ext4: Fix performance regression with mballoc

One of the side-effects of mb_optimize_scan was that the optimized functions to select next group to try were called even before we tried the goal group. As a result we no longer allocate files close to corresponding inodes as well as we don't try to expand currently allocated extent in the same group. This results in reaim regression with workfile.disk workload of upto 8% with many clients on my test machine:

baseline mb_optimize_scan

Hmean     disk-1       2114.16 (   0.00%)     2099.37 (  -0.70%)
Hmean     disk-41     87794.43 (   0.00%)    83787.47 *  -4.56%*
Hmean     disk-81    148170.73 (   0.00%)   135527.05 *  -8.53%*
Hmean     disk-121   177506.11 (   0.00%)   166284.93 *  -6.32%*
Hmean     disk-161   220951.51 (   0.00%)   207563.39 *  -6.06%*
Hmean     disk-201   208722.74 (   0.00%)   203235.59 (  -2.63%)
Hmean     disk-241   222051.60 (   0.00%)   217705.51 (  -1.96%)
Hmean     disk-281   252244.17 (   0.00%)   241132.72 *  -4.41%*
Hmean     disk-321   255844.84 (   0.00%)   245412.84 *  -4.08%*

Also this is causing huge regression (time increased by a factor of 5 or so) when untarring archive with lots of small files on some eMMC storage cards.

Fix the problem by making sure we try goal group first.

mb_set_largest_free_order() updates lists containing groups with largest chunk of free space of given order. The way it updates it leads to always moving the group to the tail of the list. Thus allocations looking for free space of given order effectively end up cycling through all groups (and due to initialization in last to first order). This spreads allocations among block groups which reduces performance for rotating disks or low-end flash media. Change mb_set_largest_free_order() to only update lists if the order of the largest free chunk in the group changed.

Quoted 9 lines · Jan Kara

mb_set_largest_free_order() updates lists containing groups with largest chunk of free space of given order. The way it updates it leads to always moving the group to the tail of the list. Thus allocations looking for free space of given order effectively end up cycling through all groups (and due to initialization in last to first order). This spreads allocations among block groups which reduces performance for rotating disks or low-end flash media. Change mb_set_largest_free_order() to only update lists if the order of the largest free chunk in the group changed.

Nice and clear explaination. Thanks :)

This change also looks good to me.

Reviewed-by Ritesh Harjani (IBM) <[email protected]>

One other thought to further optimize -
Will it make a difference if rather then adding the group to the tail of the list, we add that group to the head of sbi->s_mb_largest_free_orders[new_order].

This is because this group is the latest from where blocks were allocated/freed, and hence the next allocation should first try from this group in order to keep the files/extents blocks close to each other?
(That sometimes might help with disk firmware to avoid doing discards if the freed block can be reused?)

Or does goal block will always cover that case by default and we might never require this? Maybe in a case of a new file within the same directory where the goal group has no free blocks, but the last group attempted should be retried first?

Quoted 11 lines · Ritesh Harjani (IBM)

Quoted 9 lines · Jan Kara

mb_set_largest_free_order() updates lists containing groups with largest chunk of free space of given order. The way it updates it leads to always moving the group to the tail of the list. Thus allocations looking for free space of given order effectively end up cycling through all groups (and due to initialization in last to first order). This spreads allocations among block groups which reduces performance for rotating disks or low-end flash media. Change mb_set_largest_free_order() to only update lists if the order of the largest free chunk in the group changed.

Quoted 4 lines

Nice and clear explaination. Thanks :)

This change also looks good to me.

Reviewed-by Ritesh Harjani (IBM) <[email protected]>

Thanks for review!

Quoted 10 lines

One other thought to further optimize -
Will it make a difference if rather then adding the group to the tail of the list, we add that group to the head of sbi->s_mb_largest_free_orders[new_order].

This is because this group is the latest from where blocks were allocated/freed, and hence the next allocation should first try from this group in order to keep the files/extents blocks close to each other?
(That sometimes might help with disk firmware to avoid doing discards if the freed block can be reused?)

