[PATCH v4 0/6] sched/fair: Fix load balancing of SMT siblings with ASYM_PACKING

From: Ricardo Neri
Date: Tue Aug 10 2021 - 10:41:46 EST

Next message: Ricardo Neri: "[PATCH v4 1/6] x86/sched: Decrease further the priorities of SMT siblings"
Previous message: Nitesh Lal: "Re: [RFC] genirq: Add effective CPU index retrieving interface"
Next in thread: Ricardo Neri: "[PATCH v4 1/6] x86/sched: Decrease further the priorities of SMT siblings"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

This is v4 the series. v1, v2, and v3 patches and test results can be found
in [1], [2], and [3], respectively.

=== Problem statement ===
++ Load balancing ++

When using asymmetric packing, there exists CPU topologies with three
priority levels in which only a subset of the physical cores support SMT.
An instance of such topology is Intel Alderlake, a hybrid processor with
a mixture of Intel Core (with support for SMT) and Intel Atom CPUs.

On Alderlake, it is almost always beneficial to spread work by picking
first the Core CPUs, then the Atoms and at last the SMT siblings.

The current load balancer, however, does not behave as described when using
ASYM_PACKING. Instead, the load balancer will choose higher-priority CPUs
(an Intel Core) over medium-priority CPUs (an Intel Atom), and subsequently
overflow the load to a low priority SMT sibling CPU. This leaves medium-
priority CPUs idle while low-priority CPUs are busy.

This patchset fixes this behavior by also checking the idle state of the
SMT siblings of both the CPU doing the load balance and the busiest
candidate group when deciding whether the destination CPUs can pull tasks
from the busiest CPU.

++ Rework ASYM_PACKING priorities with ITMT
We also reworked the priority computation of the SMT siblings to ensure
that higher-numbered SMT siblings are always low priority. The current
computation may lead to situations in which in some processors those
higher-numbered SMT siblings have the same priority as the Intel Atom
CPUs.

=== Testing ===
I ran a few benchmarks with and without this version of the patchset on
an Intel Alderlake system with 8 Intel Core (with SMT) and 8 Intel
Atom CPUs.

The baseline for the results is an unmodified v5.13-rc4 kernel. Results
show a comparative percentage of improvement (positive) or degradation
(negative). Each test case is repeated eight times, and the standard
deviation among repetitions is also documented.

Table 1 shows the results when using hardware-controlled performance
performance states (HWP), a common use case. The impact of the patches
is overall positive with a few test cases showing slight degradation.
hackbench is especially difficult to assess it shows a high degree of
variability.

Thanks and BR,
Ricardo

ITMT: Intel Turbo Boost Max Technology 3.0

========
Changes since v3:
* Reworked the ITMT priority computation to further reduce the priority
of SMT siblings (patch 1).
* Clear scheduling group flags when a child scheduling level
degenerates (patch 2).
* Removed arch-specific hooks (patch 6, PeterZ)
* Removed redundant checks for the local group. (patch 5, Dietmar)
* Removed redundant check for local idle CPUs and local group
utilization. (patch 6, Joel)
* Reworded commit messages of patches 2, 3, 5, and 6 for clarity.
(Len, PeterZ)
* Added Joel's Reviewed-by tag.
* Unchanged patches: 4

Changes since v2:
* Removed arch_sched_asym_prefer_early() and re-evaluation of the
candidate busiest group in update_sd_pick_busiest(). (PeterZ)
* Introduced sched_group::flags to reflect the properties of CPUs
in the scheduling group. This helps to identify scheduling groups
whose CPUs are SMT siblings. (PeterZ)
* Modified update_sg_lb_stats() to get statistics of the scheduling
domain as an argument. This provides it with the statistics of the
local domain. (PeterZ)
* Introduced sched_asym() to decide if a busiest candidate group can
be marked for asymmetric packing.
* Reworded patch 1 for clarity. (PeterZ)

Changes since v1:
* Don't bailout in update_sd_pick_busiest() if dst_cpu cannot pull
tasks. Instead, reclassify the candidate busiest group, as it
may still be selected. (PeterZ)
* Avoid an expensive and unnecessary call to cpumask_weight() when
determining if a sched_group is comprised of SMT siblings.
(PeterZ).
* Updated test results using the v2 patches.

