transition controls
Simulation run result (Table. A.5) shows the influence of temperature
coverage control at datacenter rack or server level temperature (Section
3.2.9.1). With DPM sleep state run, rack level temperature control approach
gives better energy consumption than server level temperature control
approach but with DPM normal state run, both server and rack level
temperature control approaches have the same energy consumption.
Table. A.5. Rack vs. server level temperature coverage
Temperature
coverage
approach
Server level
temperature
control
Rack level
temperature
control
(1a)+(1b)
Datacenter SLAV
Energy
(x
(kWh)
0.0001)
(1a)+(2)
Datacenter
Energy
SLAV
(kWh)
(x 0.0001)
30.67
1.1
30.02
1.1
30.67
1.1
29.96
1.1
(1a)- Proposed Heterogeneity aware dynamic VM placement and consolidation controls
(1b)- Server DPM normal power state transition controls (off, on, SETUP)
(2)- Server DPM sleep power state transition controls (off, on, sleep, SETUP)
(3)- Dynamic computer room air conditioner (CRAC) controls
156
Simulation run results (Table. A.6) shows the influence of VM
migrations and oscillation effect controls (Section 3.2.9.2). Observe that
OASLAV metric value obtained with VM migrations and oscillation control
(VM would be considered for migration only if the last migration time is
greater than 30 seconds) is higher than the one without VM migrations and
oscillation control. This means that higher number of VMs within the server
will exhibit performance impacts due to resource constraints. Average
number of VMs impacted with OASLAV in case of VM migrations and
oscillation control is 99 as opposed to 72 when not considering VM
migrations and oscillation controls. Also, using VM migrations and oscillation
control reduces the number of migrations and hence the performance
degradation (OVMPD) due to VM migration is lesser than the run without VM
migrations and oscillation control. But here, the performance overhead due to
VM migration is known and can be addressed easily and hence is lesser of an
impact.
Based on the simulation results, though the energy consumption
using oscillation controls is better, the impact due to host resource constraint
on VM performance is unknown and would need further understanding, hence
have not considered VM migrations and oscillation controls.
Table. A.6. VM migration oscillation controls
Scenario
Datacenter
Energy
(kWh)
With VM
migration
oscillation
control
Without VM
migration
oscillation
control
(1a)+(2)
(1a)+(1b)
πŽπ•πŒππƒ
πŽπ€π’π‹π€π• (x
(%)
0.00001)
Datacenter
Energy
(kWh)
πŽπ•πŒππƒ
πŽπ€π’π‹π€π• (x
(%)
0.00001)
28.72
55.88
4.62
27.27
55.15
4.62
30.67
19.58
56.04
30.02
20.11
56.45
(1a)- Proposed Heterogeneity aware dynamic VM placement and consolidation controls
(1b)- Server DPM normal power state transition controls (off, on, SETUP)
(2)- Server DPM sleep power state transition controls (off, on, sleep, SETUP)
(3)- Dynamic computer room air conditioner (CRAC) controls
157
Simulation run result (Table. A.7) shows the influence of recent host
historical transaction utilization data counts in computing server utilization and
energy consumption (Section 3.2.9.3). With host history transaction count of 2
gives better energy consumption result. Hence, host history transaction count
as 2 is selected in the simulation run.
Table. A.7. Host history transaction based server utilization
computation
(1a)+(2)
(1a)+(1b)
Host history
transaction count
control scenario
2
5
10
20
Datacenter
Energy
(kWh)
30.67
30.70
30.74
30.78
SLAV
(x
0.0001)
1.1
1.1
1.0
1.0
Datacenter
Energy
(kWh)
SLAV
(x
0.0001)
30.02
30.02
30.02
30.05
1.1
1.1
1.1
1.0
(1a)- Proposed Heterogeneity aware dynamic VM placement and consolidation controls
(1b)- Server DPM normal power state transition controls (off, on, SETUP)
(2)- Server DPM sleep power state transition controls (off, on, sleep, SETUP)
(3)- Dynamic computer room air conditioner (CRAC) controls
A.3 Dynamic computer room air conditioner thermal controls
Simulation run results (Table. A.8) shows the influence of datacenter
rack or server level temperature coverage computation (Section 3.2.9.1). With
rack level temperature control approach considering DPM sleep state
transitions gives better energy consumption than server level temperature
control approach. Dynamic computer room air conditioner control simulation
run with rack level temperature control, Temperaturethreshold = 325 Kelvin and
Temperatureinlet_threshold = 325 Kelvin consumes higher energy than the run
with Temperaturethreshold = 350 Kelvin and Temperatureinlet_threshold = 325
Kelvin.