Quoted 4 lines

Or does goal block will always cover that case by default and we might never require this? Maybe in a case of a new file within the same directory where the goal group has no free blocks, but the last group attempted should be retried first?

So I was also wondering about this somewhat. I think that goal group will take care of keeping file data together so head/tail insertion should not matter too much for one file. Maybe if the allocation comes from a different inode, then the head/tail insertion matters but then it is not certain whether the allocation is actually related and what its order is (depending on that we might prefer same / different group) so I've decided to just keep things as they are. I agree it might be interesting to investigate and experiment with various workloads and see whether the head/tail insertion makes a difference for some workload but I think it's a separate project.

Curently we don't use any preallocation when a file is already closed when allocating blocks (from writeback code when converting delayed allocation). However for small files, using locality group preallocation

is actually desirable as that is not specific to a particular file. Rather it is a method to pack small files together to reduce fragmentation and for that the fact the file is closed is actually even stronger hint the file would benefit from packing. So change the logic to allow locality group preallocation in this case.

Signed-off-by Jan Kara <[email protected]>

Hi Jan,

https://git-scm.com/docs/git-format-patch#_base_tree_information]

url:    https://github.com/intel-lab-lkp/linux/commits/Jan-Kara/ext4-Fix-performance-regression-with-mballoc/20220907-000945
base:   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git 53e99dcff61e1523ec1c3628b2d564ba15d32eb7
config: m68k-randconfig-m041-20220906 (https://download.01.org/0day-ci/archive/20220907/202209071206.u1iHKVzB-lkp(a)intel.com/config)
compiler: m68k-linux-gcc (GCC) 12.1.0

If you fix the issue, kindly add following tag where applicable

Reported-by kernel test robot <[email protected]>

Reported-by Dan Carpenter <[email protected]>

New smatch warnings:
fs/ext4/mballoc.c:945 ext4_mb_choose_next_group_cr1() error: uninitialized symbol 'grp'.

vim +/grp +945 fs/ext4/mballoc.c

196e402adf2e4c Harshad Shirwadkar 2021-04-01  909  static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac,
196e402adf2e4c Harshad Shirwadkar 2021-04-01  910  		int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
196e402adf2e4c Harshad Shirwadkar 2021-04-01  911  {
196e402adf2e4c Harshad Shirwadkar 2021-04-01  912  	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
31b571b608cf66 Jan Kara           2022-09-06  913  	struct ext4_group_info *grp, *iter;
31b571b608cf66 Jan Kara           2022-09-06  914  	int i;
196e402adf2e4c Harshad Shirwadkar 2021-04-01  915  
196e402adf2e4c Harshad Shirwadkar 2021-04-01  916  	if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) {
196e402adf2e4c Harshad Shirwadkar 2021-04-01  917  		if (sbi->s_mb_stats)
196e402adf2e4c Harshad Shirwadkar 2021-04-01  918  			atomic_inc(&sbi->s_bal_cr1_bad_suggestions);
31b571b608cf66 Jan Kara           2022-09-06  919  	}
31b571b608cf66 Jan Kara           2022-09-06  920  
31b571b608cf66 Jan Kara           2022-09-06  921  	for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len);
31b571b608cf66 Jan Kara           2022-09-06  922  	     i < MB_NUM_ORDERS(ac->ac_sb); i++) {
31b571b608cf66 Jan Kara           2022-09-06  923  		if (list_empty(&sbi->s_mb_avg_fragment_size[i]))
31b571b608cf66 Jan Kara           2022-09-06  924  			continue;
31b571b608cf66 Jan Kara           2022-09-06  925  		read_lock(&sbi->s_mb_avg_fragment_size_locks[i]);
31b571b608cf66 Jan Kara           2022-09-06  926  		if (list_empty(&sbi->s_mb_avg_fragment_size[i])) {
31b571b608cf66 Jan Kara           2022-09-06  927  			read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
31b571b608cf66 Jan Kara           2022-09-06  928  			continue;

Smatch worries that we can hit these two continues on every iteration. Why not just initialize "grp = NULL;" at the start of the function?

