For more information .......................................................................................................................... 25
Call to action .................................................................................................................................... 25
Introduction
Server performance-per-watt continues to increase steadily. However, the number of watts-per-server also continues
to climb steadily. These increases, combined with the increasing number of servers and density in modern data
centers, make planning and managing facility power and cooling resources critically important. HP Power Capping
and HP Dynamic Power Capping are ProLiant power management tools that assist the data center administrator in
these critical tasks.
HP implements both Power Capping and HP Dynamic Power Capping in system hardware and firmware. Therefore,
they are not dependent on the operating system or applications. Power capping uses the power monitoring and
control mechanisms built into ProLiant servers. These mechanisms allow an administrator to limit, or cap, the power
consumption of a server or group of servers. Power capping lets you manage the data center parameters that server
power consumption directly influences, including data center cooling requirements and electrical provisioning.
Power capping also lets you control server power consumption in emergencies such as loss of primary AC power.
It is important to understand that power capping does not reduce the total energy consumption required for a server
to accomplish a given computational workload. Power capping simply limits the amount of power that a server can
use at any point in time. This lets you allocate data center power and cooling resources more efficiently. In general,
if a given power cap restricts the amount of power that a server would normally use to perform a task, that task will
take longer. Over time, the server will consume about the same total energy to execute the same computational
workload.
This paper outlines the use of power capping as part of a planning and provisioning strategy in the data center. It
also describes the relationship between power capping and other power management tools such as HP Power
Regulator.
Basics of server power control
The processor complex is one of the single greatest power consumers in ProLiant servers. In many common
configurations it is responsible for one-third of the power a server consumes (Figure 1). It also indirectly drives the
power consumption of other server components. A busy processor naturally increases the workload in both the
memory and peripherals. The heat generated by the increased workload causes the fans to work harder.
All HP power management technologies use this processor-driven model to control the processor’s power
consumption directly and to control overall server power consumption indirectly. The power management system
accomplishes this control using two separate mechanisms: changing the processor P-state and throttling the
processor clock.
3
Typical Server Power Usage
Processor
33%
Fans
7%
Memory
28%
PCI Slots
15%
Hard Drives
4%
Base Load
13%
Figure 1. Power use in a typical server
Processor P-states
Processor performance states, or P-states, provide a quick and effective mechanism for adjusting processor power
consumption and performance. Both Intel
®
and AMD® processors support using P-states to decrease processor
power consumption by lowering the processor’s core frequency and voltage. Tables 1 and 2 list some of the P-states
available with different processors.
Table 1. P-states of the Intel Xeon 5160 processor
P-state Core Frequency
P0 3.0 GHz
P1 2.66 GHz
P2 2.33 GHz
P3 2.0 GHz
Table 2. P-states of the AMD Opteron 2220 processor
P- state Core Frequency
P0 2.8 GHz
P1 2.6 GHz
P2 2.4 GHz
P3 2.2 GHz
P4 2.0 GHz
P5 1.8 GHz
P6 1.0 GHz
Clock throttling
Clock throttling is another method for lowering processor power consumption. Depending on the processor model,
the system BIOS can either reprogram the processor to run at a lower frequency or modulate the processor between
running periods and stopped periods. Both methods have the same net effect of lowering the processor’s overall
power consumption below the levels available using P-states. The chart in Figure 2 illustrates the relationship
between consumed power and overall performance when using P-states and clock throttling to control server power.
Using P-states clearly provides greater power reduction for a smaller loss in performance. However, using P-states
can lower power consumption only to a certain point. Reducing consumption below that point requires the use of
clock throttling.
4
DL360 G4 Power vs. Work Done
0
50
100
150
200
250
300
350
400
050010001500200025003000350040004500
Iterations
Watts
P0
P1
P2
Performance
Watts
P-state slope
Clock Throttling slope
1.00.250.7500.5
Relative Performance
Figure 2. Power versus performance characteristics for a typical Intel-based ProLiant server with three P-states
How power capping functions
Maintaining power consumption below the cap
With power capping, an administrator can set a maximum power consumption level for an individual server or for
a group of servers. The ProLiant power management system constantly monitors server power use. It adjusts P-states
and/or clock throttling to limit processor power use and control overall system power consumption. As needed, the
power capping control mechanism lowers the server’s power consumption in a controlled manner to keep it below
the cap―without affecting the server workload or environment.
