| Message ID | 20211026165100.ahz5bkx44lrrw5pt@linutronix.de (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [RFC] mm: Disable NUMA_BALANCING_DEFAULT_ENABLED and TRANSPARENT_HUGEPAGE on PREEMPT_RT | expand |
On Tue, Oct 26, 2021 at 06:51:00PM +0200, Sebastian Andrzej Siewior wrote: > In https://lore.kernel.org/all/20200304091159.GN3818@techsingularity.net/ > Mel wrote: > > | While I ack'd this, an RT application using THP is playing with fire, > | I know the RT extension for SLE explicitly disables it from being enabled > | at kernel config time. At minimum the critical regions should be mlocked > | followed by prctl to disable future THP faults that are non-deterministic, > | both from an allocation point of view, and a TLB access point of view. It's > | still reasonable to expect a smaller TLB reach for huge pages than > | base pages. > > With TRANSPARENT_HUGEPAGE enabled I haven't seen spikes > 100us > in cyclictest. I did have mlock_all() enabled but nothing else. > PR_SET_THP_DISABLE remained unchanged (enabled). Is there anything to > stress this to be sure or is mlock_all() enough to do THP but leave the > mlock() applications alone? > > Then Mel continued with: > > | It's a similar hazard with NUMA balancing, an RT application should either > | disable balancing globally or set a memory policy that forces it to be > | ignored. They should be doing this anyway to avoid non-deterministic > | memory access costs due to NUMA artifacts but it wouldn't surprise me > | if some applications got it wrong. > > Usually (often) RT applications are pinned. I would assume that on > bigger box the RT tasks are at least pinned to a node. How bad can this > get in worst case? cyclictest pins every thread to CPU. I could remove > this for testing. What would be a good test to push this to its limit? > > Cc: Mel Gorman <mgorman@techsingularity.net> > Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Somewhat tentative but Acked-by: Mel Gorman <mgorman@techsingularity.net> It's tentative because NUMA Balancing gets default disabled on PREEMPT_RT but it's still possible to enable where as THP is disabled entirely and can never be enabled. This is a little inconsistent and it would be preferable that they match either by disabling NUMA_BALANCING entirely or forbidding TRANSPARENT_HUGEPAGE_ALWAYS && PREEMPT_RT. I'm ok with either. There is the possibility that an RT application could use THP safely by using madvise() and mlock(). That way, THP is available but only if an application has explicit knowledge of THP and smart enough to do it only during the initialisation phase with diff --git a/mm/Kconfig b/mm/Kconfig index d16ba9249bc5..d6ccca216028 100644 --- a/mm/Kconfig +++ b/mm/Kconfig @@ -393,6 +393,7 @@ choice config TRANSPARENT_HUGEPAGE_ALWAYS bool "always" + depends on !PREEMPT_RT help Enabling Transparent Hugepage always, can increase the memory footprint of applications without a guaranteed There is the slight caveat that even then THP can have inconsistent latencies if it has a split THP with separate entries for base and huge pages. The responsibility would be on the person deploying the application to ensure a platform was suitable for both RT and using huge pages.
