On Fri, Apr 19, 2024 at 11:32 AM zhangpeng (AS) <[email protected]> wrote:
> On 2024/4/19 10:30, Rongwei Wang wrote:
> > On 2024/4/18 22:20, Peng Zhang wrote:
> >> From: ZhangPeng <[email protected]>
> >>
> >> Since commit f1a7941243c1 ("mm: convert mm's rss stats into
> >> percpu_counter"), the rss_stats have converted into percpu_counter,
> >> which convert the error margin from (nr_threads * 64) to approximately
> >> (nr_cpus ^ 2). However, the new percpu allocation in mm_init() causes a
> >> performance regression on fork/exec/shell. Even after commit
> >> 14ef95be6f55
> >> ("kernel/fork: group allocation/free of per-cpu counters for mm
> >> struct"),
> >> the performance of fork/exec/shell is still poor compared to previous
> >> kernel versions.
> >>
> >> To mitigate performance regression, we delay the allocation of percpu
> >> memory for rss_stats. Therefore, we convert mm's rss stats to use
> >> percpu_counter atomic mode. For single-thread processes, rss_stat is in
> >> atomic mode, which reduces the memory consumption and performance
> >> regression caused by using percpu. For multiple-thread processes,
> >> rss_stat is switched to the percpu mode to reduce the error margin.
> >> We convert rss_stats from atomic mode to percpu mode only when the
> >> second thread is created.
> > Hi, Zhang Peng
> >
> > This regression we also found it in lmbench these days. I have not
> > test your patch, but it seems will solve a lot for it.
> > And I see this patch not fix the regression in multi-threads, that's
> > because of the rss_stat switched to percpu mode?
> > (If I'm wrong, please correct me.) And It seems percpu_counter also
> > has a bad effect in exit_mmap().
> >
> > If so, I'm wondering if we can further improving it on the exit_mmap()
> > path in multi-threads scenario, e.g. to determine which CPUs the
> > process has run on (mm_cpumask()? I'm not sure).
> >
> Hi, Rongwei,
>
> Yes, this patch only fixes the regression in single-thread processes. How
> much bad effect does percpu_counter have in exit_mmap()? IMHO, the addition
> of mm counter is already in batch mode, maybe I miss something?
>
Hi, Peng Zhang, Rongwei, and all:
I've a patch series that is earlier than commit f1a7941243c1 ("mm:
convert mm's rss stats into
percpu_counter"):
https://lwn.net/ml/linux-kernel/[email protected]/
Instead of a per-mm-per-cpu cache, it used only one global per-cpu
cache, and flush it on schedule. Or, if the arch supports, flush and
fetch it use mm bitmap as an optimization (like tlb shootdown).
Unfortunately it didn't get much attention and I moved to work on other things.
I also noticed the fork regression issue, so I did a local rebase of
my previous patch, and revert f1a7941243c1.
The result is looking good, on my 32 core VM machine, I see similar
improvement as the one you posted (alloc/free on fork/exit is gone), I
also see minor improvement with database tests, memory usage is lower
by a little bit too (no more per-mm cache), and I think the error
margin in my patch should be close to zero.
I hope I can get some attention here for my idea...
On Thu, May 16, 2024 at 07:50:52PM +0800, Kairui Song wrote:
> > > On 2024/4/18 22:20, Peng Zhang wrote:
> > >> From: ZhangPeng <[email protected]>
> > >>
> > >> Since commit f1a7941243c1 ("mm: convert mm's rss stats into
> > >> percpu_counter"), the rss_stats have converted into percpu_counter,
> > >> which convert the error margin from (nr_threads * 64) to approximately
> > >> (nr_cpus ^ 2). However, the new percpu allocation in mm_init() causes a
> > >> performance regression on fork/exec/shell. Even after commit
> > >> 14ef95be6f55
> > >> ("kernel/fork: group allocation/free of per-cpu counters for mm
> > >> struct"),
> > >> the performance of fork/exec/shell is still poor compared to previous
> > >> kernel versions.