Server power consumption depends on many factors and can vary significantly over a given period. Some factors,
such as the number of options installed in the server, have a predictable and static effect on server power
consumption. Other factors have a dynamic effect on power consumption, for example the temperature in the data
center; the activity of the CPU, memory, disk drives, and I/O subsystems; and even the mix of instructions executed.
As long as the power consumption remains below the power cap, the power capping control mechanism takes no
action and there is no affect on server performance.
Minimum and maximum power consumption for a server
The power management system in each server determines both the minimum and maximum power consumption for
the server. It determines these two values during the server’s power-on self test (POST) by executing tests measuring
server power consumption in idle mode and under a simulated maximum load. Because they are determined
empirically, the two power values implicitly take into account the server configuration and its current physical
environment. The Insight Control and iLO interfaces display both values to provide key information that
administrators can use to set effective power caps.
The power management system in the server supplies one additional metric: maximum available power for the
server. For ProLiant ML and ProLiant DL servers, this value is the maximum power that the server power supply can
5
produce. However, the enclosure’s power supply array powers HP BladeSystem servers. For a blade server, the
maximum available power is the amount of power that the enclosure’s Onboard Administrator reserved for that
server blade. Both iLO and Insight Control report this value: iLO reports it as Power supply maximum power for ML
and DL servers and as Initial power-on request value for BladeSystem servers.
Minimum and maximum power consumption values for a server can change slightly while the server is running.
During normal operations, iLO and the power management system continue to check both the 10-second average
and the peak power readings for the server. iLO will raise the maximum power consumption level if it measures a
peak value above the established maximum. iLO will lower the minimum power consumption if it reads an average
power value that is below the present minimum.
Differences between HP Dynamic Power Capping and HP Power
Capping
Both HP Dynamic Power Capping and HP Power Capping maintain a server’s power consumption at or below the
cap value set by an administrator. HP Dynamic Power Capping monitors power consumption and maintains a
server’s power cap much faster than HP Power Capping. Table 3 provides a quick architectural and operational
comparison of HP Dynamic Power Capping and HP Power Capping. To avoid confusion between the two, we will
refer to HP Power Capping as basic Power Capping throughout the remainder of this paper.
Table 3. Characteristics of Dynamic Power Capping and basic Power Capping
Dynamic Power
Capping
Power capping
executed from
Control of processor
power
Power monitoring cycle More than 5 times per
Time to bring server
power consumption
back under its cap
Intended application Managing power and
Power management
microcontroller
Direct hardware
connection to
processor to control Pstate/clock throttling at
the processor
hardware level
second
Less than 0.5 seconds 10 – 30 seconds
cooling provisioning
Basic Power Capping
iLO and system ROM
BIOS
Firmware control of Pstate/clock throttling
through processor
registers
Once every 5 seconds
Managing cooling
provisioning
Power provisioning and Dynamic Power Capping
Basic Power Capping does an excellent job of maintaining average server power utilization at or below a cap
value. Administrators can use it to help manage data center cooling requirements: Limiting server power
consumption fast enough can prevent excessive heat generation. However, as the information in Table 3 illustrates,
basic Power Capping cannot respond quickly enough to limit sudden increases in server power consumption that
could trip an electrical circuit breaker.
Dynamic Power Capping operates more than 25 times faster than basic Power Capping. Dynamic Power Capping
can bring a server experiencing a sudden workload increase back under its power cap in less than one-half second.
This is fast enough to prevent any surge in power demand that could cause a typical data center circuit breaker to
6
trip. HP has tested Dynamic Power Capping to ensure that it can prevent tripping circuit breakers that have a
specified trip time of 3 seconds or longer at 50 degrees C and 150 percent overload.
Dynamic Power Capping can keep server power consumption below the power cap in real time. Therefore,
administrators can use it as an effective tool in planning and managing both electrical provisioning and cooling
requirements in the data center. An administrator can electrically provision a PDU or a rack to something less than
the full faceplate power rating of all the servers supported because Dynamic Power Capping will prevent a sudden
power demand from exceeding the power cap and tripping a circuit breaker.
