v2 mm: balance_dirty_pages. reduce calls to global_page_state to reduce cache references

Reducing the number of times balance_dirty_pages calls global_page_state reduces the cache references and so improves write performance on a variety of workloads.

'perf stats' of simple fio write tests shows the reduction in cache access.
Where the test is fio 'write,mmap,600Mb,pre_read' on AMD AthlonX2 with 3Gb memory (dirty_threshold approx 600 Mb)
running each test 10 times, dropping the fasted & slowest values then taking
the average & standard deviation

		average (s.d.) in millions (10^6)

2.6.31-rc8 648.6 (14.6)
+patch 620.1 (16.5)

Achieving this reduction is by dropping clip_bdi_dirty_limit as it rereads the counters to apply the dirty_threshold and moving this check up into balance_dirty_pages where it has already read the counters.

Also by rearrange the for loop to only contain one copy of the limit tests allows the pdflush test after the loop to use the local copies of the counters rather than rereading them.

In the common case with no throttling it now calls global_page_state 5 fewer times and bdi_stat 2 fewer.

This version includes the changes suggested by
Wu Fengguang <[email protected]>

Signed-off-by Richard Kennedy <[email protected]>

----
Thanks to everybody for the feedback & suggestions. This patch is against 2.6.31-rc8

mm/page-writeback.c +36 -63

index 81627eb..9581359 100644
@@ -260,32 +260,6 @@ static void bdi_writeout_fraction(struct backing_dev_info *bdi,
 	}
 }
 
-/*
- * Clip the earned share of dirty pages to that which is actually available.
- * This avoids exceeding the total dirty_limit when the floating averages
- * fluctuate too quickly.
- */
-static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
-		unsigned long dirty, unsigned long *pbdi_dirty)
-{
-	unsigned long avail_dirty;
-
-	avail_dirty = global_page_state(NR_FILE_DIRTY) +
-		 global_page_state(NR_WRITEBACK) +
-		 global_page_state(NR_UNSTABLE_NFS) +
-		 global_page_state(NR_WRITEBACK_TEMP);
-
-	if (avail_dirty < dirty)
-		avail_dirty = dirty - avail_dirty;
-	else
-		avail_dirty = 0;
-
-	avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
-		bdi_stat(bdi, BDI_WRITEBACK);
-
-	*pbdi_dirty = min(*pbdi_dirty, avail_dirty);
-}
-
 static inline void task_dirties_fraction(struct task_struct *tsk,
 		long *numerator, long *denominator)
 {

@@ -478,7 +452,6 @@ get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
 			bdi_dirty = dirty * bdi->max_ratio / 100;
 
 		*pbdi_dirty = bdi_dirty;
-		clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
 		task_dirty_limit(current, pbdi_dirty);
 	}
 }

@@ -499,7 +472,7 @@ static void balance_dirty_pages(struct address_space *mapping)
 	unsigned long bdi_thresh;
 	unsigned long pages_written = 0;
 	unsigned long write_chunk = sync_writeback_pages();
-
+	int dirty_exceeded;
 	struct backing_dev_info *bdi = mapping->backing_dev_info;
 
 	for (;;) {

@@ -512,16 +485,36 @@ static void balance_dirty_pages(struct address_space *mapping)
 		};
 
 		get_dirty_limits(&background_thresh, &dirty_thresh,
-				&bdi_thresh, bdi);
+				 &bdi_thresh, bdi);
 
 		nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
-					global_page_state(NR_UNSTABLE_NFS);
-		nr_writeback = global_page_state(NR_WRITEBACK);
+			global_page_state(NR_UNSTABLE_NFS);
+		nr_writeback = global_page_state(NR_WRITEBACK) +
+			global_page_state(NR_WRITEBACK_TEMP);
 
-		bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
-		bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
+		/*
+		 * In order to avoid the stacked BDI deadlock we need
+		 * to ensure we accurately count the 'dirty' pages when
+		 * the threshold is low.
+		 *
+		 * Otherwise it would be possible to get thresh+n pages
+		 * reported dirty, even though there are thresh-m pages
+		 * actually dirty; with m+n sitting in the percpu
+		 * deltas.
+		 */
+		if (bdi_thresh < 2*bdi_stat_error(bdi)) {
+			bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
+			bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
+		} else if (bdi_nr_reclaimable) {
+			bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
+			bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
+		}
 
-		if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
+		dirty_exceeded =
+			(bdi_nr_reclaimable + bdi_nr_writeback >= bdi_thresh)
+			|| (nr_reclaimable + nr_writeback >= dirty_thresh);
+
+		if (!dirty_exceeded)
 			break;
 
