pllopis / turbofreq Goto Github PK

What is this?
-------------

This kernel module provides a cpufreq p-state selection driver for Linux. 

Why?
----

Modern processors take advantage of the thermal headroom to improve performance
by opportunistically adjusting its voltage and frequency based on thermal and
energy constraints (e.g. Intel's Turbo Boost technology). This technology
introduces performance variation over time that is often undesirable under
certain conditions, especially for certain HPC codes. In addition, this dynamic
window is becoming larger and more relevant with each generation, and system
software is not thermal aware, causing thermal imbalances within a single node,
and even within a single socket. This make an homogeneous system behave like an
heterogeneous system. By being aware of this heterogeneous behavior, system
software can make a more efficient use of the hardware's resources, and expose
mechanisms to user-space to allow runtimes or user applications choose an
appropriate policy.

What does it do?
----------------

This driver implements a few policies that provide different trade-offs between
average performance and performance variability, and provides more fine-grained
control to user-space to select the appropriate behavior.

How does it work?
-----------------

After the module is inserted, it populates sysfs entries under
/sys/devices/system/cpu/turbofreq/

    pstate_available_policies: list available policies that can be fed to
    pstate_policy.  
        "boost" defaults to the maximum performance and should be
           similar to the default behavior of the intel_pstate driver.  
        "cpu" provides a stable frequency on a per-cpu basis.
        "smp" provides a stable frequency smp-wide for all cpus in the
           system.
        "noboost" disables turbo boost altogether.

    pstate_policy: used to select a target policy.

    task_boost: the task that writes "1\n" to this file will get turbo boost
         enabled on the cpu it is running on. Write "0\n" to disable and return
         to the previous policy.

    reset: resets all values to the default. This will also set internal
         frequencies that converged under "smp" or "cpu" back to the max.

    global_pstate: to manually set a stable frequency. This is meant for
         implementing cluster-wide frequency synchronization with the help of
         AMQP, but the implementation is not finished. 

There is also an interface added to each sysfs cpu to check for and
enable/disable turbo boost on a per-cpu basis under
/sys/devices/system/cpu/cpu%d/boost (depends on your topology).

How do I use this?
------------------

To change the default cpufreq driver, the easiest way is to just disable
intel_pstate and acpi-cpufreq.  To disable intel_pstate, just pass
intel_pstate=disable to the kernel's boot parameters.  In our distribution
(CentOS 6), I couldn't find any reference to acpi-cpufreq being loaded
explicitly, so I just moved the file.  After doing this, no cpufreq driver will
be registered until turbofreq.ko is inserted.