Re: [PATCH 3/3] arm64: cpuidle: Add arm_poll_idle
From: Okanovic, Haris
Date: Fri Apr 05 2024 - 16:06:26 EST
On Tue, 2024-04-02 at 18:23 +0100, Mark Rutland wrote:
> CAUTION: This email originated from outside of the organization. Do
> not click links or open attachments unless you can confirm the sender
> and know the content is safe.
>
>
>
> On Mon, Apr 01, 2024 at 08:47:06PM -0500, Haris Okanovic wrote:
> > An arm64 cpuidle driver with two states: (1) First polls for new
> > runable
> > tasks up to 100 us (by default) before (2) a wfi idle and awoken by
> > interrupt (the current arm64 behavior). It allows CPUs to return
> > from
> > idle more quickly by avoiding the longer interrupt wakeup path,
> > which
> > may require EL1/EL2 transition in certain VM scenarios.
>
> Please start off with an explanation of the problem you're trying to
> solve
> (which IIUC is to wake up more quickly in certain cases), before
> describing the
> solution. That makes it *significantly* easier for people to review
> this, since
> once you have the problem statement in mind it's much easier to
> understand how
> the solution space follows from that.
Hi Mark,
I'll try to clarify:
This change improves performance of certain multithreaded workloads
that rapidly cycles through blocked/running states, by avoiding a
relatively costly IPI wakeup path.
One example is Tensorflow as demonstated by the benchmark shared
earlier. It generates 250k reschedule IPIs per sec on AWS m7g.16xlarge
(Graviton3 SoC) in Linux 6.9.0-rc2, which drops to 120k/sec with this
change and a ~3% gain in throughput.
Some Java workloads have a similar pattern. For example, Renissance
Reactors benchmark (Reactors.IO framework app) drops from 370k to 80k
IPIs and sees a ~25% latency drop.
Renissance Reactors repro steps on Ubuntu 22.04 on AWS:
sudo apt-get install java-common
wget
https://corretto.aws/downloads/latest/amazon-corretto-17-aarch64-linux-jdk.deb
sudo dpkg -i amazon-corretto-17-aarch64-linux-jdk.deb
wget
https://github.com/renaissance-benchmarks/renaissance/releases/download/v0.15.0/renaissance-gpl-0.15.0.jar
java -jar renaissance-gpl-0.15.0.jar reactors —repetitions 1
>
> > Poll duration is optionally configured at load time via the
> > poll_limit
> > module parameter.
>
> Why should this be a configurable parameter?
>
> (note, at this point you haven't introduced any of the data below, so
> the
> trade-off isn't clear to anyone).
Some applications could improve with higher poll periods. For example,
Tensorflow throughput improves ~3% at a 100us+ poll limit and ~1% at
10us limit. A configurable knob allows users to experiment more easily.
There’s also no performance penatly for having the option. New values
are patched during load and all other code paths remain the same.
>
> > The default 100 us duration was experimentally chosen, by measuring
> > QPS
> > (queries per sec) of the MLPerf bert inference benchmark, which
> > seems
> > particularly susceptible to this change; see procedure below. 100
> > us is
> > the inflection point where QPS stopped growing in a range of tested
> > values. All results are from AWS m7g.16xlarge instances (Graviton3
> > SoC)
> > with dedicated tenancy (dedicated hardware).
> >
> > > before | 10us | 25us | 50us | 100us | 125us | 150us | 200us |
> > > 300us |
> > > 5.87 | 5.91 | 5.96 | 6.01 | 6.06 | 6.07 | 6.06 | 6.06 |
> > > 6.06 |
> >
> > Perf's scheduler benchmarks also improve with a range of poll_limit
> > values >= 10 us. Higher limits produce near identical results
> > within a
> > 3% noise margin. The following tables are `perf bench sched`
> > results,
> > run times in seconds.
