| Message ID | 20240816091327.54183-1-sunnanyong@huawei.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [RFC] mm: control mthp per process/cgroup | expand |
On Fri, Aug 16, 2024 at 05:13:27PM +0800, Nanyong Sun wrote: > Now the large folio control interfaces is system wide and tend to be > default on: file systems use large folio by default if supported, > mTHP is tend to default enable when boot [1]. > When large folio enabled, some workloads have performance benefit, > but some may not and some side effects can happen: the memory usage > may increase, direct reclaim maybe more frequently because of more > large order allocations, result in cpu usage also increases. We observed > this on a product environment which run nginx, the pgscan_direct count > increased a lot than before, can reach to 3000 times per second, and > disable file large folio can fix this. Can you share any details of your nginx workload that shows a regression? The heuristics for allocating large folios are completely untuned, so having data for a workload which performs better with small folios is very valuable.
On 2024/8/17 2:15, Matthew Wilcox wrote: > On Fri, Aug 16, 2024 at 05:13:27PM +0800, Nanyong Sun wrote: >> Now the large folio control interfaces is system wide and tend to be >> default on: file systems use large folio by default if supported, >> mTHP is tend to default enable when boot [1]. >> When large folio enabled, some workloads have performance benefit, >> but some may not and some side effects can happen: the memory usage >> may increase, direct reclaim maybe more frequently because of more >> large order allocations, result in cpu usage also increases. We observed >> this on a product environment which run nginx, the pgscan_direct count >> increased a lot than before, can reach to 3000 times per second, and >> disable file large folio can fix this. > Can you share any details of your nginx workload that shows a regression? > The heuristics for allocating large folios are completely untuned, so > having data for a workload which performs better with small folios is > very valuable. > > . The RPS(/Requests per second/) which is the performance metric of nginx workload has no regression(also no improvement)，we just observed that pgscan_direct rate is much higher with large folio. So far, we have tested some workloads' benchmark, some did not have performance improvement but also did not have regression. In a production environment, different workloads may be deployed on a machine. Therefore, do we need to add a process/cgroup level control to prevent workloads that will not have performance improvement from using mTHP? In this way, the memory overhead and direct reclaim caused by mTHP can be avoided for those process/cgroup.
On 2024/8/19 13:58, Nanyong Sun wrote: > On 2024/8/17 2:15, Matthew Wilcox wrote: >> On Fri, Aug 16, 2024 at 05:13:27PM +0800, Nanyong Sun wrote: >>> Now the large folio control interfaces is system wide and tend to be >>> default on: file systems use large folio by default if supported, >>> mTHP is tend to default enable when boot [1]. >>> When large folio enabled, some workloads have performance benefit, >>> but some may not and some side effects can happen: the memory usage >>> may increase, direct reclaim maybe more frequently because of more >>> large order allocations, result in cpu usage also increases. We observed >>> this on a product environment which run nginx, the pgscan_direct count >>> increased a lot than before, can reach to 3000 times per second, and >>> disable file large folio can fix this. >> Can you share any details of your nginx workload that shows a regression? >> The heuristics for allocating large folios are completely untuned, so >> having data for a workload which performs better with small folios is >> very valuable. >> >> . > The RPS//(Requests per second) which is the performance metric of > nginx workload has no > regression(also no improvement)，we just observed that pgscan_direct > rate is much higher > with large folio. > So far, we have tested some workloads' benchmark, some did not have > performance improvement > but also did not have regression. > In a production environment, different workloads may be deployed on a > machine. Therefore, > do we need to add a process/cgroup level control to prevent workloads > that will not have > performance improvement from using mTHP? In this way, the memory > overhead and direct reclaim > caused by mTHP can be avoided for those process/cgroup. Sorry to disturb, just a friendly ping : )