On 2021-10-27 10:12:12 [+0100], Mel Gorman wrote: > On Tue, Oct 26, 2021 at 06:51:00PM +0200, Sebastian Andrzej Siewior wrote: > > In https://lore.kernel.org/all/20200304091159.GN3818@techsingularity.net/ > > Mel wrote: > > > > | While I ack'd this, an RT application using THP is playing with fire, > > | I know the RT extension for SLE explicitly disables it from being enabled > > | at kernel config time. At minimum the critical regions should be mlocked > > | followed by prctl to disable future THP faults that are non-deterministic, > > | both from an allocation point of view, and a TLB access point of view. It's > > | still reasonable to expect a smaller TLB reach for huge pages than > > | base pages. > > > > With TRANSPARENT_HUGEPAGE enabled I haven't seen spikes > 100us > > in cyclictest. I did have mlock_all() enabled but nothing else. > > PR_SET_THP_DISABLE remained unchanged (enabled). Is there anything to > > stress this to be sure or is mlock_all() enough to do THP but leave the > > mlock() applications alone? > > > > Then Mel continued with: > > > > | It's a similar hazard with NUMA balancing, an RT application should either > > | disable balancing globally or set a memory policy that forces it to be > > | ignored. They should be doing this anyway to avoid non-deterministic > > | memory access costs due to NUMA artifacts but it wouldn't surprise me > > | if some applications got it wrong. > > > > Usually (often) RT applications are pinned. I would assume that on > > bigger box the RT tasks are at least pinned to a node. How bad can this > > get in worst case? cyclictest pins every thread to CPU. I could remove > > this for testing. What would be a good test to push this to its limit? > > > > Cc: Mel Gorman <mgorman@techsingularity.net> > > Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> > > Somewhat tentative but > > Acked-by: Mel Gorman <mgorman@techsingularity.net> > > It's tentative because NUMA Balancing gets default disabled on PREEMPT_RT > but it's still possible to enable where as THP is disabled entirely > and can never be enabled. This is a little inconsistent and it would be > preferable that they match either by disabling NUMA_BALANCING entirely or > forbidding TRANSPARENT_HUGEPAGE_ALWAYS && PREEMPT_RT. I'm ok with either. Oh. I can go either way depending on the input ;) > There is the possibility that an RT application could use THP safely by > using madvise() and mlock(). That way, THP is available but only if an > application has explicit knowledge of THP and smart enough to do it only > during the initialisation phase with Yes that was my question. So if you have "always", do mlock_all() in the application and then have other threads that same application doing malloc/ free of memory that the RT thread is not touching then bad things can still happen, right? My understanding is that all threads can be blocked in a page fault if there is some THP operation going on. You suggest that the application is using THP by setting madvice on the relevant area, mlock afterwards and then nothing bad can happen. No defrag or an optimisation happens later. The memory area uses hugepages after the madvice or not. If so, then this sounds good. > diff --git a/mm/Kconfig b/mm/Kconfig > index d16ba9249bc5..d6ccca216028 100644 > --- a/mm/Kconfig > +++ b/mm/Kconfig > @@ -393,6 +393,7 @@ choice > > config TRANSPARENT_HUGEPAGE_ALWAYS > bool "always" > + depends on !PREEMPT_RT > help > Enabling Transparent Hugepage always, can increase the > memory footprint of applications without a guaranteed > > There is the slight caveat that even then THP can have inconsistent > latencies if it has a split THP with separate entries for base and huge > pages. The responsibility would be on the person deploying the application > to ensure a platform was suitable for both RT and using huge pages. split THP? You mean latencies are different by accessing the memory depending if it is reached via the THP entry or one of the many 4kib entries? I'm more worries about locked mmap_lock while the THP operation is in progress and then a fault from the RT application has to wait until the THP operation is done. Sebastian
On Thu, Oct 28, 2021 at 02:04:29PM +0200, Sebastian Andrzej Siewior wrote: > On 2021-10-27 10:12:12 [+0100], Mel Gorman wrote: > > On Tue, Oct 26, 2021 at 06:51:00PM +0200, Sebastian Andrzej Siewior wrote: > > > In https://lore.kernel.org/all/20200304091159.GN3818@techsingularity.net/ > > > Mel wrote: > > > > > > | While I ack'd this, an RT application using THP is playing with fire, > > > | I know the RT extension for SLE explicitly disables it from being enabled > > > | at kernel config time. At minimum the critical regions should be mlocked > > > | followed by prctl to disable future THP faults that are non-deterministic, > > > | both from an allocation point of view, and a TLB access point of view. It's > > > | still reasonable to expect a smaller