> >
> > `perf bench sched messaging -l 80000`
> > > AWS instance | SoC | Before | After | % Change |
> > > c6g.16xl (VM) | Graviton2 | 18.974 | 18.400 | none |
> > > c7g.16xl (VM) | Graviton3 | 13.852 | 13.859 | none |
> > > c6g.metal | Graviton2 | 17.621 | 16.744 | none |
> > > c7g.metal | Graviton3 | 13.430 | 13.404 | none |
> >
> > `perf bench sched pipe -l 2500000`
> > > AWS instance | SoC | Before | After | % Change |
> > > c6g.16xl (VM) | Graviton2 | 30.158 | 15.181 | -50% |
> > > c7g.16xl (VM) | Graviton3 | 18.289 | 12.067 | -34% |
> > > c6g.metal | Graviton2 | 17.609 | 15.170 | -14% |
> > > c7g.metal | Graviton3 | 14.103 | 12.304 | -13% |
> >
> > `perf bench sched seccomp-notify -l 2500000`
> > > AWS instance | SoC | Before | After | % Change |
> > > c6g.16xl (VM) | Graviton2 | 28.784 | 13.754 | -52% |
> > > c7g.16xl (VM) | Graviton3 | 16.964 | 11.430 | -33% |
> > > c6g.metal | Graviton2 | 15.717 | 13.536 | -14% |
> > > c7g.metal | Graviton3 | 13.301 | 11.491 | -14% |
>
> Ok, so perf numbers for a busy workload go up.
>
> What happens for idle state residency on a mostly idle system?
It decreases by about 0.005% on AWS m7g.16xlarge (Graviton3 SoC).
>
> > Steps to run MLPerf bert inference on Ubuntu 22.04:
> > sudo apt install build-essential python3 python3-pip
> > pip install "pybind11[global]" tensorflow transformers
> > export TF_ENABLE_ONEDNN_OPTS=1
> > export DNNL_DEFAULT_FPMATH_MODE=BF16
> > git clone https://github.com/mlcommons/inference.git --recursive
> > cd inference
> > git checkout v2.0
> > cd loadgen
> > CFLAGS="-std=c++14" python3 setup.py bdist_wheel
> > pip install dist/*.whl
> > cd ../language/bert
> > make setup
> > python3 run.py --backend=tf --scenario=SingleStream
> >
> > Suggested-by: Ali Saidi <alisaidi@xxxxxxxxxx>
> > Reviewed-by: Ali Saidi <alisaidi@xxxxxxxxxx>
> > Reviewed-by: Geoff Blake <blakgeof@xxxxxxxxxx>
> > Cc: Brian Silver <silverbr@xxxxxxxxxx>
> > Signed-off-by: Haris Okanovic <harisokn@xxxxxxxxxx>
> > ---
> > drivers/cpuidle/Kconfig.arm | 13 ++
> > drivers/cpuidle/Makefile | 1 +
> > drivers/cpuidle/cpuidle-arm-polling.c | 171
> > ++++++++++++++++++++++++++
> > 3 files changed, 185 insertions(+)
> > create mode 100644 drivers/cpuidle/cpuidle-arm-polling.c
> >
> > diff --git a/drivers/cpuidle/Kconfig.arm
> > b/drivers/cpuidle/Kconfig.arm
> > index a1ee475d180d..484666dda38d 100644
> > --- a/drivers/cpuidle/Kconfig.arm
> > +++ b/drivers/cpuidle/Kconfig.arm
> > @@ -14,6 +14,19 @@ config ARM_CPUIDLE
> > initialized by calling the CPU operations init idle hook
> > provided by architecture code.
> >
> > +config ARM_POLL_CPUIDLE
> > + bool "ARM64 CPU idle Driver with polling"
> > + depends on ARM64
> > + depends on ARM_ARCH_TIMER_EVTSTREAM
> > + select CPU_IDLE_MULTIPLE_DRIVERS
> > + help
> > + Select this to enable a polling cpuidle driver for ARM64:
> > + The first state polls TIF_NEED_RESCHED for best latency on
> > short
> > + sleep intervals. The second state falls back to
> > arch_cpu_idle() to
> > + wait for interrupt. This is can be helpful in workloads
> > that
> > + frequently block/wake at short intervals or VMs where
> > wakeup IPIs
> > + are more expensive.
>
> Why is this a separate driver rather than an optional feature in the
> existing
> driver?
>
> The fact that this duplicates a bunch of code indicates to me that
> this should
> not be a separate driver.
I'll explore using Ankur Arora’s haltpoll changes on AWS Graviton; see
related thread.
