lkml.org 
[lkml]   [2022]   [Sep]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[RFC PATCH 00/23] sched: Introduce classes of tasks for load balance
Date
+++ Introduction

On hybrid processors, the microarchitectural properties of the different
types of CPUs cause them to have different instruction-per-cycle (IPC)
capabilities. IPC can be higher on some CPUs for advanced instructions
Figure 1 illustrates this concept. It plots hypothetical workloads
grouped by classes of instructions vs the IPC ratio between high and low
performance CPUs.

IPC ratio
^
| Class0 . Class1 . ClassN-1 . ClassN
| . . . +
| . . . +
| . . . +
| . . + + + + + + + +
| . . .
| . . .
| . + + + + + + + + + + + .
| + + + + + + + . .
| + . . .
| + . . .
| + . . .
| + . . .
1 |-------------------------------------------//---------------------------->
| wokloads of interest

Figure 1. Hypothetical workloads sorted by IPC ratio


The load balancer can discover the use of advanced instructions and prefer
CPUs with higher IPC for tasks running those instructions.

Hardware is free to partition its instruction set into an arbitrary number
of classes. It must provide a mechanism identify the class of the
currently running task and inform the kernel about the performance of each
class of task on each type of CPU.

This patchset introduces the concept of classes of tasks, proposes the
interfaces that hardware needs to implement and proposes changes to the
load balancer to leverage this extra information in combination with
asymmetric packing.

This patchset includes a full implementation for Intel hybrid processors
using Intel Thread Director technology [1].


+++ Structure of this series

Patches 1-6 introduce the concept of classes of tasks. They also introduce
interfaces that architectures implement to update the class of a task and
to inform the scheduler about the class-specific performance scores.

Patches 7-9 use the class of the current task of each runqueue to break
ties between two otherwise identical group_asym_packing scheduling groups.

Patches 10-16 implement support for classes of tasks on Intel hybrid
processors using Intel Thread Director technology.

Patches 17-19 introduce extra helper members to task_struct to deal with
transient classification of tasks and arch-specific implementation
vagaries.

Patches 20-22 are specific to Intel Thread Director. They reset the
classification hardware when switching to a new task.


+++ Classes of tasks

The first step to leverage the asymmetries in IPC ratios is to assign a
class label to each individual task. Hardware can monitor the instruction
stream and classify a task as it runs. At user tick, the kernel reads the
resulting classification and assigns it to the currently running task. It
stores the class in the proposed task_struct::class.


+++ Balancing load using classes of tasks. Theory of operation

Intel hybrid processors rely on asymmetric packing to place tasks on
higher performance CPUs first. The class of the current task on each
runqueue can be used to break ties between two otherwise identical
scheduling groups.

Consider these scenarios (for simplicity, assume that task-class
performance score is such that

score(Cl0) < score(Cl1) < ... < score(Cl(n-1)) < score(Cln)). (Eq I)

Boxes depict scheduling groups being considered during load balance.
Labels inside the box depict the class of rq->curr, or the CPU being
idle.

asym
packing
priorities 50 50 30 30
_____________
| i . i |
| d . d |
| l . l | _______ _______
| e . e | | Cl0 | | Cl1 |
|___________| |_____| |_____|

^
dst_cpu sgA sgB
^
busiest

Figure 2. Scenario A

In Figure 2, dst_cpu is a group of SMT siblings, has become idle, has
higher priority, and is identifying the busiest group. sgA and sgB are of
group_asym_packing type, have the same priority, have a single CPU, and
have the same number of running tasks. By checking the class of the task
currently running on both scheduling groups, it selects sgB as the busiest
because it has a class of task higher performance score if placed on
dst_cpu.

asym
packing
priorities 50 50 50 50 30
_____________ _____________
| . | | . |
| . | | . | idle
| cl0 . cl1 | | cl0 . cl2 | _______
| . | | . | | |
|___________| |___________| |_____|

^
sgA sgB dst_cpu
^
busiest group

^
busiest queue

Figure 3. Scenario B

In Figure 3, dst_cpu has become idle, has lower priority and is identifying
a busiest group. sgA and sgB are groups of SMT siblings. Both siblings are
busy and, therefore, classified as group_asym_packing [2], have the same
priority and the same number of running tasks. The load balancer computes
the class-specific performance score (scaled by the number of busy
siblings) by observing the currently running task on each runqueue.