 		/*

@@ -530,7 +523,11 @@ static void balance_dirty_pages(struct address_space *mapping)
 		 * when the bdi limits are ramping up.
 		 */
 		if (nr_reclaimable + nr_writeback <
-				(background_thresh + dirty_thresh) / 2)
+		    (background_thresh + dirty_thresh) / 2)
+			break;
+
+		/* done enough? */
+		if (pages_written >= write_chunk)
 			break;
 
 		if (!bdi->dirty_exceeded)

@@ -548,38 +545,14 @@ static void balance_dirty_pages(struct address_space *mapping)
 		if (bdi_nr_reclaimable > bdi_thresh) {
 			writeback_inodes(&wbc);
 			pages_written += write_chunk - wbc.nr_to_write;
-			get_dirty_limits(&background_thresh, &dirty_thresh,
-				       &bdi_thresh, bdi);
-		}
-
-		/*
-		 * In order to avoid the stacked BDI deadlock we need
-		 * to ensure we accurately count the 'dirty' pages when
-		 * the threshold is low.
-		 *
-		 * Otherwise it would be possible to get thresh+n pages
-		 * reported dirty, even though there are thresh-m pages
-		 * actually dirty; with m+n sitting in the percpu
-		 * deltas.
-		 */
-		if (bdi_thresh < 2*bdi_stat_error(bdi)) {
-			bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
-			bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
-		} else if (bdi_nr_reclaimable) {
-			bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
-			bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
+			/* don't wait if we've done enough */
+			if (pages_written >= write_chunk)
+				continue;
 		}
-
-		if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
-			break;
-		if (pages_written >= write_chunk)
-			break;		/* We've done our duty */
-
 		congestion_wait(BLK_RW_ASYNC, HZ/10);
 	}
 
-	if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
-			bdi->dirty_exceeded)
+	if (!dirty_exceeded && bdi->dirty_exceeded)
 		bdi->dirty_exceeded = 0;
 
 	if (writeback_in_progress(bdi))

@@ -593,10 +566,8 @@ static void balance_dirty_pages(struct address_space *mapping)
 	 * In normal mode, we start background writeout at the lower
 	 * background_thresh, to keep the amount of dirty memory low.

Quoted 24 lines · Richard Kennedy

Reducing the number of times balance_dirty_pages calls global_page_state reduces the cache references and so improves write performance on a variety of workloads.

'perf stats' of simple fio write tests shows the reduction in cache access.
Where the test is fio 'write,mmap,600Mb,pre_read' on AMD AthlonX2 with 3Gb memory (dirty_threshold approx 600 Mb)
running each test 10 times, dropping the fasted & slowest values then taking
the average & standard deviation

		average (s.d.) in millions (10^6)

2.6.31-rc8 648.6 (14.6)
+patch 620.1 (16.5)

Achieving this reduction is by dropping clip_bdi_dirty_limit as it rereads the counters to apply the dirty_threshold and moving this check up into balance_dirty_pages where it has already read the counters.

Also by rearrange the for loop to only contain one copy of the limit tests allows the pdflush test after the loop to use the local copies of the counters rather than rereading them.

Quoted 5 lines

In the common case with no throttling it now calls global_page_state 5 fewer times and bdi_stat 2 fewer.

This version includes the changes suggested by
Wu Fengguang <[email protected]>

It seems that an redundant pages_written test can be reduced by

mm/page-writeback.c +1 -5

@@ -526,10 +526,6 @@ static void balance_dirty_pages(struct a
 		    (background_thresh + dirty_thresh) / 2)
 			break;
 
-		/* done enough? */
-		if (pages_written >= write_chunk)
-			break;
-
 		if (!bdi->dirty_exceeded)
 			bdi->dirty_exceeded = 1;
 

@@ -547,7 +543,7 @@ static void balance_dirty_pages(struct a
 			pages_written += write_chunk - wbc.nr_to_write;
 			/* don't wait if we've done enough */
 			if (pages_written >= write_chunk)
-				continue;
+				break;
 		}
 		congestion_wait(BLK_RW_ASYNC, HZ/10);
 	}

Otherwise the patch looks good to me. Thank you for the nice work!

Reviewed-by Wu Fengguang <[email protected]>

Quoted 56 lines · Wu Fengguang

Quoted 24 lines · Richard Kennedy

Reducing the number of times balance_dirty_pages calls global_page_state reduces the cache references and so improves write performance on a variety of workloads.