>
> > +
> > config ARM_PSCI_CPUIDLE
> > bool "PSCI CPU idle Driver"
> > depends on ARM_PSCI_FW
> > diff --git a/drivers/cpuidle/Makefile b/drivers/cpuidle/Makefile
> > index d103342b7cfc..23c21422792d 100644
> > --- a/drivers/cpuidle/Makefile
> > +++ b/drivers/cpuidle/Makefile
> > @@ -22,6 +22,7 @@ obj-$(CONFIG_ARM_U8500_CPUIDLE) +=
> > cpuidle-ux500.o
> > obj-$(CONFIG_ARM_AT91_CPUIDLE) += cpuidle-at91.o
> > obj-$(CONFIG_ARM_EXYNOS_CPUIDLE) += cpuidle-exynos.o
> > obj-$(CONFIG_ARM_CPUIDLE) += cpuidle-arm.o
> > +obj-$(CONFIG_ARM_POLL_CPUIDLE) += cpuidle-arm-
> > polling.o
> > obj-$(CONFIG_ARM_PSCI_CPUIDLE) += cpuidle-psci.o
> > obj-$(CONFIG_ARM_PSCI_CPUIDLE_DOMAIN) += cpuidle-psci-
> > domain.o
> > obj-$(CONFIG_ARM_TEGRA_CPUIDLE) += cpuidle-tegra.o
> > diff --git a/drivers/cpuidle/cpuidle-arm-polling.c
> > b/drivers/cpuidle/cpuidle-arm-polling.c
> > new file mode 100644
> > index 000000000000..bca128568114
> > --- /dev/null
> > +++ b/drivers/cpuidle/cpuidle-arm-polling.c
> > @@ -0,0 +1,171 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * ARM64 CPU idle driver using wfe polling
> > + *
> > + * Copyright 2024 Amazon.com, Inc. or its affiliates. All rights
> > reserved.
> > + *
> > + * Authors:
> > + * Haris Okanovic <harisokn@xxxxxxxxxx>
> > + * Brian Silver <silverbr@xxxxxxxxxx>
> > + *
> > + * Based on cpuidle-arm.c
> > + * Copyright (C) 2014 ARM Ltd.
> > + * Author: Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx>
> > + */
> > +
> > +#include <linux/cpu.h>
> > +#include <linux/cpu_cooling.h>
> > +#include <linux/cpuidle.h>
> > +#include <linux/sched/clock.h>
> > +
> > +#include <asm/cpuidle.h>
> > +#include <asm/readex.h>
> > +
> > +#include "dt_idle_states.h"
> > +
> > +/* Max duration of the wfe() poll loop in us, before transitioning
> > to
> > + * arch_cpu_idle()/wfi() sleep.
> > + */
>
> /*
> * Comments should have the leading '/*' on a separate line.
> * See
> https://www.kernel.org/doc/html/v6.8/process/coding-style.html#commenting
> */
>
> > +#define DEFAULT_POLL_LIMIT_US 100
> > +static unsigned int poll_limit __read_mostly =
> > DEFAULT_POLL_LIMIT_US;
> > +
> > +/*
> > + * arm_idle_wfe_poll - Polls state in wfe loop until reschedule is
> > + * needed or timeout
> > + */
> > +static int __cpuidle arm_idle_wfe_poll(struct cpuidle_device *dev,
> > + struct cpuidle_driver *drv, int idx)
> > +{
> > + u64 time_start, time_limit;
> > +
> > + time_start = local_clock();
> > + dev->poll_time_limit = false;
> > +
> > + local_irq_enable();
>
> Why enable IRQs here? We don't do that in the regular cpuidle-arm
> driver, nor
> the cpuidle-psci driver, and there's no explanation here or in the
> commit message.
>
> How does this interact with RCU? Is that still watching or are we in
> an
> extended quiescent state? For PSCI idle states we enter an EQS, and
> it seems
> like we probably should here...
I'll circle back to your code comments (here and elsewhere) if we
proceede with this patch instead of Ankur's.
>
> > +
> > + if (current_set_polling_and_test())
> > + goto end;
> > +
> > + time_limit = cpuidle_poll_time(drv, dev);
> > +
> > + do {
> > + // exclusive read arms the monitor for wfe
> > + if (__READ_ONCE_EX(current_thread_info()->flags) &
> > _TIF_NEED_RESCHED)
> > + goto end;
> > +
> > + // may exit prematurely, see ARM_ARCH_TIMER_EVTSTREAM
> > + wfe();
> > + } while (local_clock() - time_start < time_limit);
>
> ... and if the EVTSTREAM is disabled, we'll sit in WFE forever rather
> than
> entering a deeper idle state, which doesn't seem desirable.
I took a depdnency on ARM_ARCH_TIMER_EVTSTREAM in Kconfig.arm to ensure
event stream is enabled.
>
> It's worth noting that now that we have WFET, we'll probably want to
> disable
> the EVTSTREAM by default at some point, at least in some
> configurations, since
> that'll be able to sit in a WFE state for longer while also reliably
> waking up
> when required.
