Re: [RFC PATCH 00/14] Introducing TIF_NOTIFY_IPI flag
On Wed, 6 Mar 2024, Vincent Guittot wrote:
> Hi Prateek,
>
> Adding Julia who could be interested in this patchset. Your patchset
> should trigger idle load balance instead of newly idle load balance
> now when the polling is used. This was one reason for not migrating
> task in idle CPU
My situation is roughly as follows:
The machine is an Intel 6130 with two sockets and 32 hardware threads
(subsequently referred to as cores) per socket. The test is bt.B of the
OpenMP version of the NAS benchmark suite. Initially there is one
thread per core. NUMA balancing occurs, resulting in a move, and thus 31
threads on one socket and 33 on the other.
Load balancing should result in the idle core pulling one of the threads
from the other socket. But that doesn't happen in normal load balancing,
because all 33 threads on the overloaded socket are considered to have a
preference for that socket. Active balancing could pull a thread, but it
is not triggered because the idle core is seen as being newly idle.
The question is then why a core that has been idle for up to multiple
seconds is continually seen as newly idle. Every 4ms, a scheduler tick
submits some work to try to load balance. This submission process
previously broke out of the idle loop due to a need_resched, hence the
same issue as involved in this patch series. The need_resched caused
invocation of schedule, which would then see that there was no task to
pick, making the core be considered to be newly idle. The classification
as newly idle doesn't take into account whether any task was running prior
to the call to schedule.
The load balancing work that was submitted every 4ms is also a NOP due a
test for need_resched.
This patch series no longer makes need resched be the only way out of the
idle loop. Without the need resched, the load balancing work that is
submitted every 4ms can actually try to do load balancing. The core is
not newly idle, so active balancing could in principle occur. But now
nothing happens because the work is run by ksoftirqd. The presence of
ksoftirqd on the idle core means that the core is no longer idle. Thus
there is no more need for load balancing.
So this patch series in itself doesn't solve the problem. I did 500 runs
with this patch series and 500 runs with the Linux kernel that this patch
series builds on, and there is essentially no difference in the
performance.
julia
>
> On Tue, 20 Feb 2024 at 18:15, K Prateek Nayak wrote:
> >
> > Hello everyone,
> >
> > Before jumping into the issue, let me clarify the Cc list. Everyone have
> > been cc'ed on Patch 0 through Patch 3. Respective arch maintainers,
> > reviewers, and committers returned by scripts/get_maintainer.pl have
> > been cc'ed on the respective arch side changes. Scheduler and CPU Idle
> > maintainers and reviewers have been included for the entire series. If I
> > have missed anyone, please do add them. If you would like to be dropped
> > from the cc list, wholly or partially, for the future iterations, please
> > do let me know.
> >
> > With that out of the way ...
> >
> > Problem statement
> > =
> >
> > When measuring IPI throughput using a modified version of Anton
> > Blanchard's ipistorm benchmark [1], configured to measure time taken to
> > perform a fixed number of smp_call_function_single() (with wait set to
> > 1), an increase in benchmark time was observed between v5.7 and the
> > current upstream release (v6.7-rc6 at the time of encounter).
> >
> > Bisection pointed to commit b2a02fc43a1f ("smp: Optimize
> > send_call_function_single_ipi()") as the reason behind this increase in
> > runtime.
> >
> >
> > Experiments
> > ===
> >
> > Since the commit cannot be cleanly reverted on top of the current
> > tip:sched/core, the effects of the optimizations were reverted by:
> >
> > 1. Removing the check for call_function_single_prep_ipi() in
> >send_call_function_single_ipi(). With this change
> >send_call_function_single_ipi() always calls
> >arch_send_call_function_single_ipi()
> >
> > 2. Removing the call to flush_smp_call_function_queue() in do_idle()
> >since every smp_call_function, with (1.), would unconditionally send
> >an IPI to an idle CPU in TIF_POLLING mode.
> >
> > Following is the diff of the above described changes which will be
> > henceforth referred to as the "revert":
> >
> > diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
> > index 31231925f1ec..735184d98c0f 100644
> > --- a/kernel/sched/idle.c
> > +++ b/kernel/sched/idle.c
> > @@ -332,11 +332,6 @@ static void do_idle(void)
> > */
> > smp_mb__after_atomic();
> >
> > - /*
> > -* RCU relies on this call to be done outside of an RCU read-side
> > -* critical section.