As per Eq. I, cl0+cl2 has a higher throughput than cl0+cl1. So, it selects
sgB as the busiest group. If cl2 is left to run with the whole big core to
itself, it would deliver higher throughput than cl0. Hence, the runqueue of
cl0 is selected as the busiest.


+++ Dependencies
These patches assume that both SMT siblings of a core have the same
priority, as proposed in [3]. Also, they rely on the existing support for
the Hardware Feedback Interface [4].


I look forward to your review and thank you in advance!

These patches have been Acked-by: Len Brown <len.brown@intel.com>

Thanks and BR,
Ricardo

[1]. Intel Software Developer Manual, Volume 3, Section 14.6
https://intel.com/sdm
[2]. https://lkml.kernel.org/r/20210911011819.12184-7-ricardo.neri-calderon@linux.intel.com
[3]. https://lore.kernel.org/lkml/20220825225529.26465-1-ricardo.neri-calderon@linux.intel.com/
[4]. https://lore.kernel.org/linux-pm/20220127193454.12814-1-ricardo.neri-calderon@linux.intel.com/

Ricardo Neri (23):
sched/task_struct: Introduce classes of tasks
sched: Add interfaces for classes of tasks
sched/core: Initialize the class of a new task
sched/core: Add user_tick as argument to scheduler_tick()
sched/core: Move is_core_idle() out of fair.c
sched/core: Update the classification of the current task
sched/fair: Collect load-balancing stats for task classes
sched/fair: Compute task-class performance scores for load balancing
sched/fair: Use task-class performance score to pick the busiest group
sched/fair: Use classes of tasks when selecting a busiest runqueue
thermal: intel: hfi: Introduce Hardware Feedback Interface classes
thermal: intel: hfi: Convert table_lock to use flags-handling variants
x86/cpufeatures: Add the Intel Thread Director feature definitions
thermal: intel: hfi: Update the class of the current task
thermal: intel: hfi: Report per-cpu class-specific performance scores
thermal: intel: hfi: Define a default classification for unclassified
tasks
thermal: intel: hfi: Enable the Intel Thread Director
sched/task_struct: Add helpers for task classification
sched/core: Initialize helpers of task classification
thermal: intel: hfi: Implement model-specific checks for task
classification
x86/cpufeatures: Add feature bit for HRESET
x86/hreset: Configure history reset
x86/process: Reset hardware history in context switch

arch/x86/include/asm/cpufeatures.h | 2 +
arch/x86/include/asm/disabled-features.h | 8 +-
arch/x86/include/asm/hreset.h | 30 ++++
arch/x86/include/asm/msr-index.h | 4 +
arch/x86/include/asm/topology.h | 10 ++
arch/x86/kernel/cpu/common.c | 35 +++-
arch/x86/kernel/cpu/cpuid-deps.c | 1 +
arch/x86/kernel/cpu/scattered.c | 1 +
arch/x86/kernel/process_32.c | 3 +
arch/x86/kernel/process_64.c | 3 +
drivers/thermal/intel/Kconfig | 12 ++
drivers/thermal/intel/intel_hfi.c | 215 +++++++++++++++++++++--
include/linux/sched.h | 19 +-
init/Kconfig | 9 +
kernel/sched/core.c | 10 +-
kernel/sched/fair.c | 214 ++++++++++++++++++++--
kernel/sched/sched.h | 81 +++++++++
kernel/sched/topology.c | 8 +
kernel/time/timer.c | 2 +-
19 files changed, 635 insertions(+), 32 deletions(-)
create mode 100644 arch/x86/include/asm/hreset.h

--
2.25.1

\
 
 \ /
  Last update: 2022-09-10 01:09    [W:0.362 / U:1.160 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site