On 2024/8/19 13:58, Nanyong Sun wrote: > On 2024/8/17 2:15, Matthew Wilcox wrote: > >> On Fri, Aug 16, 2024 at 05:13:27PM +0800, Nanyong Sun wrote: >>> Now the large folio control interfaces is system wide and tend to be >>> default on: file systems use large folio by default if supported, >>> mTHP is tend to default enable when boot [1]. >>> When large folio enabled, some workloads have performance benefit, >>> but some may not and some side effects can happen: the memory usage >>> may increase, direct reclaim maybe more frequently because of more >>> large order allocations, result in cpu usage also increases. We observed >>> this on a product environment which run nginx, the pgscan_direct count >>> increased a lot than before, can reach to 3000 times per second, and >>> disable file large folio can fix this. >> Can you share any details of your nginx workload that shows a regression? >> The heuristics for allocating large folios are completely untuned, so >> having data for a workload which performs better with small folios is >> very valuable. >> >> . > The RPS(/Requests per second/) which is the performance metric of nginx > workload has no > regression(also no improvement)，we just observed that pgscan_direct > rate is much higher > with large folio. > So far, we have tested some workloads' benchmark, some did not have > performance improvement > but also did not have regression. > In a production environment, different workloads may be deployed on a > machine. Therefore, > do we need to add a process/cgroup level control to prevent workloads > that will not have > performance improvement from using mTHP? In this way, the memory > overhead and direct reclaim > caused by mTHP can be avoided for those process/cgroup. OK. So no regression with mTHP, seems just some theoretical analysis. IMHO, it would be better to evaluate your 'per-cgroup mTHP control' idea on some real workloads, and gather some data to evaluation, which can be more convincing. Just my 2 cents:)
On 02.09.24 11:36, Baolin Wang wrote: > > > On 2024/8/19 13:58, Nanyong Sun wrote: >> On 2024/8/17 2:15, Matthew Wilcox wrote: >> >>> On Fri, Aug 16, 2024 at 05:13:27PM +0800, Nanyong Sun wrote: >>>> Now the large folio control interfaces is system wide and tend to be >>>> default on: file systems use large folio by default if supported, >>>> mTHP is tend to default enable when boot [1]. >>>> When large folio enabled, some workloads have performance benefit, >>>> but some may not and some side effects can happen: the memory usage >>>> may increase, direct reclaim maybe more frequently because of more >>>> large order allocations, result in cpu usage also increases. We observed >>>> this on a product environment which run nginx, the pgscan_direct count >>>> increased a lot than before, can reach to 3000 times per second, and >>>> disable file large folio can fix this. >>> Can you share any details of your nginx workload that shows a regression? >>> The heuristics for allocating large folios are completely untuned, so >>> having data for a workload which performs better with small folios is >>> very valuable. >>> >>> . >> The RPS(/Requests per second/) which is the performance metric of nginx >> workload has no >> regression(also no improvement)，we just observed that pgscan_direct >> rate is much higher >> with large folio. >> So far, we have tested some workloads' benchmark, some did not have >> performance improvement >> but also did not have regression. >> In a production environment, different workloads may be deployed on a >> machine. Therefore, >> do we need to add a process/cgroup level control to prevent workloads >> that will not have >> performance improvement from using mTHP? In this way, the memory >> overhead and direct reclaim >> caused by mTHP can be avoided for those process/cgroup. > > OK. So no regression with mTHP, seems just some theoretical analysis. > > IMHO, it would be better to evaluate your 'per-cgroup mTHP control' idea > on some real workloads, and gather some data to evaluation, which can be > more convincing. Agreed!
diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h index e25d9ebfdf89..8c0b62b732b7 100644 --- a/include/linux/huge_mm.h +++ b/include/linux/huge_mm.h @@ -247,6 +247,9 @@ unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma, if ((tva_flags & TVA_ENFORCE_SYSFS) && vma_is_anonymous(vma)) { unsigned long mask = READ_ONCE(huge_anon_orders_always); + if (test_bit(MMF_DISABLE_ANON_MTHP, &vma->vm_mm->flags)) + return 0; + if (vm_flags & VM_HUGEPAGE) mask |= READ_ONCE(huge_anon_orders_madvise); if (hugepage_global_always() || diff --git a/include/linux/sched/coredump.h b/include/linux/sched/coredump.h index e62ff805cfc9..0935b4790e6f 100644 --- a/include/linux/sched/coredump.h +++ b/include/linux/sched/coredump.h @@ -56,6 +56,10 @@ static inline int get_dumpable(struct mm_struct *mm) # define MMF_DUMP_MASK_DEFAULT_ELF 0 #endif /* leave room for more dump flags */ +#define MMF_DISABLE_ANON_MTHP 13 +#define MMF_DISABLE_SHMEM_MTHP 14 +#define MMF_DISABLE_FILE_MTHP 15 +#define MMF_DISABLE_MTHP_MASK (7 << MMF_DISABLE_ANON_MTHP) #define MMF_VM_MERGEABLE 16 /* KSM may merge identical pages */ #define MMF_VM_HUGEPAGE 17 /* set when mm is available for khugepaged */ @@ -96,8 +100,9 @@ static inline int get_dumpable(struct mm_struct *mm) #define MMF_TOPDOWN_MASK (1 << MMF_TOPDOWN) #define MMF_INIT_MASK (MMF_DUMPABLE_MASK | MMF_DUMP_FILTER_MASK |\ - MMF_DISABLE_THP_MASK | MMF_HAS_MDWE_MASK |\ - MMF_VM_MERGE_ANY_MASK | MMF_TOPDOWN_MASK) + MMF_DISABLE_THP_MASK | MMF_DISABLE_MTHP_MASK |\ + MMF_HAS_MDWE_MASK | MMF_VM_MERGE_ANY_MASK |\ + MMF_TOPDOWN_MASK) static inline unsigned long mmf_init_flags(unsigned long flags) { diff --git a/include/uapi/linux/prctl.h b/include/uapi/linux/prctl.h index 35791791a879..584ac45f4ec8 100644 --- a/include/uapi/linux/prctl.h +++ b/include/uapi/linux/prctl.h @@ -178,6 +178,11 @@ struct prctl_mm_map { #define PR_GET_TID_ADDRESS 40 #define PR_SET_THP_DISABLE 41 +# define PR_DISABLE_ANON_MTHP (1UL << 1) +# define PR_DISABLE_SHMEM_MTHP (1UL << 2) +# define PR_DISABLE_FILE_MTHP (1UL << 3) +# define DISABLE_MTHP_ALL_MASK (PR_DISABLE_ANON_MTHP | PR_DISABLE_SHMEM_MTHP |\ + PR_DISABLE_FILE_MTHP) #define PR_GET_THP_DISABLE 42 /* diff --git a/kernel/sys.c b/kernel/sys.c index 3a2df1bd9f64..06f2b1de46a7 100644 --- a/kernel/sys.c +++ b/kernel/sys.c @@ -2627,17 +2627,41 @@ SYSCALL_DEFINE5(prctl, int, option, unsigned long, arg2, unsigned long, arg3, case PR_GET_THP_DISABLE: if (arg2 || arg3 || arg4 || arg5) return -EINVAL; - error = !!test_bit(MMF_DISABLE_THP, &me->mm->flags); + if (test_bit(MMF_DISABLE_THP, &me->mm->flags)) + error = 1; + if (test_bit(MMF_DISABLE_ANON_MTHP, &me->mm->flags)) + error |= PR_DISABLE_ANON_MTHP; + if (test_bit(MMF_DISABLE_SHMEM_MTHP, &me->mm->flags)) + error |= PR_DISABLE_SHMEM_MTHP; + if (test_bit(MMF_DISABLE_FILE_MTHP, &me->mm->flags)) + error |= PR_DISABLE_FILE_MTHP; break; case PR_SET_THP_DISABLE: - if (arg3 || arg4 || arg5) + if (arg4 || arg5) + return -EINVAL; + if (arg3 && (arg3 & ~DISABLE_MTHP_ALL_MASK)) return -EINVAL; if (mmap_write_lock_killable(me->mm)) return -EINTR; - if (arg2) - set_bit(MMF_DISABLE_THP, &me->mm->flags); - else - clear_bit(MMF_DISABLE_THP, &me->mm->flags); + if (arg2) { + if (!arg3) + set_bit(MMF_DISABLE_THP, &me->mm->flags); + if (arg3 & PR_DISABLE_ANON_MTHP) + set_bit(MMF_DISABLE_ANON_MTHP, &me->mm->flags); + if (arg3 & PR_DISABLE_SHMEM_MTHP) + set_bit(MMF_DISABLE_SHMEM_MTHP, &me->mm->flags); + if (arg3 & PR_DISABLE_FILE_MTHP) + set_bit(MMF_DISABLE_FILE_MTHP, &me->mm->flags); + } else { + if (!arg3) + clear_bit(MMF_DISABLE_THP, &me->mm->flags); + if (arg3 & PR_DISABLE_ANON_MTHP) + clear_bit(MMF_DISABLE_ANON_MTHP, &me->mm->flags); + if (arg3 & PR_DISABLE_SHMEM_MTHP) + clear_bit(MMF_DISABLE_SHMEM_MTHP, &me->mm->flags); + if (arg3 & PR_DISABLE_FILE_MTHP) + clear_bit(MMF_DISABLE_FILE_MTHP, &me->mm->flags); + } mmap_write_unlock(me->mm); break; case PR_MPX_ENABLE_MANAGEMENT: diff --git a/mm/shmem.c b/mm/shmem.c index 5a77acf6ac6a..f4272883df77 100644 --- a/mm/shmem.c +++ b/mm/shmem.c @@ -556,7 +556,9 @@ static bool __shmem_is_huge(struct inode *inode, pgoff_t index, if (!S_ISREG(inode->i_mode)) return false; - if (mm && ((vm_flags & VM_NOHUGEPAGE) || test_bit(MMF_DISABLE_THP, &mm->flags))) + if (mm && ((vm_flags & VM_NOHUGEPAGE) || + test_bit(MMF_DISABLE_THP, &mm->flags) || + test_bit(MMF_DISABLE_SHMEM_MTHP, &mm->flags))) return false; if (shmem_huge == SHMEM_HUGE_DENY) return false; @@ -1633,7 +1635,8 @@ unsigned long shmem_allowable_huge_orders(struct inode *inode, int order; if ((vm_flags & VM_NOHUGEPAGE) || - test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)) + test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags) || + test_bit(MMF_DISABLE_SHMEM_MTHP, &vma->vm_mm->flags)) return 0; /* If the hardware/firmware marked hugepage support disabled. */
Now the large folio control interfaces is system wide and tend to be default on: file systems use large folio by default if supported, mTHP is tend to default enable when boot [1]. When large folio enabled, some workloads have performance benefit, but some may not and some side effects can happen: the memory usage may increase, direct reclaim maybe more frequently because of more large order allocations, result in cpu usage also increases. We observed this on a product environment which run nginx, the pgscan_direct count increased a lot than before, can reach to 3000 times per second, and disable file large folio can fix this. Now the anon/shmem/file mthp control interfaces is system wide, so the control api per process or cgroup may be necessary, for example, in one machine, some containers can use large folios and some can disable them. This patch present a possible solution: Extend the existing prctl api(PR_SET_THP_DISABLE), use the third argument to specify which kind of mTHP to disable for this process. For example: prctl(PR_SET_THP_DISABLE, 1, 0, 0, 0); //keep the original semantics prctl(PR_SET_THP_DISABLE, 1, PR_DISABLE_ANON_MTHP, 0, 0); prctl(PR_SET_THP_DISABLE, 1, PR_DISABLE_SHMEM_MTHP, 0, 0); prctl(PR_SET_THP_DISABLE, 1, PR_DISABLE_FILE_MTHP, 0, 0); The child processes will inherit the setting so if a seed process had configured, all processes in the cgroup will inherit the setting. This patch does not implement control over file mTHP, this is planed to do after pagecache folio sizes control done[2]. [1] https://lore.kernel.org/linux-mm/20240814085409.124466-1-21cnbao@gmail.com/T/ [2] https://lore.kernel.org/lkml/20240717071257.4141363-1-ryan.roberts@arm.com/T/ Signed-off-by: Nanyong Sun <sunnanyong@huawei.com> --- This patch is more to discuss the possible directions. I am not sure occupy 3 bits in the mm flag hole is a good way, do we need add a new flag in mm_struct to do this? I think another possible solution(not in this patch) is to go in the opposite direction of this patch, i.e. only allow the processes/cgroups who really need large folio to use the mTHP. We can add a new prctl api to mark the process who can use some specific sizes of mTHP. TODO: Need modify reference manual for the change of prctl PR_SET_THP_DISABLE. include/linux/huge_mm.h | 3 +++ include/linux/sched/coredump.h | 9 +++++++-- include/uapi/linux/prctl.h | 5 +++++ kernel/sys.c | 36 ++++++++++++++++++++++++++++------ mm/shmem.c | 7 +++++-- 5 files changed, 50 insertions(+), 10 deletions(-)