>
> I suspect we want something like an smp_load_acquire_timeout() here
> to do the
> wait in arch code (allowing us to use WFET), and enabling this state
> will
> depend on either having WFET or EVTSTREAM.
>
> > +
> > + dev->poll_time_limit = true;
> > +
> > +end:
> > + current_clr_polling();
> > + return idx;
> > +}
> > +
> > +/*
> > + * arm_idle_wfi - Places cpu in lower power state until interrupt,
> > + * a fallback to polling
> > + */
> > +static int __cpuidle arm_idle_wfi(struct cpuidle_device *dev,
> > + struct cpuidle_driver *drv, int idx)
> > +{
> > + if (current_clr_polling_and_test()) {
> > + local_irq_enable();
> > + return idx;
> > + }
>
> Same as above, why enable IRQs here?
>
> > + arch_cpu_idle();
> > + return idx;
>
> ... and if we need to enable IRQs in the other cases above, why do we
> *not*
> need to enable them here?
>
> > +}
> > +
> > +static struct cpuidle_driver arm_poll_idle_driver __initdata = {
> > + .name = "arm_poll_idle",
> > + .owner = THIS_MODULE,
> > + .states = {
> > + {
> > + .enter = arm_idle_wfe_poll,
> > + .exit_latency = 0,
> > + .target_residency = 0,
> > + .exit_latency_ns = 0,
> > + .power_usage = UINT_MAX,
> > + .flags =
> > CPUIDLE_FLAG_POLLING,
> > + .name = "WFE",
> > + .desc = "ARM WFE",
> > + },
> > + {
> > + .enter = arm_idle_wfi,
> > + .exit_latency =
> > DEFAULT_POLL_LIMIT_US,
> > + .target_residency =
> > DEFAULT_POLL_LIMIT_US,
> > + .power_usage = UINT_MAX,
> > + .name = "WFI",
> > + .desc = "ARM WFI",
> > + },
> > + },
> > + .state_count = 2,
> > +};
>
> How does this interact with the existing driver?
>
> How does DEFAULT_POLL_LIMIT_US compare with PSCI idle states?
>
> > +
> > +/*
> > + * arm_poll_init_cpu - Initializes arm cpuidle polling driver for
> > one cpu
> > + */
> > +static int __init arm_poll_init_cpu(int cpu)
> > +{
> > + int ret;
> > + struct cpuidle_driver *drv;
> > +
> > + drv = kmemdup(&arm_poll_idle_driver, sizeof(*drv),
> > GFP_KERNEL);
> > + if (!drv)
> > + return -ENOMEM;
> > +
> > + drv->cpumask = (struct cpumask *)cpumask_of(cpu);
> > + drv->states[1].exit_latency = poll_limit;
> > + drv->states[1].target_residency = poll_limit;
> > +
> > + ret = cpuidle_register(drv, NULL);
> > + if (ret) {
> > + pr_err("failed to register driver: %d, cpu %d\n",
> > ret, cpu);
> > + goto out_kfree_drv;
> > + }
> > +
> > + pr_info("registered driver cpu %d\n", cpu);
>
> This does not need to be printed for each CPU.
>
> Mark.
>
> > +
> > + cpuidle_cooling_register(drv);
> > +
> > + return 0;
> > +
> > +out_kfree_drv:
> > + kfree(drv);
> > + return ret;
> > +}
> > +
> > +/*
> > + * arm_poll_init - Initializes arm cpuidle polling driver
> > + */
> > +static int __init arm_poll_init(void)
> > +{
> > + int cpu, ret;
> > + struct cpuidle_driver *drv;
> > + struct cpuidle_device *dev;
> > +
> > + for_each_possible_cpu(cpu) {
> > + ret = arm_poll_init_cpu(cpu);
> > + if (ret)
> > + goto out_fail;
> > + }
> > +
> > + return 0;
> > +
> > +out_fail:
> > + pr_info("de-register all");
> > + while (--cpu >= 0) {
> > + dev = per_cpu(cpuidle_devices, cpu);
> > + drv = cpuidle_get_cpu_driver(dev);
> > + cpuidle_unregister(drv);
> > + kfree(drv);
> > + }
> > +
> > + return ret;
> > +}
> > +
> > +module_param(poll_limit, uint, 0444);
> > +device_initcall(arm_poll_init);
> > --
> > 2.34.1
> >
> >
Thanks,
Haris Okanovic