Using rbtree for sorting groups by average fragment size is relatively expensive (needs rbtree update on every block freeing or allocation) and leads to wide spreading of allocations because selection of block group is very sentitive both to changes in free space and amount of blocks allocated. Furthermore selecting group with the best matching average fragment size is not necessary anyway, even more so because the variability of fragment sizes within a group is likely large so average is not telling much. We just need a group with large enough average fragment size so that we have high probability of finding large enough free extent and we don't want average fragment size to be too big so that we are likely to find free extent only somewhat larger than what we need.

So instead of maintaing rbtree of groups sorted by fragment size keep bins (lists) or groups where average fragment size is in the interval [2^i, 2^(i+1)). This structure requires less updates on block allocation / freeing, generally avoids chaotic spreading of allocations into block groups, and still is able to quickly (even faster that the rbtree) provide a block group which is likely to have a suitably sized free space extent.

This patch reduces number of block groups used when untarring archive with medium sized files (size somewhat above 64k which is default mballoc limit for avoiding locality group preallocation) to about half and thus improves write speeds for eMMC flash significantly.

Signed-off-by Jan Kara <[email protected]>

---
 fs/ext4/ext4.h    |  10 +-
 fs/ext4/mballoc.c | 252 +++++++++++++++++++---------------------------
 fs/ext4/mballoc.h |   1 -
 3 files changed, 110 insertions(+), 153 deletions(-)

fs/ext4/ext4.h +5 -5

index 9bca5565547b..3bf9a6926798 100644
@@ -167,8 +167,6 @@ enum SHIFT_DIRECTION {
 #define EXT4_MB_CR0_OPTIMIZED		0x8000
 /* Avg fragment size rb tree lookup succeeded at least once for cr = 1 */
 #define EXT4_MB_CR1_OPTIMIZED		0x00010000
-/* Perform linear traversal for one group */
-#define EXT4_MB_SEARCH_NEXT_LINEAR	0x00020000
 struct ext4_allocation_request {
 	/* target inode for block we're allocating */
 	struct inode *inode;

@@ -1600,8 +1598,8 @@ struct ext4_sb_info {
 	struct list_head s_discard_list;
 	struct work_struct s_discard_work;
 	atomic_t s_retry_alloc_pending;
-	struct rb_root s_mb_avg_fragment_size_root;
-	rwlock_t s_mb_rb_lock;
+	struct list_head *s_mb_avg_fragment_size;
+	rwlock_t *s_mb_avg_fragment_size_locks;
 	struct list_head *s_mb_largest_free_orders;
 	rwlock_t *s_mb_largest_free_orders_locks;
 

@@ -3413,6 +3411,8 @@ struct ext4_group_info {
 	ext4_grpblk_t	bb_first_free;	/* first free block */
 	ext4_grpblk_t	bb_free;	/* total free blocks */
 	ext4_grpblk_t	bb_fragments;	/* nr of freespace fragments */
+	int		bb_avg_fragment_size_order;	/* order of average
+							   fragment in BG */
 	ext4_grpblk_t	bb_largest_free_order;/* order of largest frag in BG */
 	ext4_group_t	bb_group;	/* Group number */
 	struct          list_head bb_prealloc_list;

@@ -3420,7 +3420,7 @@ struct ext4_group_info {
 	void            *bb_bitmap;
 #endif
 	struct rw_semaphore alloc_sem;
-	struct rb_node	bb_avg_fragment_size_rb;
+	struct list_head bb_avg_fragment_size_node;
 	struct list_head bb_largest_free_order_node;
 	ext4_grpblk_t	bb_counters[];	/* Nr of free power-of-two-block
 					 * regions, index is order.