======== Table 1. Test results of patches with HWP ========
=======================================================================
hackbench
=========
case load baseline(std%) compare%( std%)
process-pipe group-1 1.00 ( 1.45) -2.99 ( 4.35)
process-pipe group-2 1.00 ( 18.32) +2.14 ( 8.33)
process-pipe group-4 1.00 ( 17.27) -10.68 ( 15.85)
process-pipe group-8 1.00 ( 12.33) +2.26 ( 13.28)
process-pipe group-12 1.00 ( 6.52) -4.07 ( 7.97)
process-pipe group-16 1.00 ( 9.70) -7.71 ( 6.01)
process-pipe group-20 1.00 ( 2.52) -4.15 ( 6.35)
process-pipe group-24 1.00 ( 4.84) +1.04 ( 4.60)
process-pipe group-32 1.00 ( 4.79) +1.72 ( 5.13)
process-pipe group-48 1.00 ( 6.77) +4.68 ( 4.24)
process-sockets group-1 1.00 ( 1.89) +0.53 ( 3.48)
process-sockets group-2 1.00 ( 7.57) -6.16 ( 4.52)
process-sockets group-4 1.00 ( 14.62) +4.93 ( 7.11)
process-sockets group-8 1.00 ( 7.69) +3.15 ( 7.19)
process-sockets group-12 1.00 ( 4.97) +2.49 ( 2.80)
process-sockets group-16 1.00 ( 3.93) -1.57 ( 3.86)
process-sockets group-20 1.00 ( 2.56) -3.63 ( 2.88)
process-sockets group-24 1.00 ( 3.00) +0.74 ( 3.01)
process-sockets group-32 1.00 ( 7.63) +1.79 ( 3.67)
process-sockets group-48 1.00 ( 4.15) -0.44 ( 3.70)
threads-pipe group-1 1.00 ( 2.34) -0.55 ( 3.78)
threads-pipe group-2 1.00 ( 12.74) -2.24 ( 12.96)
threads-pipe group-4 1.00 ( 10.03) +5.80 ( 16.02)
threads-pipe group-8 1.00 ( 7.45) -12.09 ( 22.91)
threads-pipe group-12 1.00 ( 5.00) -15.25 ( 10.86)
threads-pipe group-16 1.00 ( 7.41) +1.95 ( 11.73)
threads-pipe group-20 1.00 ( 7.31) -1.72 ( 5.17)
threads-pipe group-24 1.00 ( 4.48) +0.43 ( 6.39)
threads-pipe group-32 1.00 ( 3.75) -0.62 ( 3.87)
threads-pipe group-48 1.00 ( 1.56) -3.69 ( 5.99)
threads-sockets group-1 1.00 ( 2.27) +3.51 ( 3.79)
threads-sockets group-2 1.00 ( 6.86) -8.42 ( 11.39)
threads-sockets group-4 1.00 ( 5.28) -14.35 ( 8.73)
threads-sockets group-8 1.00 ( 11.74) +5.04 ( 5.18)
threads-sockets group-12 1.00 ( 3.29) -6.15 ( 7.08)
threads-sockets group-16 1.00 ( 5.07) -1.40 ( 6.49)
threads-sockets group-20 1.00 ( 4.38) -5.44 ( 5.68)
threads-sockets group-24 1.00 ( 7.20) +3.67 ( 3.99)
threads-sockets group-32 1.00 ( 3.30) +0.10 ( 3.08)
threads-sockets group-48 1.00 ( 4.83) +4.83 ( 3.61)