'perf stats' of simple fio write tests shows the reduction in cache access.
Where the test is fio 'write,mmap,600Mb,pre_read' on AMD AthlonX2 with 3Gb memory (dirty_threshold approx 600 Mb)
running each test 10 times, dropping the fasted & slowest values then taking
the average & standard deviation

		average (s.d.) in millions (10^6)

2.6.31-rc8 648.6 (14.6)
+patch 620.1 (16.5)

Achieving this reduction is by dropping clip_bdi_dirty_limit as it rereads the counters to apply the dirty_threshold and moving this check up into balance_dirty_pages where it has already read the counters.

Also by rearrange the for loop to only contain one copy of the limit tests allows the pdflush test after the loop to use the local copies of the counters rather than rereading them.

Quoted 5 lines

In the common case with no throttling it now calls global_page_state 5 fewer times and bdi_stat 2 fewer.

This version includes the changes suggested by
Wu Fengguang <[email protected]>

It seems that an redundant pages_written test can be reduced by

mm/page-writeback.c +1 -5

@@ -526,10 +526,6 @@ static void balance_dirty_pages(struct a
 		    (background_thresh + dirty_thresh) / 2)
 			break;
 
-		/* done enough? */
-		if (pages_written >= write_chunk)
-			break;
-
 		if (!bdi->dirty_exceeded)
 			bdi->dirty_exceeded = 1;
 

@@ -547,7 +543,7 @@ static void balance_dirty_pages(struct a
 			pages_written += write_chunk - wbc.nr_to_write;
 			/* don't wait if we've done enough */
 			if (pages_written >= write_chunk)
-				continue;
+				break;
 		}
 		congestion_wait(BLK_RW_ASYNC, HZ/10);
 	}

Quoted 4 lines

Otherwise the patch looks good to me. Thank you for the nice work!

Reviewed-by Wu Fengguang <[email protected]>

Thank you.

I'll give your suggestion a try & run some tests. I think you're right it should be better. Not re-reading the global counters again should be of some benefit!
regards
Richard

Reducing the number of times balance_dirty_pages calls global_page_state reduces the cache references and so improves write performance on a variety of workloads.

'perf stats' of simple fio write tests shows the reduction in cache access.
Where the test is fio 'write,mmap,600Mb,pre_read' on AMD AthlonX2 with 3Gb memory (dirty_threshold approx 600 Mb)
running each test 10 times, dropping the fasted & slowest values then taking
the average & standard deviation

		average (s.d.) in millions (10^6)

2.6.31-rc8 648.6 (14.6)
+patch 620.1 (16.5)

Quoted 12 lines · Richard Kennedy

Reducing the number of times balance_dirty_pages calls global_page_state reduces the cache references and so improves write performance on a variety of workloads.

'perf stats' of simple fio write tests shows the reduction in cache access.
Where the test is fio 'write,mmap,600Mb,pre_read' on AMD AthlonX2 with 3Gb memory (dirty_threshold approx 600 Mb)
running each test 10 times, dropping the fasted & slowest values then taking
the average & standard deviation

Quoted 3 lines

		average (s.d.) in millions (10^6)

2.6.31-rc8 648.6 (14.6)
+patch 620.1 (16.5)

This patch looks good to me, I have workloads too here where up to 10% of the time is spent in balance_dirty_pages() because of this. I'll give this patch a go on the box and test in question tomorrow, but it looks promising.

Quoted 13 lines · Jens Axboe

Quoted 12 lines · Richard Kennedy

Reducing the number of times balance_dirty_pages calls global_page_state reduces the cache references and so improves write performance on a variety of workloads.

'perf stats' of simple fio write tests shows the reduction in cache access.
Where the test is fio 'write,mmap,600Mb,pre_read' on AMD AthlonX2 with 3Gb memory (dirty_threshold approx 600 Mb)
running each test 10 times, dropping the fasted & slowest values then taking
the average & standard deviation

Quoted 9 lines

Quoted 3 lines

		average (s.d.) in millions (10^6)

2.6.31-rc8 648.6 (14.6)
+patch 620.1 (16.5)

This patch looks good to me, I have workloads too here where up to 10% of the time is spent in balance_dirty_pages() because of this. I'll give this patch a go on the box and test in question tomorrow, but it looks promising.

Thanks Jens,

It will be interesting to see how it works on different hardware & workload. How many cores are you going to run it on?
wow 10% in balance_dirty_pages! Is that on a large server? or do you think its peculiar to your workload?