Professional Documents
Culture Documents
Eddie Pettis
Google Inc.
Mountain View, CA
epettis@google.com
I.
INTRODUCTION
242
II.
RELATED WORK
243
10
Centralized SoA
Distributed SoA
8
6
4
2
0.48 0.46
0
285 295 305 315 325
Per server magnitude of peak
power pulse (W)
TABLE I.
Hierarchy Level
Server
Rack
PDU
Cluster
Size of a group
1
20
200
1000
244
Effective Capacity
(Ah)
20
(1)
!"#$ %
'()
&
*+%
),,
20
Current (A)
40
100000
LFP
LA
10000
1000
100
0
50
DoD level
100
terms of hours and obtain its value from the data sheets,
which is generally 20 hours [13]. Peukerts exponent is
shown by k, which changes depending on the battery type.
For LA batteries, the typical value is around 1.15 whereas for
LFP batteries it is 1.05 [23]. Effective capacity is also scaled
with SoH value to reflect the capacity loss as the battery is
used. The DoD is subtracted from the SoC at the end of each
interval. When discharging ends, we save the total DoD
value during that discharge period,
- as (100
/)%.
Peukerts law states the effective capacity of a battery
decreases with higher discharge current. Figure 4 shows this
negative effect on 20Ah LA and LFP batteries. The
horizontal and vertical axes show the effective battery
capacity and discharging current respectively. The effective
capacity of the LA battery decreases faster due to its greater
nonlinear behavior, represented by a larger Peukert
exponent. At 40A, corresponding to 2C rate for both of these
batteries, the LA battery loses 42% of its nominal capacity,
but the LFP battery loses only 15%.
We update the battery SoH after a complete
charge/discharge cycle [22]. This update depends on the
battery chemistry, determining Peukerts exponent,
and
- . The number of charge/discharge cycles
decreases with deeper discharges, represented by a larger
- value. We use a lookup table derived from
effective capacity graphs provided in commonly available
battery data sheets; similar to Figure 5 for each battery
chemistry to define the effects of
- .
In Figure 5, the horizontal axis shows the DoD level for
charge/discharge at 20h discharge rate, which is defined as
the current that drains the battery in 20h. The vertical axis is
on a log scale and illustrates the number of cycles a battery
can provide for a particular DoD level. As the battery is
discharged deeper in each cycle, the available number of
charge/discharge cycles decreases exponentially. LFP
batteries provide 5x more cycle life compared to LA
batteries in average.
We normalize the effect of one cycle with
- value to calculate its impact on the battery lifetime.
The battery lifetime is defined as the interval in which
battery SoH is greater than a state of health value which
determines when the battery is dead, /4 . Battery
manufacturers generally recommend 80% for this value [24]
[25], i.e. the battery is considered dead if the maximum
capacity it can provide falls below 80% of its rated capacity.
If the battery has 56789:+:_
- cycles with
-
value, the battery SoH is updated as:
(2)
10
where
is the length of the time interval and
is
the discharge current. DoD is computed as:
=
15
LFP
LA
Number of Cycles
(3)
245
)
;<><_!?
nonlinear nature of Peukerts Law, despite additional DCAC conversion losses in the centralized UPS. It obtains
improved battery lifetime and requires significantly less
communication overhead than SoA distributed designs.
(4)
Battery
Li-Ion5
Li-Ion6
Li-Ion7
Error
4.35% 2.05%
5.83% 3.60%
3.84% 2.75%
ED = K
NO
(6)
246
= K
NO Q
Placement
Selective Discharge
Adaptive Current
Battery Model
Coordination
Medium
SoA [2]
centralized
Centralized
X
X
X
N/A
V.
Our design
Centralized
9
9
9
Dedicated
network
COST MODELS
(7)
9
H+H = 6R GGH+H +
O
T
UH+H
(8)
(9)
where 9
H+H (\W
8]Y89)
represents total cost with and without batteries, respectively.