fs/ext4/mballoc.c +23 -42

index af1e49c3603f..213d2d0750dd 100644
@@ -140,13 +140,15 @@
  *    number of buddy bitmap orders possible) number of lists. Group-infos are
  *    placed in appropriate lists.
  *
- * 2) Average fragment size rb tree (sbi->s_mb_avg_fragment_size_root)
+ * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size)
  *
- *    Locking: sbi->s_mb_rb_lock (rwlock)
+ *    Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks)
  *
- *    This is a red black tree consisting of group infos and the tree is sorted
- *    by average fragment sizes (which is calculated as ext4_group_info->bb_free
- *    / ext4_group_info->bb_fragments).
+ *    This is an array of lists where in the i-th list there are groups with
+ *    average fragment size >= 2^i and < 2^(i+1). The average fragment size
+ *    is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments.
+ *    Note that we don't bother with a special list for completely empty groups
+ *    so we only have MB_NUM_ORDERS(sb) lists.
  *
  * When "mb_optimize_scan" mount option is set, mballoc consults the above data
  * structures to decide the order in which groups are to be traversed for

@@ -160,7 +162,8 @@
  *
  * At CR = 1, we only consider groups where average fragment size > request
  * size. So, we lookup a group which has average fragment size just above or
- * equal to request size using our rb tree (data structure 2) in O(log N) time.
+ * equal to request size using our average fragment size group lists (data
+ * structure 2) in O(1) time.
  *
  * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in
  * linear order which requires O(N) search time for each CR 0 and CR 1 phase.

@@ -802,65 +805,51 @@ static void ext4_mb_mark_free_simple(struct super_block *sb,
 	}
 }
 
-static void ext4_mb_rb_insert(struct rb_root *root, struct rb_node *new,
-			int (*cmp)(struct rb_node *, struct rb_node *))
+static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len)
 {
-	struct rb_node **iter = &root->rb_node, *parent = NULL;
+	int order;
 
-	while (*iter) {
-		parent = *iter;
-		if (cmp(new, *iter) > 0)
-			iter = &((*iter)->rb_left);
-		else
-			iter = &((*iter)->rb_right);
-	}
-
-	rb_link_node(new, parent, iter);
-	rb_insert_color(new, root);
-}
-
-static int
-ext4_mb_avg_fragment_size_cmp(struct rb_node *rb1, struct rb_node *rb2)
-{
-	struct ext4_group_info *grp1 = rb_entry(rb1,
-						struct ext4_group_info,
-						bb_avg_fragment_size_rb);
-	struct ext4_group_info *grp2 = rb_entry(rb2,
-						struct ext4_group_info,
-						bb_avg_fragment_size_rb);
-	int num_frags_1, num_frags_2;
-
-	num_frags_1 = grp1->bb_fragments ?
-		grp1->bb_free / grp1->bb_fragments : 0;
-	num_frags_2 = grp2->bb_fragments ?
-		grp2->bb_free / grp2->bb_fragments : 0;
-
-	return (num_frags_2 - num_frags_1);
+	/*
+	 * We don't bother with a special lists groups with only 1 block free
+ 	 * extents and for completely empty groups.
+	 */
+	order = fls(len) - 2;
+	if (order < 0)
+		return 0;
+	if (order == MB_NUM_ORDERS(sb))
+		order--;
+	return order;
 }
 
-/*
- * Reinsert grpinfo into the avg_fragment_size tree with new average
- * fragment size.
- */
+/* Move group to appropriate avg_fragment_size list */
 static void
 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
 {
 	struct ext4_sb_info *sbi = EXT4_SB(sb);
+	int new_order;
 
 	if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0)

Quoted 13 lines · Jan Kara

Using rbtree for sorting groups by average fragment size is relatively expensive (needs rbtree update on every block freeing or allocation) and leads to wide spreading of allocations because selection of block group is very sentitive both to changes in free space and amount of blocks allocated. Furthermore selecting group with the best matching average fragment size is not necessary anyway, even more so because the variability of fragment sizes within a group is likely large so average is not telling much. We just need a group with large enough average fragment size so that we have high probability of finding large enough free extent and we don't want average fragment size to be too big so that we are likely to find free extent only somewhat larger than what we need.

Quoted 7 lines

So instead of maintaing rbtree of groups sorted by fragment size keep bins (lists) or groups where average fragment size is in the interval [2^i, 2^(i+1)). This structure requires less updates on block allocation / freeing, generally avoids chaotic spreading of allocations into block groups, and still is able to quickly (even faster that the rbtree) provide a block group which is likely to have a suitably sized free space extent.