TLB reach for huge pages than > > > | base pages. > > > > > > With TRANSPARENT_HUGEPAGE enabled I haven't seen spikes > 100us > > > in cyclictest. I did have mlock_all() enabled but nothing else. > > > PR_SET_THP_DISABLE remained unchanged (enabled). Is there anything to > > > stress this to be sure or is mlock_all() enough to do THP but leave the > > > mlock() applications alone? > > > > > > Then Mel continued with: > > > > > > | It's a similar hazard with NUMA balancing, an RT application should either > > > | disable balancing globally or set a memory policy that forces it to be > > > | ignored. They should be doing this anyway to avoid non-deterministic > > > | memory access costs due to NUMA artifacts but it wouldn't surprise me > > > | if some applications got it wrong. > > > > > > Usually (often) RT applications are pinned. I would assume that on > > > bigger box the RT tasks are at least pinned to a node. How bad can this > > > get in worst case? cyclictest pins every thread to CPU. I could remove > > > this for testing. What would be a good test to push this to its limit? > > > > > > Cc: Mel Gorman <mgorman@techsingularity.net> > > > Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> > > > > Somewhat tentative but > > > > Acked-by: Mel Gorman <mgorman@techsingularity.net> > > > > It's tentative because NUMA Balancing gets default disabled on PREEMPT_RT > > but it's still possible to enable where as THP is disabled entirely > > and can never be enabled. This is a little inconsistent and it would be > > preferable that they match either by disabling NUMA_BALANCING entirely or > > forbidding TRANSPARENT_HUGEPAGE_ALWAYS && PREEMPT_RT. I'm ok with either. > > Oh. I can go either way depending on the input ;) > > > There is the possibility that an RT application could use THP safely by > > using madvise() and mlock(). That way, THP is available but only if an > > application has explicit knowledge of THP and smart enough to do it only > > during the initialisation phase with > > Yes that was my question. So if you have "always", do mlock_all() in the > application and then have other threads that same application doing > malloc/ free of memory that the RT thread is not touching then bad > things can still happen, right? > My understanding is that all threads can be blocked in a page fault if > there is some THP operation going on. > Hmm, it could happen if all the memory used by the RT thread was not hugepage-aligned and potentially khugepaged could interfere. khugepaged can be disabled if tuned properly but the alignment requirement would be tricky. Probably safer to just disable it like it has been historically. For consistently, force NUMA_BALANCING to be disabled too because it introduces non-deterministic latencies even if memory regions are locked and bound. > > There is the slight caveat that even then THP can have inconsistent > > latencies if it has a split THP with separate entries for base and huge > > pages. The responsibility would be on the person deploying the application > > to ensure a platform was suitable for both RT and using huge pages. > > split THP? Sorry, "split TLB" where part of the TLB only handles base pages and another part handles huge pages.
On 2021-10-28 13:52:24 [+0100], Mel Gorman wrote: > > Yes that was my question. So if you have "always", do mlock_all() in the > > application and then have other threads that same application doing > > malloc/ free of memory that the RT thread is not touching then bad > > things can still happen, right? > > My understanding is that all threads can be blocked in a page fault if > > there is some THP operation going on. > > > > Hmm, it could happen if all the memory used by the RT thread was not > hugepage-aligned and potentially khugepaged could interfere. khugepaged > can be disabled if tuned properly but the alignment requirement would be > tricky. Probably safer to just disable it like it has been historically. > For consistently, force NUMA_BALANCING to be disabled too because it > introduces non-deterministic latencies even if memory regions are locked > and bound. Okay. I don't mind disabling it or keeping it enabled under some restrictions. I just need it to document it so people are aware why it is disabled so if they want to enable they know what the areas that need attention. THP disable due to alignment issues and potential defragmentation by khugepaged. Understood. Workaround: Use hugepages. NUMA_BALANCING. It looks like it replaces the physical page while keeping the virtual address. This kind of page migration does not look good if it happens for everyone since it involves mmap_lock. Let me write that up and post properly. Thank you. > > > There is the slight caveat that even then THP can have inconsistent > > > latencies if it has a split THP with separate entries for base and huge > > > pages. The responsibility would be on the person deploying the application > > > to ensure a platform was suitable for both RT and using huge pages. > > > > split THP? > > Sorry, "split TLB" where part of the TLB only handles base pages and > another part handles huge pages. ah okay. Sebastian