> > -*/
> > - flush_smp_call_function_queue();
> > schedule_idle();
> >
> > if (unlikely(klp_patch_pending(current)))
> > diff --git a/kernel/smp.c b/kernel/smp.c
> > index
Re: [RFC PATCH 00/14] Introducing TIF_NOTIFY_IPI flag
On Wed, 6 Mar 2024 at 11:18, K Prateek Nayak wrote:
>
> Hello Vincent,
>
> Thank you for taking a look at the series.
>
> On 3/6/2024 3:29 PM, Vincent Guittot wrote:
> > Hi Prateek,
> >
> > Adding Julia who could be interested in this patchset. Your patchset
> > should trigger idle load balance instead of newly idle load balance
> > now when the polling is used. This was one reason for not migrating
> > task in idle CPU
>
> Thank you.
>
> >
> > On Tue, 20 Feb 2024 at 18:15, K Prateek Nayak
> > wrote:
> >>
> >> Hello everyone,
> >>
> >> [..snip..]
> >>
> >>
> >> Skipping newidle_balance()
> >> ==
> >>
> >> In an earlier attempt to solve the challenge of the long IRQ disabled
> >> section, newidle_balance() was skipped when a CPU waking up from idle
> >> was found to have no runnable tasks, and was transitioning back to
> >> idle [2]. Tim [3] and David [4] had pointed out that newidle_balance()
> >> may be viable for CPUs that are idling with tick enabled, where the
> >> newidle_balance() has the opportunity to pull tasks onto the idle CPU.
> >>
> >> Vincent [5] pointed out a case where the idle load kick will fail to
> >> run on an idle CPU since the IPI handler launching the ILB will check
> >> for need_resched(). In such cases, the idle CPU relies on
> >> newidle_balance() to pull tasks towards itself.
> >
> > Calling newidle_balance() instead of the normal idle load balance
> > prevents the CPU to pull tasks from other groups
>
> Thank you for the correction.
>
> >
> >>
> >> Using an alternate flag instead of NEED_RESCHED to indicate a pending
> >> IPI was suggested as the correct approach to solve this problem on the
> >> same thread.
> >>
> >>
> >> Proposed solution: TIF_NOTIFY_IPI
> >> =
> >>
> >> Instead of reusing TIF_NEED_RESCHED bit to pull an TIF_POLLING CPU out
> >> of idle, TIF_NOTIFY_IPI is a newly introduced flag that
> >> call_function_single_prep_ipi() sets on a target TIF_POLLING CPU to
> >> indicate a pending IPI, which the idle CPU promises to process soon.
> >>
> >> On architectures that do not support the TIF_NOTIFY_IPI flag (this
> >> series only adds support for x86 and ARM processors for now),
> >
> > I'm surprised that you are mentioning ARM processors because they
> > don't use TIF_POLLING.
>
> Yup I just realised that after Linus Walleij pointed it out on the
> thread.
>
> >
> >> call_function_single_prep_ipi() will fallback to setting
> >> TIF_NEED_RESCHED bit to pull the TIF_POLLING CPU out of idle.
> >>
> >> Since the pending IPI handlers are processed before the call to
> >> schedule_idle() in do_idle(), schedule_idle() will only be called if the
> >> IPI handler have woken / migrated a new task on the idle CPU and has set
> >> TIF_NEED_RESCHED bit to indicate the same. This avoids running into the
> >> long IRQ disabled section in schedule_idle() unnecessarily, and any
> >> need_resched() check within a call function will accurately notify if a
> >> task is waiting for CPU time on the CPU handling the IPI.
> >>
> >> Following is the crude visualization of how the situation changes with
> >> the newly introduced TIF_NOTIFY_IPI flag:
> >> --
> >> CPU0CPU1
> >>
> >> do_idle() {
> >>
> >> __current_set_polling();
> >> ...
> >>
> >> monitor(addr);
> >> if
> >> (!need_resched_or_ipi())
> >>
> >> mwait() {
> >> /*
> >> Waiting */
> >> smp_call_function_single(CPU1, func, wait = 1) {
> >> ...
> >> ...
> >> ...
> >> set_nr_if_polling(CPU1) {
> >> ...
> >> /* Realizes CPU1 is polling */
> >> ...
> >> try_cmpxchg(addr,
> >> ...
> >> ,
> >> ...
> >> val | _TIF_NOTIFY_IPI);
> >> ...
> >> } /* Does not send an IPI */
> >> ...
> >> ... }
> >> /* mwait exit due to write at addr */
> >> csd_lock_wait() { ...
> >> /* Waiting */
> >>
Re: [RFC PATCH 00/14] Introducing TIF_NOTIFY_IPI flag
Hello Vincent,
Thank you for taking a look at the series.