> > >>
> > >> To mitigate performance regression, we delay the allocation of percpu
> > >> memory for rss_stats. Therefore, we convert mm's rss stats to use
> > >> percpu_counter atomic mode. For single-thread processes, rss_stat is in
> > >> atomic mode, which reduces the memory consumption and performance
> > >> regression caused by using percpu. For multiple-thread processes,
> > >> rss_stat is switched to the percpu mode to reduce the error margin.
> > >> We convert rss_stats from atomic mode to percpu mode only when the
> > >> second thread is created.
>
> I've a patch series that is earlier than commit f1a7941243c1 ("mm:
> convert mm's rss stats into
> percpu_counter"):
>
> https://lwn.net/ml/linux-kernel/[email protected]/
>
> Instead of a per-mm-per-cpu cache, it used only one global per-cpu
> cache, and flush it on schedule. Or, if the arch supports, flush and
> fetch it use mm bitmap as an optimization (like tlb shootdown).
>
I just spotted this thread.
I have a rather long rant to write about the entire ordeal, but don't
have the time at the moment. I do have time to make some remarks though.
Rolling with a centralized counter and only distributing per-cpu upon
creation of a thread is something which was discussed last time and
which I was considering doing. Then life got it in the way and in the
meantime I managed to conclude it's a questionable idea anyway.
The state prior to the counters moving to per-cpu was not that great to
begin with, with quite a few serialization points. As far as allocating
stuff goes one example is mm_alloc_cid, with the following:
mm->pcpu_cid = alloc_percpu(struct mm_cid);
Converting the code to avoid per-cpu rss counters in the common case or
the above patchset only damage-control the state back to what it was,
don't do anything to push things further.
Another note is that unfortunately userspace is increasingly
multithreaded for no good reason, see the Rust ecosystem as an example.
All that to say is that the multithreaded case is what has to get
faster, as a side effect possibly obsoleting both approaches proposed
above. I concede if there is nobody wiling to commit to doing the work
in the foreseeable future then indeed a damage-controlling solution
should land.
On that note in check_mm there is this loop:
for (i = 0; i < NR_MM_COUNTERS; i++) {
long x = percpu_counter_sum(&mm->rss_stat[i]);
This avoidably walks all cpus 4 times with a preemption and lock trip
for each round. Instead one can observe all modifications are supposed
to have already stopped and that this is allocated in a banch. A
routine, say percpu_counter_sum_many_unsafe, could do one iteration
without any locks or interrupt play and return an array. This should be
markedly faster and I perhaps will hack it up.
A part of The Real Solution(tm) would make counter allocations scale
(including mcid, not just rss) or dodge them (while maintaining the
per-cpu distribution, see below for one idea), but that boils down to
balancing scalability versus total memory usage. It is trivial to just
slap together a per-cpu cache of these allocations and have the problem
go away for benchmarking purposes, while being probably being too memory
hungry for actual usage.
I was pondering an allocator with caches per some number of cores (say 4
or 8). Microbenchmarks aside I suspect real workloads would not suffer
from contention at this kind of granularity. This would trivially reduce
memory usage compared to per-cpu caching. I suspect things like
mm_struct, task_struct, task stacks and similar would be fine with it.
Suppose mm_struct is allocated from a more coarse grained allocator than
per-cpu. Total number of cached objects would be lower than it is now.
That would also mean these allocated but not currently used mms could
hold on to other stuff, for example per-cpu rss and mcid counters. Then
should someone fork or exit, alloc/free_percpu would be avoided for most
cases. This would scale better and be faster single-threaded than the
current state.
(believe it or not this is not the actual long rant I have in mind)
I can't commit to work on the Real Solution though.
In the meantime I can submit percpu_counter_sum_many_unsafe as described
above if Denis likes the idea.
