opr0R3UT.pdf

Performance loss during normal operation in a Dell Latitude E6500

laptop due to processor and bus clock throttling

Randall Cotton - recotton@earthlink.net

Revision 1.1 - July 31, 2009

Introduction:

This article describes how processor performance can decline by more than 95% in a Dell Latitude E6500

running Windows XP Professional (Service Pack 3) by running routine software at normal operating

temperatures due to activation of the following clock throttling mechanisms present in Intel Core 2 Duo

processors:

Performance State Transitions

II.

Software-controlled Clock Modulation (also called On-Demand Clock Modulation)

These clock throttling features are designed to be invoked to conserve energy when a processor is idle or to

head off an overheating condition. However, in the tested Dell Latitude E6500, they are engaged at normal

operating temperatures even when demand for processing power is high. This can result in severe loss of

performance under routine use at normal ambient room temperatures. This loss of performance can remain until

the system is rebooted (and, since it is temperature-dependent, the performance loss can conceivably remain

across reboots).

To illustrate and reliably reproduce this phenomenon, various methods are presented that raise internal system

operating temperatures enough to cause the throttling mechanisms to be invoked while monitoring key

parameters such as processor temperatures and operating frequencies. In some cases, freely available third-party

processor-stressing software is used, but in other cases, even simple Windows-native processes like Calculator

and a 4-line looped batch file that outputs scrolling lines of text are enough to trigger substantial loss of

processor performance. In the worst case, all of the following takes place:

I. Reduction of internal CPU clock frequency from 2261 MHz to 798 MHz

II. Reduction of FSB (Front Side Bus) clock frequency from 266MHz to 133MHz

III. In addition to the above, overall reduction in the CPU effective clock rate by a factor of

eight due to Software-controlled Clock Modulation.

note: the term CPU stands for Central Processing Unit, referring to a PC’s main processor chip

note: the term Front Side Bus is the technical name for the main data path between the CPU and the rest of the system

Even setting aside the negative performance effect of FSB downshifting in II above, the effective processing

power is reduced to 1/8 of 798 Mhz = 100 MHz. This is a reduction to less than 5% of full capacity (100/2261 =

4.42%).

It is noteworthy that all performance loss takes place with neither notice nor explanation to the user.

Background: Processor Clock-throttling Mechanisms

The Dell Latitude E6500 features an Intel Core 2 Duo processor, which offers several mechanisms that can be

used (typically by operating systems such as Microsoft Windows) for both thermal management and power

conservation. Background information follows on the two mechanisms already mentioned.

A. Performance State Transitions

Many of Intel’s CPUs are designed to run at a variety of performance states. A performance state, or P-state,

is generally a combination of processor frequency (such as 2.26GHz) and a voltage level (such as 1.2V) 1 .

Lower frequencies provide less processing performance, but require lower voltages supplied to the

processor and therefore consume less electrical power. Typically, Intel CPU’s in PCs use 4 performance

states. There can be a wide range of processor performance among these states. In the Dell Latitude E6500

tested for this article under Windows XP, the following states are used by default:

Performance

State

FSB Clock

Frequency

CPU Clock Frequency

Est. CPU

Voltage

266 MHz

2261 MHz (8.5 x FSB frequency)

with Intel Dynamic Acceleration

1.2V

266 MHz

2261 MHz (8.5 x FSB frequency)

without Intel Dynamic Acceleration

1.1375V

266 MHz

1596 MHz (6 x FSB frequency)

1.0V

133 MHz

798 MHz (6 x FSB frequency)

0.925V

Note: The terminology P0, P1, etc., referring to P-state 0, P-state 1, etc., is from the ACPI standard – Advanced Configuration and Power

Interface. That standard and its relevance are described later in this article.

Note: Intel Dynamic Acceleration refers to a feature whereby one of the two processing cores is clocked a little higher than the other when that

other processor is not under much load – in this case, the boost to one of the cores is to 2394 MHz (9 x 266 MHz FSB frequency) 2 .

Note: CPU clock frequencies are derived from the Front Side Bus frequency. Traditionally, the CPU frequency has been an integer multiple of

the FSB frequency, though some Core 2 Duo processors support half-step multipliers such as 8.5 in this case.

Note: Performance State P3 is achieved not by slowing just the CPU clock, but rather by slowing the Front Side Bus (which necessarily slows the

CPU clock by the same factor). This provides for even greater power savings (and even lower performance) than by just slowing the CPU clock

alone. Intel calls this technique Dynamic FSB Frequency Switching 3 .