On 3/6/2024 3:29 PM, Vincent Guittot wrote:
> Hi Prateek,
>
> Adding Julia who could be interested in this patchset. Your patchset
> should trigger idle load balance instead of newly idle load balance
> now when the polling is used. This was one reason for not migrating
> task in idle CPU
Thank you.
>
> On Tue, 20 Feb 2024 at 18:15, K Prateek Nayak wrote:
>>
>> Hello everyone,
>>
>> [..snip..]
>>
>>
>> Skipping newidle_balance()
>> ==
>>
>> In an earlier attempt to solve the challenge of the long IRQ disabled
>> section, newidle_balance() was skipped when a CPU waking up from idle
>> was found to have no runnable tasks, and was transitioning back to
>> idle [2]. Tim [3] and David [4] had pointed out that newidle_balance()
>> may be viable for CPUs that are idling with tick enabled, where the
>> newidle_balance() has the opportunity to pull tasks onto the idle CPU.
>>
>> Vincent [5] pointed out a case where the idle load kick will fail to
>> run on an idle CPU since the IPI handler launching the ILB will check
>> for need_resched(). In such cases, the idle CPU relies on
>> newidle_balance() to pull tasks towards itself.
>
> Calling newidle_balance() instead of the normal idle load balance
> prevents the CPU to pull tasks from other groups
Thank you for the correction.
>
>>
>> Using an alternate flag instead of NEED_RESCHED to indicate a pending
>> IPI was suggested as the correct approach to solve this problem on the
>> same thread.
>>
>>
>> Proposed solution: TIF_NOTIFY_IPI
>> =
>>
>> Instead of reusing TIF_NEED_RESCHED bit to pull an TIF_POLLING CPU out
>> of idle, TIF_NOTIFY_IPI is a newly introduced flag that
>> call_function_single_prep_ipi() sets on a target TIF_POLLING CPU to
>> indicate a pending IPI, which the idle CPU promises to process soon.
>>
>> On architectures that do not support the TIF_NOTIFY_IPI flag (this
>> series only adds support for x86 and ARM processors for now),
>
> I'm surprised that you are mentioning ARM processors because they
> don't use TIF_POLLING.
Yup I just realised that after Linus Walleij pointed it out on the
thread.
>
>> call_function_single_prep_ipi() will fallback to setting
>> TIF_NEED_RESCHED bit to pull the TIF_POLLING CPU out of idle.
>>
>> Since the pending IPI handlers are processed before the call to
>> schedule_idle() in do_idle(), schedule_idle() will only be called if the
>> IPI handler have woken / migrated a new task on the idle CPU and has set
>> TIF_NEED_RESCHED bit to indicate the same. This avoids running into the
>> long IRQ disabled section in schedule_idle() unnecessarily, and any
>> need_resched() check within a call function will accurately notify if a
>> task is waiting for CPU time on the CPU handling the IPI.
>>
>> Following is the crude visualization of how the situation changes with
>> the newly introduced TIF_NOTIFY_IPI flag:
>> --
>> CPU0CPU1
>>
>> do_idle() {
>>
>> __current_set_polling();
>> ...
>>
>> monitor(addr);
>> if
>> (!need_resched_or_ipi())
>>
>> mwait() {
>> /*
>> Waiting */
>> smp_call_function_single(CPU1, func, wait = 1) {
>>...
>> ...
>>...
>> set_nr_if_polling(CPU1) {
>>...
>> /* Realizes CPU1 is polling */
>>...
>> try_cmpxchg(addr,
>>...
>> ,
>>...
>> val | _TIF_NOTIFY_IPI);
>>...
>> } /* Does not send an IPI */
>>...
>> ... } /*
>> mwait exit due to write at addr */
>> csd_lock_wait() { ...
>> /* Waiting */
>> preempt_fold_need_resched(); /* fold if NEED_RESCHED */
>> ...
>> __current_clr_polling();
>> ...
>> flush_smp_call_function_queue() {
>> ...
Re: [RFC PATCH 00/14] Introducing TIF_NOTIFY_IPI flag
Hi Prateek,
Adding Julia who could be interested in this patchset. Your patchset
should trigger idle load balance instead of newly idle load balance
now when the polling is used. This was one reason for not migrating
task in idle CPU
On Tue, 20 Feb 2024 at 18:15, K Prateek Nayak wrote:
>
> Hello everyone,
>
> Before jumping into the issue, let me clarify the Cc list. Everyone have
> been cc'ed on Patch 0 through Patch 3. Respective arch maintainers,
> reviewers, and committers returned by scripts/get_maintainer.pl have
> been cc'ed on the respective arch side changes. Scheduler and CPU Idle
> maintainers and reviewers have been included for the entire series. If I
> have missed anyone, please do add them. If you would like to be dropped
> from the cc list, wholly or partially, for the future iterations, please
> do let me know.