Mateusz Guzik <[email protected]> 于 2024年5月16日周四 23:14写道：
>
> On Thu, May 16, 2024 at 07:50:52PM +0800, Kairui Song wrote:
> > > > On 2024/4/18 22:20, Peng Zhang wrote:
> > > >> From: ZhangPeng <[email protected]>
> > > >>
> > > >> Since commit f1a7941243c1 ("mm: convert mm's rss stats into
> > > >> percpu_counter"), the rss_stats have converted into percpu_counter,
> > > >> which convert the error margin from (nr_threads * 64) to approximately
> > > >> (nr_cpus ^ 2). However, the new percpu allocation in mm_init() causes a
> > > >> performance regression on fork/exec/shell. Even after commit
> > > >> 14ef95be6f55
> > > >> ("kernel/fork: group allocation/free of per-cpu counters for mm
> > > >> struct"),
> > > >> the performance of fork/exec/shell is still poor compared to previous
> > > >> kernel versions.
> > > >>
> > > >> To mitigate performance regression, we delay the allocation of percpu
> > > >> memory for rss_stats. Therefore, we convert mm's rss stats to use
> > > >> percpu_counter atomic mode. For single-thread processes, rss_stat is in
> > > >> atomic mode, which reduces the memory consumption and performance
> > > >> regression caused by using percpu. For multiple-thread processes,
> > > >> rss_stat is switched to the percpu mode to reduce the error margin.
> > > >> We convert rss_stats from atomic mode to percpu mode only when the
> > > >> second thread is created.
> >
> > I've a patch series that is earlier than commit f1a7941243c1 ("mm:
> > convert mm's rss stats into
> > percpu_counter"):
> >
> > https://lwn.net/ml/linux-kernel/[email protected]/
> >
> > Instead of a per-mm-per-cpu cache, it used only one global per-cpu
> > cache, and flush it on schedule. Or, if the arch supports, flush and
> > fetch it use mm bitmap as an optimization (like tlb shootdown).
> >
>
> I just spotted this thread.
>
> I have a rather long rant to write about the entire ordeal, but don't
> have the time at the moment. I do have time to make some remarks though.
>
> Rolling with a centralized counter and only distributing per-cpu upon
> creation of a thread is something which was discussed last time and
> which I was considering doing. Then life got it in the way and in the
> meantime I managed to conclude it's a questionable idea anyway.
>
> The state prior to the counters moving to per-cpu was not that great to
> begin with, with quite a few serialization points. As far as allocating
> stuff goes one example is mm_alloc_cid, with the following:
> mm->pcpu_cid = alloc_percpu(struct mm_cid);
>
> Converting the code to avoid per-cpu rss counters in the common case or
> the above patchset only damage-control the state back to what it was,
> don't do anything to push things further.
>
> Another note is that unfortunately userspace is increasingly
> multithreaded for no good reason, see the Rust ecosystem as an example.
>
> All that to say is that the multithreaded case is what has to get
> faster, as a side effect possibly obsoleting both approaches proposed
> above. I concede if there is nobody wiling to commit to doing the work
> in the foreseeable future then indeed a damage-controlling solution
> should land.
Hi, Mateusz,
Which patch are you referencing? My series didn't need any allocations
on thread creation or destruction. Also RSS update is extremely
lightweight (pretty much just read GS and do a few ADD/INC, that's
all), performance is better than all even with micro benchmarks. RSS
read only collects info from CPUs that may contain real updates.
I understand you may not have time to go through my series... but I
think I should add some details here.
> On that note in check_mm there is this loop:
> for (i = 0; i < NR_MM_COUNTERS; i++) {
> long x = percpu_counter_sum(&mm->rss_stat[i]);
>
> This avoidably walks all cpus 4 times with a preemption and lock trip
> for each round. Instead one can observe all modifications are supposed
> to have already stopped and that this is allocated in a banch. A
> routine, say percpu_counter_sum_many_unsafe, could do one iteration
> without any locks or interrupt play and return an array. This should be
> markedly faster and I perhaps will hack it up.
Which is similar to the RSS read in my earlier series... It is based
on the assumption that updates are likely stopped so just read the
counter "unsafely" with a double (and fast) check to ensure no race.
And even more, when coupled with mm shootdown
(CONFIG_ARCH_PCP_RSS_USE_CPUMASK), it doesn't need to collect RSS info
on thread exit at all.