This makes sense because we anyways maintain buddy bitmap for MB_NUM_ORDERS bitmaps. Hence our data structure to maintain different lists of groups, with their average fragments size can be bounded within MB_NUM_ORDERS lists. This also makes it for amortized O(1) search time for finding the right group in CR1 search.

Quoted 1 line

Quoted 5 lines

This patch reduces number of block groups used when untarring archive with medium sized files (size somewhat above 64k which is default mballoc limit for avoiding locality group preallocation) to about half and thus improves write speeds for eMMC flash significantly.

Indeed a nice change. More inline with the how we maintain sbi->s_mb_largest_free_orders lists.

I think as you already noted there are few minor checkpatch errors, other than that one small query below.

Quoted 14 lines

fs/ext4/mballoc.c +2 -4

-/*
- * Reinsert grpinfo into the avg_fragment_size tree with new average
- * fragment size.
- */
+/* Move group to appropriate avg_fragment_size list */
 static void
 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
 {
 	struct ext4_sb_info *sbi = EXT4_SB(sb);
+	int new_order;
 
 	if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0)
 		return;

fs/ext4/mballoc.c static void ext4_mb_mark_free_simple(struct super_block *sb, · line 835

Quoted 8 lines

fs/ext4/mballoc.c +2 -6

-	write_lock(&sbi->s_mb_rb_lock);
-	if (!RB_EMPTY_NODE(&grp->bb_avg_fragment_size_rb)) {
-		rb_erase(&grp->bb_avg_fragment_size_rb,
-				&sbi->s_mb_avg_fragment_size_root);
-		RB_CLEAR_NODE(&grp->bb_avg_fragment_size_rb);
-	}
+	new_order = mb_avg_fragment_size_order(sb,
+					grp->bb_free / grp->bb_fragments);

Previous rbtree change was always checking for if grp->bb_fragments for 0. Can grp->bb_fragments be 0 here?

Quoted 35 lines · Ritesh Harjani (IBM)

Quoted 13 lines · Jan Kara

Using rbtree for sorting groups by average fragment size is relatively expensive (needs rbtree update on every block freeing or allocation) and leads to wide spreading of allocations because selection of block group is very sentitive both to changes in free space and amount of blocks allocated. Furthermore selecting group with the best matching average fragment size is not necessary anyway, even more so because the variability of fragment sizes within a group is likely large so average is not telling much. We just need a group with large enough average fragment size so that we have high probability of finding large enough free extent and we don't want average fragment size to be too big so that we are likely to find free extent only somewhat larger than what we need.

Quoted 7 lines

So instead of maintaing rbtree of groups sorted by fragment size keep bins (lists) or groups where average fragment size is in the interval [2^i, 2^(i+1)). This structure requires less updates on block allocation / freeing, generally avoids chaotic spreading of allocations into block groups, and still is able to quickly (even faster that the rbtree) provide a block group which is likely to have a suitably sized free space extent.

This makes sense because we anyways maintain buddy bitmap for MB_NUM_ORDERS bitmaps. Hence our data structure to maintain different lists of groups, with their average fragments size can be bounded within MB_NUM_ORDERS lists. This also makes it for amortized O(1) search time for finding the right group in CR1 search.

Quoted 6 lines

This patch reduces number of block groups used when untarring archive with medium sized files (size somewhat above 64k which is default mballoc limit for avoiding locality group preallocation) to about half and thus improves write speeds for eMMC flash significantly.

Quoted 2 lines

Indeed a nice change. More inline with the how we maintain sbi->s_mb_largest_free_orders lists.

I didn't really find more comments than the one below?

Quoted 3 lines

I think as you already noted there are few minor checkpatch errors, other than that one small query below.