When calculating the total cost without the batteries, we can
just neglect the battery related parts of Equation 8. The main
purpose of peak shaving in this case is to reduce the
provisioned power level so that the co-location renters can
contract for less power.
B. Total Cost of Ownership (TCO) Model
There are several companies that own their data centers,
where they still make power contracts based on their peak
power consumption to reduce their cost of energy. However,
247
values for both this model and CLR model. Further details of
the TCO model are not in the scope of this paper and
covered in detail in [5].
Our grid-tie design requires more power distribution
infrastructure than the distributed design because we keep
transmitting power throughout the data center, even if the
power is not drawn from the utility. For example, a 10MW
data center may have 1MW worth of additional servers due
to peak shaving. In our case, the extra power is provided
from the UPS through the data center power infrastructure to
the servers. In the distributed case, this extra power is not
provided through the data center power infrastructure.
Although all the servers are connected to the main power
infrastructure, during a peak pulse some of them may
disconnect themselves from the main power infrastructure
and get power locally from the on-board UPS. Therefore, the
provisioned power infrastructure is sufficient. This means
that our approach has constant power infrastructure
depreciation, whereas the distributed design decreases this
depreciation with more servers. But, our design does not
require a custom PSU or power distribution, as opposed to
the DC architecture. This makes it practical for the existing
data centers. The additional peak shaving opportunities from
our approach outweigh the additional infrastructure costs.
Input
Battery unit price -rated with 20h
Per server capacity
Peukerts exponent
Battery nominal voltage
Data center depreciation time
Server depreciation time
Utility energy price
Utility power price
LA Value
LFP value
2 $/Ah [30]
5 $/Ah [29]
20 Ah [5]
40 Ah [5]
1.15 [23]
1.05 [23]
12V [11]
10 years [31]
4 years [31]
4.7/kWh [32]
12 $/kW [2]
VI.
METHODOLOGY
TCO/SERVER BREAK-DOWN IN DIFFERENT DESIGNS [5]. THE COMPONENTS WITH DIFFERENT TRENDS ARE HIGHLIGHTED.
TCO Component
w/o peak
shaving
$3.40
$0.13
$5.94
$2.46
$8.97
$7.49
$31.25
$1.56
$1.94
$8.71
$4.20
$76.04
248
Grid-tie Design
TCO/server trend
Breakwith extra servers
down
Decreasing
$2.72
Constant
$3.33
Constant
$5.94
Decreasing
$1.96
Decreasing
$7.17
Decreasing
$5.99
Constant
$31.25
Constant
$1.56
Constant
$1.94
Constant
$8.71
Decreasing
$3.36
Decreasing
$73.94
TABLE VII.
Average Error
3%
6%
8%
WORKLOAD PARAMETERS
Workload
Search [6]
Social Networking [34]
MapReduce [37]
Average Time
Service Interarrival
50ms
42ms
1sec
445ms
2 min
3.3 min
20
Discharge Cur.(A)
Parameter
Avg. Power Consumption
90th%tile Services QoS
Avg. MapReduce Comp. Time
Discharge Cur.(A)
TABLE VI.
15
10
5
0
0
24
48
Time (h)
72
20
15
10
5
0
0
24
48
Time (h)
72
Figure 9. Avg. discharging current for the distributed design (left) and gridtie design (right) over a 3 day period, with LFP.
249
LA
3 years
1.4 years
LFP
10 years
4.1 years
TABLE IX.
Cost
Lifetime
1
2
3
4
7
10
LA distributed design
2 $/Ah
3 $/Ah
4 $/Ah
CLR Savings (%)
0.9
-3.6
-8.1
5.5
3.2
0.9
6.4
4.6
2.7
8.2
5.9
4.6
5 $/Ah
-12.7
-1.3
1
3.2
Not possible
TABLE X.