158
Table. A.8. Rack vs. server level temperature coverage
Temperature coverage approach
Server level temperature control
with dynamic computer room air
conditioner control (3) with
Temperaturethreshold = 350 Kelvin
and
Temperatureinlet_threshold
= 325 Kelvin
Rack level temperature control with
dynamic computer room air
conditioner control (3) with
Temperaturethreshold = 350 Kelvin
and
Temperatureinlet_threshold
= 325 Kelvin
Server level temperature control
with dynamic computer room air
conditioner control (3) with
Temperaturethreshold = 325 Kelvin
and
Temperatureinlet_threshold
= 325 Kelvin
Rack level temperature control with
dynamic computer room air
conditioner control (3) with
Temperaturethreshold = 325 Kelvin
and
Temperatureinlet_threshold
= 325 Kelvin
(1a)+(1b)+(3)
(1a)+(2)+(3)
Datacenter
Energy
(kWh)
SLAV
(x
0.0001)
Datacenter
Energy
(kWh)
SLAV
(x
0.0001)
26.92
1.16
26.37
1.09
26.90
1.15
26.41
1.11
29.78
3.68
29.24
3.19
126.08
0.230
52.96
0.09
(1a)- Proposed Heterogeneity aware dynamic VM placement and consolidation controls
(1b)- Server DPM normal power state transition controls (off, on, SETUP)
(2)- Server DPM sleep power state transition controls (off, on, sleep, SETUP)
(3)- Dynamic computer room air conditioner (CRAC) controls
Simulation run results (Table. A.9) shows the influence of VM
migrations and oscillation effect controls (Section 3.2.9.2). Observe that
OASLAV metric value obtained with VM migrations and oscillation control
(VM would be considered for migration only if the last migration time is
greater than 30 seconds) is higher than the one without VM migrations and
159
oscillation control. This means that higher number of VMs within the server
will exhibit performance impacts due to resource constraints. Average
number of VMs impacted with OASLAV in case of VM migrations and
oscillation control is 99 as opposed to 72 when not considering VM
migrations and oscillation controls. Also, using VM migrations and oscillation
control reduces the number of migrations and hence the performance
degradation (OVMPD) due to VM migration is lesser than the run without VM
migrations and oscillation control. But here, the performance overhead due to
VM migration is known and can be addressed easily and hence is lesser of an
impact.
Based on the simulation results, impact due to host resource
constraint on VM performance is unknown and hence VM migrations and
oscillation controls is not selected.
Table. A.9. VM migration oscillation controls
Scenario
Dynamic
computer room
air conditioner
control (3) With
VM migration
oscillation
control
Dynamic
computer room
air conditioner
control (3)
Without VM
migration
oscillation
control
(1a)+(2)+(3)
(1a)+(1b)+(3)
Datacent
er
Energy
(kWh)
πŽπ•πŒππƒ
πŽπ€π’π‹π€π• (x
(%)
0.00001)
Datacent
er
Energy
(kWh)
πŽπ•πŒππƒ
πŽπ€π’π‹π€π• (x
(%)
0.00001)
25.25
55.06
4.69
23.91
55.88
4.62
26.92
20.02
57.49
26.37
19.67
55.09
(1a)- Proposed Heterogeneity aware dynamic VM placement and consolidation controls
(1b)- Server DPM normal power state transition controls (off, on, SETUP)
(2)- Server DPM sleep power state transition controls (off, on, sleep, SETUP)
(3)- Dynamic computer room air conditioner (CRAC) controls
Simulation run result (Table. A.10) shows the influence of recent host
historical transaction utilization data counts in computing server utilization and
160
energy consumption (Section 3.2.9.3). With host history transaction count of 2
gives better energy consumption result. Hence, host history transaction count
as 2 is selected in the simulation run.
Table. A.10. Host history transaction based server utilization
computation
Host history
transaction
count control
scenario
2
5
10
20
(1a)+(1b)+(3)
(1a)+(2)+(3)
Datacenter
SLAV
Datacenter
SLAV
Energy
(x
Energy
(x
(kWh)
0.0001)
(kWh)
0.0001)
26.92
1.2
26.37
1.1
26.96
1.2
26.35
1.1
26.96
1.1
26.37
1.0
27.03
0.98
26.30
0.96
(1a)- Proposed Heterogeneity aware dynamic VM placement and consolidation controls
(1b)- Server DPM normal power state transition controls (off, on, SETUP)
(2)- Server DPM sleep power state transition controls (off, on, sleep, SETUP)
(3)- Dynamic computer room air conditioner (CRAC) controls