On Thu, Jul 22, 2010 at 12:06:34AM +0800, Minchan Kim wrote:
> On Tue, Jul 20, 2010 at 03:18:44PM +0800, Shaohua Li wrote:
> > The zone->lru_lock is heavily contented in workload where activate_page()
> > is frequently used. We could do batch activate_page() to reduce the lock
> > contention. The batched pages will be added into zone list when the pool
> > is full or page reclaim is trying to drain them.
> >
> > For example, in a 4 socket 64 CPU system, create a sparse file and 64 processes,
> > processes shared map to the file. Each process read access the whole file and
> > then exit. The process exit will do unmap_vmas() and cause a lot of
> > activate_page() call. In such workload, we saw about 58% total time reduction
> > with below patch.
>
> Great :)
> >
> > But we did see some strange regression. The regression is small (usually < 2%)
> > and most are from multithread test and none heavily use activate_page(). For
> > example, in the same system, we create 64 threads. Each thread creates a private
> > mmap region and does read access. We measure the total time and saw about 2%
> > regression. But in such workload, 99% time is on page fault and activate_page()
> > takes no time. Very strange, we haven't a good explanation for this so far,
> > hopefully somebody can share a hint.
>
> Mabye it might be due to lru_add_drain.
> You are adding cost in lru_add_drain and it is called several place.
> So if we can't get the gain in there, it could make a bit of regression.
> I might be wrong and it's a just my guessing.
The workload with regression doesn't invoke too many activate_page, so
basically activate_page_drain_cpu() is a nop, it should not take too much.
> > Signed-off-by: Shaohua Li <shaohua.li@xxxxxxxxx>
> >
> > diff --git a/mm/swap.c b/mm/swap.c
> > index 3ce7bc3..4a3fd7f 100644
> > --- a/mm/swap.c
> > +++ b/mm/swap.c
> > @@ -39,6 +39,7 @@ int page_cluster;
> >
> > static DEFINE_PER_CPU(struct pagevec[NR_LRU_LISTS], lru_add_pvecs);
> > static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
> > +static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
> >
> > /*
> > * This path almost never happens for VM activity - pages are normally
> > @@ -175,11 +176,10 @@ static void update_page_reclaim_stat(struct zone *zone, struct page *page,
> > /*
> > * FIXME: speed this up?
> > */
> Couldn't we remove above comment by this patch?
ha, yes.
> > -void activate_page(struct page *page)
> > +static void __activate_page(struct page *page)
> > {
> > struct zone *zone = page_zone(page);
> >
> > - spin_lock_irq(&zone->lru_lock);
> > if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
> > int file = page_is_file_cache(page);
> > int lru = page_lru_base_type(page);
> > @@ -192,7 +192,46 @@ void activate_page(struct page *page)
> >
> > update_page_reclaim_stat(zone, page, file, 1);
> > }
> > - spin_unlock_irq(&zone->lru_lock);
> > +}
> > +
> > +static void activate_page_drain_cpu(int cpu)
> > +{
> > + struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
> > + struct zone *last_zone = NULL, *zone;
> > + int i, j;
> > +
> > + for (i = 0; i < pagevec_count(pvec); i++) {
> > + zone = page_zone(pvec->pages[i]);
> > + if (zone == last_zone)
> > + continue;
> > +
> > + if (last_zone)
> > + spin_unlock_irq(&last_zone->lru_lock);
> > + last_zone = zone;
> > + spin_lock_irq(&last_zone->lru_lock);
> > +
> > + for (j = i; j < pagevec_count(pvec); j++) {
> > + struct page *page = pvec->pages[j];
> > +
> > + if (last_zone != page_zone(page))
> > + continue;
> > + __activate_page(page);
> > + }
> > + }
> > + if (last_zone)
> > + spin_unlock_irq(&last_zone->lru_lock);
> > + release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
> > + pagevec_reinit(pvec);
>
> In worst case(DMA->NORMAL->HIGHMEM->DMA->NORMA->HIGHMEM->......),
> overhead would is big than old. how about following as?
> static DEFINE_PER_CPU(struct pagevec[MAX_NR_ZONES], activate_page_pvecs);
> Is it a overkill?
activate_page_drain_cpu is a two level loop. In you case, the drain order
will be DMA->DMA->NORMAL->NORMAL->HIGHMEM->HIGHMEM. Since pagevec size is
14, the loop should finish quickly.
Thanks,
Shaohua
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>