Cost
Lifetime
1
2
3
4
7
10
12 $/Ah
<-50
-34
-24
-13
-1.8
9.1
LA distributed design
2 $/Ah
3 $/Ah
4 $/Ah
TCO/server Savings (%)
0.02
-2.08
-4.18
2.21
1.2
0.2
2.65
1.86
1.07
3.09
2.52
1.95
5 $/Ah
-6.27
-0.81
0.28
1.38
Not possible
Cost
Lifetime
1
2
3
4
7
10
250
TABLE XIII.
EFFICIENCY OF CENTRALIZED VS. DISTRIBUTED DESIGNS CONSIDERING DIFFERENT POWER EQUIPMENT AND DELIVERY OPTIONS
Unit
Centralized double
conversion UPS
AC distribution PDU
Server AC PSU
DC UPS
DC distribution PDU
Server DC PSU
Filter + Rectifier
Grid-tie inverter
Efficiency
common
Efficiency
best
85% [47]
90% [19]
98% [19]
75% [46]
90% [19]
92% [21], [47]
95% [19]
99% [19]
92% [19]
95% [19]
97% [48]
95% [45]
Design
Distributed
w/ AC
power
Distributed
w/ DC
power
SoA
Centralized
5%
2%
35%
23%
2%
2%
12%
8%
38%
15%
5%
5%
251
TABLE XIV.
10000
25
1000
20
100
15
10
10
0.1
Peak Shaving %
0
Rack
PDU
Cluster
Distibuted Design
Total delay - ideal
Total delay - normal congestion
Grid-tie
Design
VIII. CONCLUSION
Peak shaving with batteries has gained significant
importance because of its ease of applicability and great
performance. In this paper, we address the key challenges of
architecting a cost and energy efficient battery-based peak
shaving design. We first use a detailed battery model to
capture the negative effects of high discharging currents. Our
results indicates that not having a detailed battery model
overestimates the battery lifetime up to 2.44x and leads to
3.35x error in cost saving estimates. Second, we propose a
new grid-tie based design which preserves the advantages of
the existing designs, such as individual control of the
batteries, and eliminates the key drawbacks, such as
adaptively selecting the discharge current. It can use a fast,
dedicated network to coordinate the batteries, reducing the
communication overhead by 4x compared to the distributed
design. Our design achieves up to 78% longer battery
lifetime and doubles the savings compared to the state-ofthe-art designs.
TABLE XV. GRID-TIE VS. DISTRIBUTED DESIGN, HIGHLIGHTED BEST VALUES. (EB = EXTRA BATTERIES, BL = BATTERY LIFETIME, PS = PEAK SHAVING, ES = EXTRA SERVERS)
LFP
Design - Policy
Grid-tie all_battery
Grid-tie
discrete_current
Distributed -N/A
EB
8%
0%
8%
0%
N/A
LA
BL
PS
ES
CLR
savings
6.4
yrs
5.4
yrs
20%
16%
20%
16%
19%
25%
19%
25%
19%
24%
11%
7.7%
9.4%
6.1%
6.4%
4 yrs
252
TCO/server
savings
2.77%
1.36%
2.77%
1.36%
1.86%
BL
2.5 yrs
2.2 yrs
1.4 yrs
PS
ES
CLR
savings
TCO/server
savings
9.9%
7%
9.9%
7%
9.5%
11%
8%
11%
8%
10.5%
5.5%
2.8%
4.9%
2.2%
2.7%
1.87%
1.14%
1.44%
0.42%
0.9%
ACKNOWLEDGEMENTS
coulomb counting method for estimating state-of-charge and state-ofhealth of lithium-ion batteries," Applied Energy, vol. 86, no. 9, 2009.
[23] F. Harvey, "Table with Peukert's exponent for different battery
models,"2001.
http://www.electricmotorsport.com/store/ems_ev_parts_batteries.php.
[24] PowerSonic, "Technical manual of LA batteries,"
http://www.power-sonic.com/technical.php.
