lkml.org 
[lkml]   [2013]   [May]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH v2 1/2] Make the batch size of the percpu_counter configurable
On Tue, 21 May 2013 16:27:29 -0700 Tim Chen <tim.c.chen@linux.intel.com> wrote:

> Will something like the following work if we get rid of the percpu
> counter changes and use __percpu_counter_add(..., batch)? In
> benchmark with a lot of memory changes via brk, this makes quite
> a difference when we go to a bigger batch size.

That looks pretty close.

> Tim
>
> Change batch size for memory accounting to be proportional to memory available.
>
> Currently the per cpu counter's batch size for memory accounting is
> configured as twice the number of cpus in the system. However,
> for system with very large memory, it is more appropriate to make it
> proportional to the memory size per cpu in the system.
>
> For example, for a x86_64 system with 64 cpus and 128 GB of memory,
> the batch size is only 2*64 pages (0.5 MB). So any memory accounting
> changes of more than 0.5MB will overflow the per cpu counter into
> the global counter. Instead, for the new scheme, the batch size
> is configured to be 0.4% of the memory/cpu = 8MB (128 GB/64 /256),
> which is more inline with the memory size.
>
> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
> ---
> include/linux/mman.h | 5 +++++
> mm/mmap.c | 14 ++++++++++++++
> mm/nommu.c | 14 ++++++++++++++
> 3 files changed, 33 insertions(+)
>
> diff --git a/include/linux/mman.h b/include/linux/mman.h
> index 9aa863d..11d5ce9 100644
> --- a/include/linux/mman.h
> +++ b/include/linux/mman.h
> @@ -10,12 +10,17 @@
> extern int sysctl_overcommit_memory;
> extern int sysctl_overcommit_ratio;
> extern struct percpu_counter vm_committed_as;
> +extern int vm_committed_as_batch;
>
> unsigned long vm_memory_committed(void);
>
> static inline void vm_acct_memory(long pages)
> {
> +#ifdef CONFIG_SMP
> + __percpu_counter_add(&vm_committed_as, pages, vm_committed_as_batch);
> +#else
> percpu_counter_add(&vm_committed_as, pages);
> +#endif
> }

I think we could use __percpu_counter_add() unconditionally here and
just do

#ifdef CONFIG_SMP
#define vm_committed_as_batch 0
#else
int vm_committed_as_batch;
#endif

The EXPORT_SYMBOL(vm_committed_as_batch) is unneeded.

> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -3145,11 +3145,25 @@ void mm_drop_all_locks(struct mm_struct *mm)
> /*
> * initialise the VMA slab
> */
> +
> +int vm_committed_as_batch;
> +EXPORT_SYMBOL(vm_committed_as_batch);
> +
> +static int mm_compute_batch(void)
> +{
> + int nr = num_present_cpus();
> + int batch = max(32, nr*2);
> +
> + /* batch size set to 0.4% of (total memory/#cpus) */
> + return max((int) (totalram_pages/nr) / 256, batch);
> +}

Change this to do the assignment to vm_committed_as_batch then put this
code inside #ifdef CONFIG_SMP and do

#else /* CONFIG_SMP */
static inline void mm_compute_batch(void)
{
}
#endif

> void __init mmap_init(void)
> {
> int ret;
>
> ret = percpu_counter_init(&vm_committed_as, 0);
> + vm_committed_as_batch = mm_compute_batch();

This becomes just

mm_compute_batch();

> VM_BUG_ON(ret);
> }
>
> diff --git a/mm/nommu.c b/mm/nommu.c
> index 298884d..1b7008a 100644
> --- a/mm/nommu.c
> +++ b/mm/nommu.c
> @@ -527,11 +527,25 @@ SYSCALL_DEFINE1(brk, unsigned long, brk)
> /*
> * initialise the VMA and region record slabs
> */
> +
> +int vm_committed_as_batch;
> +EXPORT_SYMBOL(vm_committed_as_batch);
> +
> +static int mm_compute_batch(void)
> +{
> + int nr = num_present_cpus();
> + int batch = max(32, nr*2);
> +
> + /* batch size set to 0.4% of (total memory/#cpus) */
> + return max((int) (totalram_pages/nr) / 256, batch);
> +}
> +
> void __init mmap_init(void)
> {
> int ret;
>
> ret = percpu_counter_init(&vm_committed_as, 0);
> + vm_committed_as_batch = mm_compute_batch();
> VM_BUG_ON(ret);
> vm_region_jar = KMEM_CACHE(vm_region, SLAB_PANIC);

I'm not sure that CONFIG_MMU=n && CONFIG_SMP=y even exists. Perhaps it
does. But there's no point in ruling out that option here.

The nommu code becomes identical to the mmu code so we should put it in
a shared file. I suppose mmap.c would be as good a place as any.

We could make mm_compute_batch() __init and call it from mm_init().
But really it should be __meminit and there should be a memory-hotplug
notifier handler which adjusts vm_committed_as_batch's value.




\
 
 \ /
  Last update: 2013-05-22 02:01    [W:0.082 / U:0.876 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site