Yep, some checkpatch errors + procfs file handling bugs + one bad unlock in an error recovery path. All fixed up locally :)

Quoted 15 lines

Quoted 15 lines

fs/ext4/mballoc.c +2 -4

-/*
- * Reinsert grpinfo into the avg_fragment_size tree with new average
- * fragment size.
- */
+/* Move group to appropriate avg_fragment_size list */
 static void
 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
 {
 	struct ext4_sb_info *sbi = EXT4_SB(sb);
+	int new_order;
 
 	if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0)
 		return;

Quoted 11 lines

fs/ext4/mballoc.c static void ext4_mb_mark_free_simple(struct super_block *sb, · line 835

Quoted 8 lines

fs/ext4/mballoc.c +2 -6

-	write_lock(&sbi->s_mb_rb_lock);
-	if (!RB_EMPTY_NODE(&grp->bb_avg_fragment_size_rb)) {
-		rb_erase(&grp->bb_avg_fragment_size_rb,
-				&sbi->s_mb_avg_fragment_size_root);
-		RB_CLEAR_NODE(&grp->bb_avg_fragment_size_rb);
-	}
+	new_order = mb_avg_fragment_size_order(sb,
+					grp->bb_free / grp->bb_fragments);

Previous rbtree change was always checking for if grp->bb_fragments for 0. Can grp->bb_fragments be 0 here?

Since grp->bb_free is greater than zero, there should be at least one fragment...

** snip **
/*
* Choose next group by traversing average fragment size tree. Updates *new_cr

Hi Jan,

Quoted 4 lines · Jan Kara

** snip **
/*
* Choose next group by traversing average fragment size tree. Updates *new_cr

Maybe we can change this to "average fragment size list of suitable order"

Quoted 7 lines

- * if cr lvel needs an update. Sets EXT4_MB_SEARCH_NEXT_LINEAR to indicate that
- * the linear search should continue for one iteration since there's lock
- * contention on the rb tree lock.
+ * if cr level needs an update.
*/
static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac,
int *new_cr, ext4_group_t *group, ext4_group_t ngroups)

fs/ext4/mballoc.c static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac, · line 929

Quoted 5 lines

{

fs/ext4/mballoc.c +0 -3

 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
-	int avg_fragment_size, best_so_far;
-	struct rb_node *node, *found;
-	struct ext4_group_info *grp;

Other than that, this patch along with the updated mb_structs_summary proc file change looks good to me.

Hello,

Here is a second version of my mballoc improvements to avoid spreading allocations with mb_optimize_scan=1. The patches fix the performance regression I was able to reproduce with reaim on my test machine:

mb_optimize_scan=0 mb_optimize_scan=1 patched

Hmean     disk-1       2076.12 (   0.00%)     2099.37 (   1.12%)     2032.52 (  -2.10%)
Hmean     disk-41     92481.20 (   0.00%)    83787.47 *  -9.40%*    90308.37 (  -2.35%)
Hmean     disk-81    155073.39 (   0.00%)   135527.05 * -12.60%*   154285.71 (  -0.51%)
Hmean     disk-121   185109.64 (   0.00%)   166284.93 * -10.17%*   185298.62 (   0.10%)
Hmean     disk-161   229890.53 (   0.00%)   207563.39 *  -9.71%*   232883.32 *   1.30%*
Hmean     disk-201   223333.33 (   0.00%)   203235.59 *  -9.00%*   221446.93 (  -0.84%)
Hmean     disk-241   235735.25 (   0.00%)   217705.51 *  -7.65%*   239483.27 *   1.59%*
Hmean     disk-281   266772.15 (   0.00%)   241132.72 *  -9.61%*   263108.62 (  -1.37%)
Hmean     disk-321   265435.50 (   0.00%)   245412.84 *  -7.54%*   267277.27 (   0.69%)

The changes also significanly reduce spreading of allocations for small / moderately sized files. I'm not able to measure a performance difference resulting from this but on eMMC storage this seems to be the main culprit of reduced performance. Untarring of raspberry-pi archive touches following numbers of groups:

mb_optimize_scan=0 mb_optimize_scan=1 patched
groups 4 22 7

To achieve this I have added two more changes on top of v1 - patches 4 and 5. Patch 4 makes sure we use locality group preallocation even for files that are not likely to grow anymore (previously we have disabled all preallocations for such files, however locality group preallocation still makes a lot of sense for such files). This patch reduced spread of a small file allocations but larger file allocations were still spread significantly because they avoid locality group preallocation and as they are not power-of-two in size, they also immediately start with cr=1 scan. To address that I've changed the data structure for looking up the best block group to allocate from (see patch 5 for details).