netperf
=======
case load baseline(std%) compare%( std%)
TCP_RR thread-1 1.00 ( 0.00) +1.20 ( 0.71)
TCP_RR thread-2 1.00 ( 3.03) +0.25 ( 3.05)
TCP_RR thread-4 1.00 ( 1.99) +1.21 ( 3.40)
TCP_RR thread-8 1.00 ( 1.36) -2.00 ( 6.79)
TCP_RR thread-12 1.00 ( 4.53) +1.68 ( 0.95)
TCP_RR thread-16 1.00 ( 0.56) -1.20 ( 1.27)
TCP_RR thread-20 1.00 ( 2.41) +0.68 ( 0.36)
TCP_RR thread-24 1.00 ( 1.27) -0.16 ( 0.29)
TCP_RR thread-32 1.00 ( 11.14) -0.35 ( 11.10)
TCP_RR thread-48 1.00 ( 18.56) -0.12 ( 19.98)
UDP_RR thread-1 1.00 ( 0.00) +1.85 ( 1.53)
UDP_RR thread-2 1.00 ( 2.78) -2.02 ( 3.14)
UDP_RR thread-4 1.00 ( 2.26) -0.54 ( 2.42)
UDP_RR thread-8 1.00 ( 1.46) -1.49 ( 5.31)
UDP_RR thread-12 1.00 ( 0.70) +0.21 ( 1.64)
UDP_RR thread-16 1.00 ( 1.26) -1.90 ( 2.87)
UDP_RR thread-20 1.00 ( 0.29) +0.29 ( 0.27)
UDP_RR thread-24 1.00 ( 2.75) +2.61 ( 0.97)
UDP_RR thread-32 1.00 ( 11.16) -2.90 ( 11.26)
UDP_RR thread-48 1.00 ( 19.22) +3.12 ( 17.32)

tbench
======
case load baseline(std%) compare%( std%)
loopback thread-1 1.00 ( 1.33) +1.41 ( 0.38)
loopback thread-2 1.00 ( 1.03) +0.06 ( 1.08)
loopback thread-4 1.00 ( 1.05) -0.74 ( 1.65)
loopback thread-8 1.00 ( 5.81) +6.66 ( 0.31)
loopback thread-12 1.00 ( 0.60) +0.32 ( 0.20)
loopback thread-16 1.00 ( 1.38) +1.67 ( 0.34)
loopback thread-20 1.00 ( 2.49) +0.24 ( 2.96)
loopback thread-24 1.00 ( 2.26) +0.75 ( 0.29)
loopback thread-32 1.00 ( 2.54) +0.07 ( 2.83)
loopback thread-48 1.00 ( 2.40) +1.90 ( 2.32)

schbench
========
case load baseline(std%) compare%( std%)
normal mthread-1 1.00 ( 3.51) +1.03 ( 2.96)
normal mthread-2 1.00 ( 2.33) +6.35 ( 6.19)
normal mthread-4 1.00 ( 9.58) +0.12 ( 8.02)
normal mthread-8 1.00 ( 1.59) +0.40 ( 0.07)
normal mthread-12 1.00 ( 3.18) +1.51 ( 4.19)
normal mthread-16 1.00 ( 1.72) +0.00 ( 1.72)
normal mthread-20 1.00 ( 1.43) +0.80 ( 3.16)
normal mthread-24 1.00 ( 1.38) -1.69 ( 1.71)
normal mthread-32 1.00 ( 1.86) -0.78 ( 2.25)
normal mthread-48 1.00 ( 3.50) -2.81 ( 2.82)
========

[1]. https://lore.kernel.org/lkml/20210406041108.7416-1-ricardo.neri-calderon@xxxxxxxxxxxxxxx/
[2]. https://lore.kernel.org/lkml/20210414020436.12980-1-ricardo.neri-calderon@xxxxxxxxxxxxxxx/
[3]. https://lore.kernel.org/lkml/20210513154909.6385-1-ricardo.neri-calderon@xxxxxxxxxxxxxxx/

Ricardo Neri (6):
x86/sched: Decrease further the priorities of SMT siblings
sched/topology: Introduce sched_group::flags
sched/fair: Optimize checking for group_asym_packing
sched/fair: Provide update_sg_lb_stats() with sched domain statistics
sched/fair: Carve out logic to mark a group for asymmetric packing
sched/fair: Consider SMT in ASYM_PACKING load balance

arch/x86/kernel/itmt.c | 2 +-
kernel/sched/fair.c | 122 ++++++++++++++++++++++++++++++++++++----
kernel/sched/sched.h | 1 +
kernel/sched/topology.c | 21 ++++++-
4 files changed, 132 insertions(+), 14 deletions(-)

--
2.17.1

Next message: Ricardo Neri: "[PATCH v4 1/6] x86/sched: Decrease further the priorities of SMT siblings"
Previous message: Nitesh Lal: "Re: [RFC] genirq: Add effective CPU index retrieving interface"
Next in thread: Ricardo Neri: "[PATCH v4 1/6] x86/sched: Decrease further the priorities of SMT siblings"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]