>
> With that out of the way ...
>
> Problem statement
> =
>
> When measuring IPI throughput using a modified version of Anton
> Blanchard's ipistorm benchmark [1], configured to measure time taken to
> perform a fixed number of smp_call_function_single() (with wait set to
> 1), an increase in benchmark time was observed between v5.7 and the
> current upstream release (v6.7-rc6 at the time of encounter).
>
> Bisection pointed to commit b2a02fc43a1f ("smp: Optimize
> send_call_function_single_ipi()") as the reason behind this increase in
> runtime.
>
>
> Experiments
> ===
>
> Since the commit cannot be cleanly reverted on top of the current
> tip:sched/core, the effects of the optimizations were reverted by:
>
> 1. Removing the check for call_function_single_prep_ipi() in
>send_call_function_single_ipi(). With this change
>send_call_function_single_ipi() always calls
>arch_send_call_function_single_ipi()
>
> 2. Removing the call to flush_smp_call_function_queue() in do_idle()
>since every smp_call_function, with (1.), would unconditionally send
>an IPI to an idle CPU in TIF_POLLING mode.
>
> Following is the diff of the above described changes which will be
> henceforth referred to as the "revert":
>
> diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
> index 31231925f1ec..735184d98c0f 100644
> --- a/kernel/sched/idle.c
> +++ b/kernel/sched/idle.c
> @@ -332,11 +332,6 @@ static void do_idle(void)
> */
> smp_mb__after_atomic();
>
> - /*
> -* RCU relies on this call to be done outside of an RCU read-side
> -* critical section.
> -*/
> - flush_smp_call_function_queue();
> schedule_idle();
>
> if (unlikely(klp_patch_pending(current)))
> diff --git a/kernel/smp.c b/kernel/smp.c
> index f085ebcdf9e7..2ff100c41885 100644
> --- a/kernel/smp.c
> +++ b/kernel/smp.c
> @@ -111,11 +111,9 @@ void __init call_function_init(void)
> static __always_inline void
> send_call_function_single_ipi(int cpu)
> {
> - if (call_function_single_prep_ipi(cpu)) {
> - trace_ipi_send_cpu(cpu, _RET_IP_,
> -
> generic_smp_call_function_single_interrupt);
> - arch_send_call_function_single_ipi(cpu);
> - }
> + trace_ipi_send_cpu(cpu, _RET_IP_,
> + generic_smp_call_function_single_interrupt);
> + arch_send_call_function_single_ipi(cpu);
> }
>
> static __always_inline void
> --
>
> With the revert, the time taken to complete a fixed set of IPIs using
> ipistorm improves significantly. Following are the numbers from a dual
> socket 3rd Generation EPYC system (2 x 64C/128T) (boost on, C2 disabled)
> running ipistorm between CPU8 and CPU16:
>
> cmdline: insmod ipistorm.ko numipi=10 single=1 offset=8 cpulist=8 wait=1
>
> (tip:sched/core at tag "sched-core-2024-01-08" for all the testing done
> below)
>
> ==
> Test : ipistorm (modified)
> Units : Normalized runtime
> Interpretation: Lower is better
> Statistic : AMean
> ==
> kernel: time [pct imp]
> tip:sched/core1.00 [0.00]
> tip:sched/core + revert 0.81 [19.36]
>
> Although the revert improves ipistorm performance, it also regresses
> tbench and netperf, supporting the validity of the optimization.
> Following are netperf and tbench numbers from the same machine comparing
> vanilla tip:sched/core and the revert applied on top:
>
> ==
> Test : tbench
> Units : Normalized throughput
> Interpretation: Higher is better
> Statistic : AMean
> ==
> Clients:tip[pct imp](CV) revert[pct imp](CV)
> 1 1.00 [ 0.00]( 0.24) 0.91 [ -8.96]( 0.30)
> 2 1.00 [ 0.00]( 0.25) 0.92 [ -8.20]( 0.97)
> 4 1.00 [ 0.00]( 0.23) 0.91 [ -9.20]( 1.75)
[RFC PATCH 00/14] Introducing TIF_NOTIFY_IPI flag
Hello everyone,
Before jumping into the issue, let me clarify the Cc list. Everyone have
been cc'ed on Patch 0 through Patch 3. Respective arch maintainers,
reviewers, and committers returned by scripts/get_maintainer.pl have
been cc'ed on the respective arch side changes. Scheduler and CPU Idle
maintainers and reviewers have been included for the entire series. If I
have missed anyone, please do add them. If you would like to be dropped
from the cc list, wholly or partially, for the future iterations, please
do let me know.
