Specifications

April 2012 v1 AMD Opteron™ 6200 Linux Tuning Guide
21
P3 := Freq: 2300 MHz
P4 := Freq: 1400 MHz
In this example, P0 is the highest non-boosted P state in normal operation. Pb1 is the first boosted frequency.
Having HPC P-state mode enabled reconfigures P1 through P3 to be set to the same frequency as P0. This
prevents slower performance as long as the system TDP design allows continuous operation at these
frequencies.
4.3.1 Check OS Power Management Default
Linux in general (upstream) uses ‘cpufreq’ subsystem for power management. The power management modes
are ondemand, performance, conservative, powersave, and userspace. In the performance mode, the kernel
keeps the cores in P-state P0. By default on servers, ondemand policy is used, which scales up frequency to
the highest software P-state possible when there are tasks to run. When there is nothing in the run queue for a
core to do, in ondemand, the kernel will move to a lower performance P-state (e.g., P1 – P6).
An Opteron™ 6200 or 4200 CPU has two levels of boost: PB0 and PB1. Depending on the C-states of the sibling
cores and the TDP utilization, the processor will transparently boost to PB0 (highest) when possible or PB1.
This happens transparently to the kernel. All of the power management modes except powersave mode allow
CPUs to boost transparently.
The following are Linux Distribution/Version defaults:
• Red Hat RHEL 6.1 and 6.2 by default uses ondemand mode (governor). 6.1/6.2 has cpuidle support that
can also put cores into C6 states. Hence, highest boost levels can be achieved transparently by default.
• Red Hat RHEL 5.7/5.8 by default uses ondemand mode, but RHEL 5.7/5.8 do not support deep C-states;
therefore, they will not enable the processor to do “big” boost (Pb0).
• Red Hat RHEL 5.x line does not support deep C-states. Hence, although boost happens transparently,
the processor cannot boost to the highest boost P-state (Pb0).
• SUSE Linux Enterprise Server (SLES) 11 SP2 is similar to mainline. ondemand is the default mode. The OS
supports cpuidle and highest boost states can be achieved transparently.
Note: The frequency in /proc/cpuinfo does not show any effect of boost. The turbostat utility, if in your
distribution, can be used to actually read the boosted frequency of CPUs.
A useful public tool package for monitoring CPU boost is the cpufreq-utils package that can be used to display
useful information about the current state of the CPUs with the cpufreq-info tool and allows you to watch the
current speed of the CPUs and identify if and when they boost. This handy package is located at http://rpmfind.
net/linux/rpm2html/search.php?query=cpufrequtils.
4.3.2 Trade-offs and Consequences
No discussion of changing processor clocking modes and power management features would be complete
without a discussion of the consequences and trade-offs. The default settings should be suitable for many
applications and certainly best for machines being deployed for general purpose computing with a wide variety
of applications. There are some special circumstances where it provides a benefit to make these changes, for
example, in some benchmarking scenarios. Keep the following points in mind when deciding when to use APM,
HPC P-state mode, and other modifications to default power management settings.