Note: These aren’t the only P-states that are technically possible with this processor. They just constitute the states being used by the Dell

Latitude E6500 under test.

Note: P1 is sometimes referred to as HFM (Highest Frequency Mode) by Intel documentation. Likewise, P2 is sometimes referred to as LFM

(Lowest Frequency Mode) and P3 is referred to as SLFM (Super Low Frequency Mode) with the “Super” referring to Dynamic FSB Frequency

Switching.

Sometimes, transitions to lower-power P-states are made in order to decrease power consumption when

processing demand is low. As soon as processing demand increases enough to warrant a higher-performance

P-state, the necessary upclocking transition is made nearly instantaneously. Intel calls this power-saving

technique Speedstep Technology 4 . But lower-performance P-states can also be used on a sustained basis for

the purpose of lowering operating temperatures even under high processing demand.

B. Software-controlled Clock Modulation

This is a mechanism to reduce the effective clock rate of the CPU by slicing up time into tiny equal intervals

and then only running the CPU clock (and hence, the processor) during a fraction of each time slice. The

fraction can be one of eight settings 1/8 through 8/8 = 1. That is, the clock runs either 8/8 of the time (all the

time), 7/8 of the time, etc. down to 1/8 of the time, depending on what setting is chosen by software. So in

the extreme case, if the setting is 1/8, the processor is effectively only running at 1/8 th or 12.5% of the speed

it was before Software-controlled Clock Modulation was engaged 5 .

Normal operating temperatures

In general, CPUs and GPUs (Graphics Processing Units) generate the highest temperatures in a PC. However,

these chips are both designed to run without problem even at fairly high temperatures. Normal operating

temperatures range from around 35°C when idle to the 80s or even 90s Celsius under heavy load. All processors

of any type are designed to operate normally under a specified range of temperatures. For example, the Intel

Core 2 Duo P8400 processor in the system under test is rated for normal operation up to 105° Celsius 6 . Though

specifications for the NVIDIA Quadro NVS 160M processor are not available online, GPU chip technology is

similar to CPU technology and GPUs are designed to operate in similar temperature ranges (sometimes at even

higher temperatures than a system’s CPU).

While high temperatures beyond the rated operating range can result in processing errors and extreme

temperatures can result in physical damage, modern-day CPUs and GPUs are commonly designed with built-in

temperature sensors, automatic clock throttling capabilities and emergency shutdown mechanisms. For example,

the Intel Core 2 Duo P8400 processor under test is capable of automatically engaging both clock-throttling

mechanisms described above in the event of overheating (once the internally measured temperature exceeds the

rated maximum of 105° Celsius). The processor also has an emergency shutdown feature, which halts all

operation (and requires power to be cut) once the internally measured temperature reaches around 125° Celsius 7 .

How sustained clock throttling occurs at normal operating temperatures

In the system under test, once internal temperatures reach a given point (still well within normal operating

range), the system is throttled first by transitioning to lower-performance P-states. As long as the temperature

remains elevated enough, lower and lower performance P-states are activated in succession at 30 second

intervals. Assuming the system starts in state P0, it first transitions to P1, then 30 seconds later to P2 and 30

seconds after that to P3. It’s possible for the system to throttle down only to P1 or P2 and then hold there,

foregoing any further throttling so long as the internal temperature declines enough. However, if the lowest-

performing P-state is reached and the system still determines the temperature is too high, then it will begin

Software-controlled Clock Throttling, successively cutting the effective CPU clock (and thus, the CPU’s

performance) in the available 1/8 increments, as previously described. This also happens at intervals of 30

seconds and continues as long as the system determines the temperature is too high. Once the system progresses

through all available Software-controlled Clock Throttling settings, the system effects no further throttling. In

total, there are 10 incremental steps, progressively applied at 30 second intervals, that are used to throttle the

processing power of the system:

Throttling action

Effective processing power (frequency)