On Thu, Oct 28, 2021 at 03:56:47PM +0200, Sebastian Andrzej Siewior wrote: > On 2021-10-28 13:52:24 [+0100], Mel Gorman wrote: > > > Yes that was my question. So if you have "always", do mlock_all() in the > > > application and then have other threads that same application doing > > > malloc/ free of memory that the RT thread is not touching then bad > > > things can still happen, right? > > > My understanding is that all threads can be blocked in a page fault if > > > there is some THP operation going on. > > > > > > > Hmm, it could happen if all the memory used by the RT thread was not > > hugepage-aligned and potentially khugepaged could interfere. khugepaged > > can be disabled if tuned properly but the alignment requirement would be > > tricky. Probably safer to just disable it like it has been historically. > > For consistently, force NUMA_BALANCING to be disabled too because it > > introduces non-deterministic latencies even if memory regions are locked > > and bound. > > Okay. I don't mind disabling it or keeping it enabled under some > restrictions. I just need it to document it so people are aware why it > is disabled so if they want to enable they know what the areas that need > attention. > > THP disable due to alignment issues and potential defragmentation by > khugepaged. Understood. Workaround: Use hugepages. > > NUMA_BALANCING. It looks like it replaces the physical page while > keeping the virtual address. This kind of page migration does not look > good if it happens for everyone since it involves mmap_lock. > Let me write that up and post properly. > In case it helps; TRANSPARENT_HUGEPAGE: There are potential non-determinstic delays to an RT thread if a critical memory region is not THP-aligned and a non-RT buffer is located in the same hugepage-aligned region. It's also possible for an unrelated thread to migrate pages belonging to an RT task incurring unexpected page faults due to memory defragmentation even if khugepaged is disabled. NUMA_BALANCING: There is a non-determinstic delay to mark PTEs PROT_NONE to gather NUMA fault samples, increased page faults of regions even if mlocked and non-deterministic delays when migrating pages.
On 2021-10-28 15:14:52 [+0100], Mel Gorman wrote:
> In case it helps;
Yes. Thank you.
Sebastian
diff --git a/init/Kconfig b/init/Kconfig index edc0a0228f143..8e96817d507c3 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -922,7 +922,7 @@ config NUMA_BALANCING config NUMA_BALANCING_DEFAULT_ENABLED bool "Automatically enable NUMA aware memory/task placement" default y - depends on NUMA_BALANCING + depends on NUMA_BALANCING && !PREEMPT_RT help If set, automatic NUMA balancing will be enabled if running on a NUMA machine. diff --git a/mm/Kconfig b/mm/Kconfig index c150a0c6fce2c..5c5508fafcec5 100644 --- a/mm/Kconfig +++ b/mm/Kconfig @@ -374,7 +374,7 @@ config NOMMU_INITIAL_TRIM_EXCESS config TRANSPARENT_HUGEPAGE bool "Transparent Hugepage Support" - depends on HAVE_ARCH_TRANSPARENT_HUGEPAGE + depends on HAVE_ARCH_TRANSPARENT_HUGEPAGE && !PREEMPT_RT select COMPACTION select XARRAY_MULTI help
In https://lore.kernel.org/all/20200304091159.GN3818@techsingularity.net/ Mel wrote: | While I ack'd this, an RT application using THP is playing with fire, | I know the RT extension for SLE explicitly disables it from being enabled | at kernel config time. At minimum the critical regions should be mlocked | followed by prctl to disable future THP faults that are non-deterministic, | both from an allocation point of view, and a TLB access point of view. It's | still reasonable to expect a smaller TLB reach for huge pages than | base pages. With TRANSPARENT_HUGEPAGE enabled I haven't seen spikes > 100us in cyclictest. I did have mlock_all() enabled but nothing else. PR_SET_THP_DISABLE remained unchanged (enabled). Is there anything to stress this to be sure or is mlock_all() enough to do THP but leave the mlock() applications alone? Then Mel continued with: | It's a similar hazard with NUMA balancing, an RT application should either | disable balancing globally or set a memory policy that forces it to be | ignored. They should be doing this anyway to avoid non-deterministic | memory access costs due to NUMA artifacts but it wouldn't surprise me | if some applications got it wrong. Usually (often) RT applications are pinned. I would assume that on bigger box the RT tasks are at least pinned to a node. How bad can this get in worst case? cyclictest pins every thread to CPU. I could remove this for testing. What would be a good test to push this to its limit? Cc: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> --- init/Kconfig | 2 +- mm/Kconfig | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)