With that out of the way ...
Problem statement
=
When measuring IPI throughput using a modified version of Anton
Blanchard's ipistorm benchmark [1], configured to measure time taken to
perform a fixed number of smp_call_function_single() (with wait set to
1), an increase in benchmark time was observed between v5.7 and the
current upstream release (v6.7-rc6 at the time of encounter).
Bisection pointed to commit b2a02fc43a1f ("smp: Optimize
send_call_function_single_ipi()") as the reason behind this increase in
runtime.
Experiments
===
Since the commit cannot be cleanly reverted on top of the current
tip:sched/core, the effects of the optimizations were reverted by:
1. Removing the check for call_function_single_prep_ipi() in
send_call_function_single_ipi(). With this change
send_call_function_single_ipi() always calls
arch_send_call_function_single_ipi()
2. Removing the call to flush_smp_call_function_queue() in do_idle()
since every smp_call_function, with (1.), would unconditionally send
an IPI to an idle CPU in TIF_POLLING mode.
Following is the diff of the above described changes which will be
henceforth referred to as the "revert":
diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
index 31231925f1ec..735184d98c0f 100644
--- a/kernel/sched/idle.c
+++ b/kernel/sched/idle.c
@@ -332,11 +332,6 @@ static void do_idle(void)
*/
smp_mb__after_atomic();
- /*
-* RCU relies on this call to be done outside of an RCU read-side
-* critical section.
-*/
- flush_smp_call_function_queue();
schedule_idle();
if (unlikely(klp_patch_pending(current)))
diff --git a/kernel/smp.c b/kernel/smp.c
index f085ebcdf9e7..2ff100c41885 100644
--- a/kernel/smp.c
+++ b/kernel/smp.c
@@ -111,11 +111,9 @@ void __init call_function_init(void)
static __always_inline void
send_call_function_single_ipi(int cpu)
{
- if (call_function_single_prep_ipi(cpu)) {
- trace_ipi_send_cpu(cpu, _RET_IP_,
- generic_smp_call_function_single_interrupt);
- arch_send_call_function_single_ipi(cpu);
- }
+ trace_ipi_send_cpu(cpu, _RET_IP_,
+ generic_smp_call_function_single_interrupt);
+ arch_send_call_function_single_ipi(cpu);
}
static __always_inline void
--
With the revert, the time taken to complete a fixed set of IPIs using
ipistorm improves significantly. Following are the numbers from a dual
socket 3rd Generation EPYC system (2 x 64C/128T) (boost on, C2 disabled)
running ipistorm between CPU8 and CPU16:
cmdline: insmod ipistorm.ko numipi=10 single=1 offset=8 cpulist=8 wait=1
(tip:sched/core at tag "sched-core-2024-01-08" for all the testing done
below)
==
Test : ipistorm (modified)
Units : Normalized runtime
Interpretation: Lower is better
Statistic : AMean
==
kernel: time [pct imp]
tip:sched/core1.00 [0.00]
tip:sched/core + revert 0.81 [19.36]
Although the revert improves ipistorm performance, it also regresses
tbench and netperf, supporting the validity of the optimization.
Following are netperf and tbench numbers from the same machine comparing
vanilla tip:sched/core and the revert applied on top:
==
Test : tbench
Units : Normalized throughput
Interpretation: Higher is better
Statistic : AMean
==
Clients:tip[pct imp](CV) revert[pct imp](CV)
1 1.00 [ 0.00]( 0.24) 0.91 [ -8.96]( 0.30)
2 1.00 [ 0.00]( 0.25) 0.92 [ -8.20]( 0.97)
4 1.00 [ 0.00]( 0.23) 0.91 [ -9.20]( 1.75)
8 1.00 [ 0.00]( 0.69) 0.91 [ -9.48]( 1.56)
16 1.00 [ 0.00]( 0.66) 0.92 [ -8.49]( 2.43)
32 1.00 [ 0.00]( 0.96) 0.89 [-11.13]( 0.96)
64 1.00 [ 0.00]( 1.06) 0.90 [ -9.72]( 2.49)
128 1.00 [ 0.00]( 0.70) 0.92 [ -8.36]( 1.26)
256 1.00 [ 0.00]( 0.72) 0.97 [ -3.30]( 1.10)
512 1.00 [ 0.00]( 0.42) 0.98 [ -1.73]( 0.37)
1024 1.00 [ 0.00]( 0.28) 0.99 [ -1.39]( 0.43)