Support for Power Capping in ProLiant servers
ProLiant servers with power measurement circuitry support basic Power Capping:
Basic Power Capping requires the following system firmware:
• iLO 2 version 1.30 or later
• System BIOS 2007.05.01 or later
Support for Dynamic Power Capping requires a certain level of ProLiant hardware, as well upgrades to the
following system firmware:
• System BIOS 2008.11.01 or later
• iLO 2 version 1.70 or later
• Onboard Administrator firmware version 2.32 or later (for HP BladeSystem enclosures)
At introduction, support for Dynamic Power Capping is available on a limited set of ProLiant servers and a larger set
of ProLiant c-Class server blades. Many ProLiant G5 servers can support Dynamic Power Capping if they have fully
qualified BIOS and iLO firmware. Please consult the most recent
support matrix.
Group power capping for servers through Insight Control
One of the most powerful uses of power capping is monitoring and controlling the power use of an entire group of
servers. This capability is available through Insight Control. Administrators can apply a group power cap to any
group of servers that they can select within Insight Control, including Insight collections.
Insight Control displays the aggregated minimum and maximum power consumption for an entire group of servers
and their aggregated power supply rating. Each of these numbers is simply the sum of the respective values for the
individual servers in the group. Using the Insight Control interface, an administrator can apply to the server group a
power cap that is between the minimum power and the power supply rating of the entire group.
Insight Control assigns an individual power cap to each server in the group. It is a proportional allocation of the
group power cap. The total of the individual power caps equals the group cap. The individual power caps for the
servers continue until an administrator changes them through the iLO or Insight Control interface.
Figure 3 shows a group consisting of four servers. The left side of the figure shows the measured minimum and
maximum power consumption for each server. The right side of the figure shows that the aggregated maximum
power consumption for this group is 1375 watts. The aggregated minimum allowable power consumption is 725
watts. In this example, an administrator has applied a power cap of 1115 watts to the group. This group cap limits
the group power consumption to 60 percent of the wattage between the aggregated minimum and maximum. To
implement this group power cap, Insight Control applies to each server a power cap that is 60 percent of the
wattage between that server’s minimum and maximum power consumption. This results in individual power caps of
320, 170, 305, and 320 watts respectively.
7
Group Power Capping
Individual Servers
400 watts (Maximum)
200 watts (Minimum)
Servers as a Group
320 watts (Apportioned Cap)
200
125
170
375
200
400
200
305
320
1375 watts
(Group maximum)
725 watts
(Group minimum)
1115 watts
(User set group power cap)
1000 watts (Power supply maximum)
1000
1000
500
3500 watts
(Group power supply
maximum)
Figure 3. Apportioning a group power cap to individual servers in the group
Group power capping apportionment works exactly the same way on ProLiant ML and DL servers with either
Dynamic Power Capping or basic Power Capping. For server blades, there is a more advanced feature called
Enclosure Dynamic Power Capping.
Enclosure Dynamic Power Capping
Enclosure Dynamic Power Capping is a special implementation of Dynamic Power Capping for HP BladeSystem
enclosures. In one sense, it is a higher level of power management functionality, since an administrator sets and
maintains a power cap at the enclosure level rather than directly at the server or blade level. In another sense, it is a
more powerful implementation of group power capping for an enclosure: Setting a power cap for the enclosure
indirectly creates power caps for the server blades within it. The Onboard Administrator (OA) then actively
manages these power caps and reallocates power as workloads change over time.
Elements of an enclosure power cap
With Enclosure Dynamic Power Capping, an administrator sets a power cap for an entire BladeSystem enclosure,
not simply for the server blades in the enclosure. Total power consumption for an enclosure is the sum of the power
used by all of these components:
• Server blades
• I/O peripherals for the enclosure (interconnects, etc.)
• Cooling fans for the enclosure
• Onboard Administrator(s)
8
Loading...
+ 17 hidden pages
You need points to download manuals.
1 point = 1 manual.
You can buy points or you can get point for every manual you upload.