[25] P. Rong and M. Pedram, "An analytical model for predicting the
remaining battery capacity of lithium-ion batteries," IEEE
Transactions on VLSI Systems 2006.
[26] B. Saha and K. Goebel, "Battery Data Set, NASA Ames Prognostics
Data Repository," 2007.
http://ti.arc.nasa.gov/tech/dash/pcoe/prognostic-data-repository/.
[27] H. Qian , W. Yu , J.-S. Lai and J. Zhang , "A High-Efficiency Grid-Tie
Battery Energy Storage System," in IEEE TPEL 2011.
[28] APC, "InfraStruxure Total Cost of Ownership, Infrastructure cost
report," 2008. http://www.apc.com/tools/isx/tco/.
[29] A. P. P. Corp, "Portable Power Product design, assemble and quality
control," 2000. http://www.batteryspace.com/lifepo4cellspacks.aspx.
[30] E. motor sport, "EV construction, thundersky batteries," 2001.
http://www.electricmotorsport.com/store/ems_ev_parts_batteries.php.
[31] L. A. Barosso and U. Holzle, The data center as a computer: An
Introduction to the Design of Warehouse-Scale Machines, 2009.
[32] D. E. Carolinas, "Utility bill tariff," 2009.
http://www.duke-energy.com/pdfs/scscheduleopt.pdf.
[33] "RUBiS," http://rubis.ow2.org/.
[34] Apache, http://incubator.apache.org/olio/.
[35] "Hadoop," http://hadoop.apache.org/.
[36] B. Aksanli, J. Venkatesh, L. Zhang and T. Rosing, "Utilizing green
energy prediction to schedule mixed batch and service jobs in data
centers," in HotPower 2011.
[37] Y. Chen, A. Ganapathi, R. Griffith and R. Katz, "The case for
evaluating MapReduce performance using workload suites," In Tech.
Report No. UCB/EECS-2011-21, 2011.
[38] "Google Transparency Report,"
http://www.google.com/transparencyreport/traffic.
[39] "Battery
prices
at
different
current
rates,"
http://www.batteryspace.com/.
[40] D. Kliazovich, P. Bouvry, Y. Audzevich and S. Khan, "GreenCloud: A
Packet-Level Simulator of Energy-Aware Cloud Computing Data
Centers," in GLOBECOM 2010
[41] M. Alizadeh, A. Kabbani, T. Edsall, B. Prabhakar, A. Vahdat and M.
Yasuda, "Less is More: Trading a little Bandwidth for Ultra-Low
Latency in the Data Center," in USENIX NSDI 2012
[42] "Priority ow control: Build reliable layer 2 infrastructure,"
http://www.cisco.com/en/US/prod/collateral/switches/ps9441/ps9670/
white_paper_c11-542809.pdf.
[43] M. Swierczynski, R. Teodorescu and P. Rodriguez, "Lifetime
investigations of a lithium iron phosphate (LFP) battery system
connected to a wind turbine for forecast improvement and output
power gradient reduction," in In BatCon'08., 2008.
[44] Windsun, "Lead-acid batteries: Lifetime vs Depth of discharge," 2009.
http://www.windsun.com/Batteries/Battery_FAQ.htm.
[45] US Department of Energy, "Grid-tie Energy Efficiency," 2010.
www1.eere.energy.gov/solar/review_meeting/pdfs/prm2010_apollo.pdf.
[46] C. S. Computing. www.climatesaverscomputing.org/resources/certification,.
[47] Energy Star, "Uninterruptible Power Supply Energy Efficiency
Values," www.energystar.gov/index.cfm?c=specs.uninterruptible_power_supplies.
[48] Emerson Network Power, "Rectifier Energy Efficiency,"
http://www.emersonnetworkpower.com/enUS/Brands/EnergySystems/Pages/ensys_eSureRectifiers.aspx.
http://www.google.com/corporate/datacenter/events/dc-summit-2009.html.
253