1. Transition from state P0 to P1

2261 MHz

2. Transition from state P1 to P2

1596 MHz

3. Transition from state P2 to P3

798 MHz (with FSB frequency cut in half)

4. Clock Throttling at 7/8 capacity of state P3

700 MHz (with FSB frequency still reduced by half)

5. Clock Throttling at 6/8 capacity of state P3

600 MHz (with FSB frequency still reduced by half)

6. Clock Throttling at 5/8 capacity of state P3

500 MHz (with FSB frequency still reduced by half)

7. Clock Throttling at 4/8 capacity of state P3

400 MHz (with FSB frequency still reduced by half)

8. Clock Throttling at 3/8 capacity of state P3

300 MHz (with FSB frequency still reduced by half)

9. Clock Throttling at 2/8 capacity of state P3

200 MHz (with FSB frequency still reduced by half)

10. Clock Throttling at 1/8 capacity of state P3

100 MHz (with FSB frequency still reduced by half)

While this aggressive degradation of the system’s capabilities might be warranted if internal temperatures were

so high that they might cause damage or malfunction, unfortunately in the system under test, these steps are

activated at normal internal operating temperatures, in some cases even well below 70° Celsius. It’s particularly

problematic that the user receives no notification regarding these drastic changes which severely impact

performance. As a further complication, even though the throttling action brings down temperatures

dramatically, performance levels are sometimes not restored for a long time, requiring a system reboot to clear

the throttled condition.

The possible role of ACPI

The Advanced Configuration and Power Interface is an interface specification standard promoted by Intel,

Microsoft, Phoenix Technologies and other corporations which allows operating systems to communicate in a

standardized way with physical hardware devices in a computer system. It provides a robust and uniform way

for operating systems such as Windows to detect, monitor and configure a wide spectrum of devices including

very low-level hardware such as power supplies, batteries, cooling fans and processor chips. In a system that is

fully compliant with ACPI, configuration of such devices is performed exclusively through the operating

system via the ACPI interface 8 .

The throttling behavior observed would be consistent with an ACPI “passive cooling” policy. Passive cooling is

defined in ACPI as a method whereby the “OS reduces the power consumption of devices at the cost of system

performance to reduce the temperature of the machine.” In the standard, a temperature threshold _PSV is

defined beyond which passive cooling is engaged. ACPI suggests providing hysteresis by changing _PSV to a

lower temperature once the initial threshold is reached and using that new value as a target temperature while

passive cooling is engaged. An equation is even suggested for how to adjust the system’s performance to

achieve the lower temperature (using vendor-provided constants _TC1 and _TC2). The standard provides for

periodic polling of the system temperature for the purposes of passive cooling using a sampling period _TSP 9 .

While it’s not practically possible to independently confirm conclusively that this system is fully compliant with

ACPI, or even that this system’s performance loss is being effected through ACPI (particularly since passive

cooling is not an absolute requirement for ACPI compliance), there are a number of facts that strongly suggest

both may be the case.

2. Microsoft (the operating system vendor), Intel (the CPU vendor) and Phoenix Technologies (the BIOS

vendor) are all promoters of the ACPI standard. Phoenix Technologies can be confirmed as the BIOS

vendor for the Dell Latitude E6500 using a BIOS scan at address 0xF0000 (see screen capture image

below)

1. The strings “_TC1”, “_TC2”, “_PSV” and “_TSP”, which are particular to the “passive cooling” option

of the ACPI standard, all appear in the file ACPI.SYS file under

C:\WINDOWS\SYSTEM32\DRIVERS.

3. The Phoenix BIOS in the Dell Latitude E6500 contains the expected ACPI tables and code required by

the standard (Root System Description Pointer at address 0xFB9C0, Extended System Description Table

at address 0xDF451E00, Fixed ACPI Description Table at address 0xDF451C9C, Differentiated System

Descriptor Table at address 0xDF452400, Secondary System Description Table at address 0xDF45032D,

etc.). A screen capture image of the beginning of the Differentiated System Descriptor Table appears

below:

Illustration: maximum performance loss due to clock throttling at normal operating temperatures

Following is an illustrated example of exactly how the Dell Latitude E6500 is progressively throttled to the

point that it is operating at less than 5% of processor capacity. Screen captures are used to chronicle the process.

First, here is a more complete description of the test system:

Dell Latitude E6500

Intel Core 2 Duo P8400 processor, 2.26Ghz

4GB RAM

NVIDIA Quadro NVS 160M graphics

Dell BIOS version A13

E/Port Plus Advanced Port Replicator (Docking Station)

dual Dell E248WFP monitors (only one monitor used in some tests)

Windows XP Service Pack 3

Configuration note: A key adjustment for these tests was to switch to the “Always On” power scheme in the “Power Options”

Control Panel window. Choosing the “Always On” power scheme disables Intel’s Speedstep feature which temporarily

downclocks the processor (by transitioning among the various P-states) when it is idle. To avoid confusion about what caused

a P-state change, it was important to disable Speedstep.

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika: