21. Hardware Feedback Interface For Hetero Core Scheduling On AMD Platform

Copyright:

2025 Advanced Micro Devices, Inc. All Rights Reserved.

Author:

Perry Yuan <perry.yuan@amd.com>

Author:

Mario Limonciello <mario.limonciello@amd.com>

21.1. Overview

AMD Heterogeneous Core implementations are comprised of more than one architectural class and CPUs are comprised of cores of various efficiency and power capabilities: performance-oriented classic cores and power-efficient dense cores. As such, power management strategies must be designed to accommodate the complexities introduced by incorporating different core types. Heterogeneous systems can also extend to more than two architectural classes as well. The purpose of the scheduling feedback mechanism is to provide information to the operating system scheduler in real time such that the scheduler can direct threads to the optimal core.

The goal of AMD’s heterogeneous architecture is to attain power benefit by sending background threads to the dense cores while sending high priority threads to the classic cores. From a performance perspective, sending background threads to dense cores can free up power headroom and allow the classic cores to optimally service demanding threads. Furthermore, the area optimized nature of the dense cores allows for an increasing number of physical cores. This improved core density will have positive multithreaded performance impact.

21.2. AMD Heterogeneous Core Driver

The amd_hfi driver delivers the operating system a performance and energy efficiency capability data for each CPU in the system. The scheduler can use the ranking data from the HFI driver to make task placement decisions.

21.3. Thread Classification and Ranking Table Interaction

The thread classification is used to select into a ranking table that describes an efficiency and performance ranking for each classification.

Threads are classified during runtime into enumerated classes. The classes represent thread performance/power characteristics that may benefit from special scheduling behaviors. The below table depicts an example of thread classification and a preference where a given thread should be scheduled based on its thread class. The real time thread classification is consumed by the operating system and is used to inform the scheduler of where the thread should be placed.

21.3.1. Thread Classification Example Table

class ID

Classification

Preferred scheduling behavior

Preemption priority

Counter

0

Default

Performant

Highest

1

Non-scalable

Efficient

Lowest

PMCx1A1

2

I/O bound

Efficient

Lowest

PMCx044

Thread classification is performed by the hardware each time that the thread is switched out. Threads that don’t meet any hardware specified criteria are classified as “default”.

21.4. AMD Hardware Feedback Interface

The Hardware Feedback Interface provides to the operating system information about the performance and energy efficiency of each CPU in the system. Each capability is given as a unit-less quantity in the range [0-255]. A higher performance value indicates higher performance capability, and a higher efficiency value indicates more efficiency. Energy efficiency and performance are reported in separate capabilities in the shared memory based ranking table.

These capabilities may change at runtime as a result of changes in the operating conditions of the system or the action of external factors. Power Management firmware is responsible for detecting events that require a reordering of the performance and efficiency ranking. Table updates happen relatively infrequently and occur on the time scale of seconds or more.

The following events trigger a table update:
  • Thermal Stress Events

  • Silent Compute

  • Extreme Low Battery Scenarios

The kernel or a userspace policy daemon can use these capabilities to modify task placement decisions. For instance, if either the performance or energy capabilities of a given logical processor becomes zero, it is an indication that the hardware recommends to the operating system to not schedule any tasks on that processor for performance or energy efficiency reasons, respectively.

21.5. Implementation details for Linux

The implementation of threads scheduling consists of the following steps:

  1. A thread is spawned and scheduled to the ideal core using the default heterogeneous scheduling policy.

  2. The processor profiles thread execution and assigns an enumerated classification ID. This classification is communicated to the OS via logical processor scope MSR.

  3. During the thread context switch out the operating system consumes the workload (WL) classification which resides in a logical processor scope MSR.

  4. The OS triggers the hardware to clear its history by writing to an MSR, after consuming the WL classification and before switching in the new thread.

  5. If due to the classification, ranking table, and processor availability, the thread is not on its ideal processor, the OS will then consider scheduling the thread on its ideal processor (if available).

21.6. Ranking Table

The ranking table is a shared memory region that is used to communicate the performance and energy efficiency capabilities of each CPU in the system.

The ranking table design includes rankings for each APIC ID in the system and rankings both for performance and efficiency for each workload classification.

struct amd_shmem_info

Shared memory table for AMD HFI

Definition:

struct amd_shmem_info {
    struct acpi_pcct_ext_pcc_shared_memory header;
    u32 version_number          :8, n_logical_processors    :8, n_capabilities          :8, table_update_context    :8;
    u32 n_bitmaps               :8, reserved                :24;
    u32 table_data[];
};

Members

header

The PCCT table header including signature, length flags and command.

version_number

Version number of the table

n_logical_processors

Number of logical processors

n_capabilities

Number of ranking dimensions (performance, efficiency, etc)

table_update_context

Command being sent over the subspace

n_bitmaps

Number of 32-bit bitmaps to enumerate all the APIC IDs This is based on the maximum APIC ID enumerated in the system

reserved

24 bit spare

table_data

Bit Map(s) of enabled logical processors Followed by the ranking data for each logical processor

21.7. Ranking Table update

The power management firmware issues an platform interrupt after updating the ranking table and is ready for the operating system to consume it. CPUs receive such interrupt and read new ranking table from shared memory which PCCT table has provided, then amd_hfi driver parses the new table to provide new consume data for scheduling decisions.