>
> A part of The Real Solution(tm) would make counter allocations scale
> (including mcid, not just rss) or dodge them (while maintaining the
> per-cpu distribution, see below for one idea), but that boils down to
> balancing scalability versus total memory usage. It is trivial to just
> slap together a per-cpu cache of these allocations and have the problem
> go away for benchmarking purposes, while being probably being too memory
> hungry for actual usage.
>
> I was pondering an allocator with caches per some number of cores (say 4
> or 8). Microbenchmarks aside I suspect real workloads would not suffer
> from contention at this kind of granularity. This would trivially reduce
> memory usage compared to per-cpu caching. I suspect things like
> mm_struct, task_struct, task stacks and similar would be fine with it.
>
> Suppose mm_struct is allocated from a more coarse grained allocator than
> per-cpu. Total number of cached objects would be lower than it is now.
> That would also mean these allocated but not currently used mms could
> hold on to other stuff, for example per-cpu rss and mcid counters. Then
> should someone fork or exit, alloc/free_percpu would be avoided for most
> cases. This would scale better and be faster single-threaded than the
> current state.
And what is the issue with using only one CPU cache, and flush on mm
switch? No more alloc after boot, and the total (and fixed) memory
usage is just about a few unsigned long per CPU, which should be even
lower that the old RSS cache solution (4 unsigned long per task). And
it scaled very well with many kinds of microbench or workload I've
tested.
Unless the workload keeps doing something like "alloc one page then
switch to another mm", I think the performance will be horrible
already due to cache invalidations and many switch_*()s, RSS isn't
really a concern there.
>
> (believe it or not this is not the actual long rant I have in mind)
>
> I can't commit to work on the Real Solution though.
>
> In the meantime I can submit percpu_counter_sum_many_unsafe as described
> above if Denis likes the idea.
On Fri, May 17, 2024 at 11:29:57AM +0800, Kairui Song wrote:
> Mateusz Guzik <[email protected]> 于 2024年5月16日周四 23:14写道：
> > A part of The Real Solution(tm) would make counter allocations scale
> > (including mcid, not just rss) or dodge them (while maintaining the
> > per-cpu distribution, see below for one idea), but that boils down to
> > balancing scalability versus total memory usage. It is trivial to just
> > slap together a per-cpu cache of these allocations and have the problem
> > go away for benchmarking purposes, while being probably being too memory
> > hungry for actual usage.
> >
> > I was pondering an allocator with caches per some number of cores (say 4
> > or 8). Microbenchmarks aside I suspect real workloads would not suffer
> > from contention at this kind of granularity. This would trivially reduce
> > memory usage compared to per-cpu caching. I suspect things like
> > mm_struct, task_struct, task stacks and similar would be fine with it.
> >
> > Suppose mm_struct is allocated from a more coarse grained allocator than
> > per-cpu. Total number of cached objects would be lower than it is now.
> > That would also mean these allocated but not currently used mms could
> > hold on to other stuff, for example per-cpu rss and mcid counters. Then
> > should someone fork or exit, alloc/free_percpu would be avoided for most
> > cases. This would scale better and be faster single-threaded than the
> > current state.
>
> And what is the issue with using only one CPU cache, and flush on mm
> switch? No more alloc after boot, and the total (and fixed) memory
> usage is just about a few unsigned long per CPU, which should be even
> lower that the old RSS cache solution (4 unsigned long per task). And
> it scaled very well with many kinds of microbench or workload I've
> tested.
>
> Unless the workload keeps doing something like "alloc one page then
> switch to another mm", I think the performance will be horrible
> already due to cache invalidations and many switch_*()s, RSS isn't
> really a concern there.
>
I only skimmed through your patchset. I do think it has a legitimate
approach, but personally I would not do it like that due to the extra
work on context switches. However, I have 0 say about this, so you will
need to prod the mm overlords to get this moving forward.
Maybe I was not clear enough in my opening e-mail, so I'm going to
reiterate some bits: there are scalability problems in execve even with
your patchset or the one which uses atomics. One of them concerns
another bit which allocates per-cpu memory (the mcid thing).
I note that sorting it out would possibly also take care of the rss
problem, outlining an example approach above.