Stefan, can you please test whether these patches fix the problem for you as well? Comments & review welcome.

Honza
Previous versions:

Link http://lore.kernel.org/r/[email protected] # v1 open here

Am 06.09.22 um 17:29 schrieb Jan Kara:

Quoted 37 lines

Hello,

Here is a second version of my mballoc improvements to avoid spreading allocations with mb_optimize_scan=1. The patches fix the performance regression I was able to reproduce with reaim on my test machine:

mb_optimize_scan=0 mb_optimize_scan=1 patched

Hmean     disk-1       2076.12 (   0.00%)     2099.37 (   1.12%)     2032.52 (  -2.10%)
Hmean     disk-41     92481.20 (   0.00%)    83787.47 *  -9.40%*    90308.37 (  -2.35%)
Hmean     disk-81    155073.39 (   0.00%)   135527.05 * -12.60%*   154285.71 (  -0.51%)
Hmean     disk-121   185109.64 (   0.00%)   166284.93 * -10.17%*   185298.62 (   0.10%)
Hmean     disk-161   229890.53 (   0.00%)   207563.39 *  -9.71%*   232883.32 *   1.30%*
Hmean     disk-201   223333.33 (   0.00%)   203235.59 *  -9.00%*   221446.93 (  -0.84%)
Hmean     disk-241   235735.25 (   0.00%)   217705.51 *  -7.65%*   239483.27 *   1.59%*
Hmean     disk-281   266772.15 (   0.00%)   241132.72 *  -9.61%*   263108.62 (  -1.37%)
Hmean     disk-321   265435.50 (   0.00%)   245412.84 *  -7.54%*   267277.27 (   0.69%)

The changes also significanly reduce spreading of allocations for small / moderately sized files. I'm not able to measure a performance difference resulting from this but on eMMC storage this seems to be the main culprit of reduced performance. Untarring of raspberry-pi archive touches following numbers of groups:

mb_optimize_scan=0 mb_optimize_scan=1 patched
groups 4 22 7

To achieve this I have added two more changes on top of v1 - patches 4 and 5. Patch 4 makes sure we use locality group preallocation even for files that are not likely to grow anymore (previously we have disabled all preallocations for such files, however locality group preallocation still makes a lot of sense for such files). This patch reduced spread of a small file allocations but larger file allocations were still spread significantly because they avoid locality group preallocation and as they are not power-of-two in size, they also immediately start with cr=1 scan. To address that I've changed the data structure for looking up the best block group to allocate from (see patch 5 for details).

Quoted 2 lines

Stefan, can you please test whether these patches fix the problem for you as well? Comments & review welcome.

this looks amazing \o/

With this patch v2 applied the untar with mb_optimize_scan=1 is now faster than mb_optimize_scan=0.

mb_optimize_scan=0 -> almost 5 minutes

mb_optimize_scan=1 -> almost 1 minute

The original scenario (firmware download) with mb_optimize_scan=1 is now fast as mb_optimize_scan=0.

Here the iostat as usual:

https://github.com/lategoodbye/mb_optimize_scan_regress/commit/f4ad188e0feee60bffa23a8e1ad254544768c3bd

There is just one thing, but not sure this if this comes from these patches. If i call

cat /proc/fs/ext4/mmcblk1p2/mb_structs_summary

The kernel throw a NULL pointer derefence in
ext4_mb_seq_structs_summary_show

Quoted 7 lines

Here the iostat as usual:

https://github.com/lategoodbye/mb_optimize_scan_regress/commit/f4ad188e0feee60bffa23a8e1ad254544768c3bd

There is just one thing, but not sure this if this comes from these patches. If i call

Quoted 4 lines

cat /proc/fs/ext4/mmcblk1p2/mb_structs_summary

The kernel throw a NULL pointer derefence in
ext4_mb_seq_structs_summary_show

Yeah, likely a bug in my last patch. I didn't test my changes to the sysfs interface. Thanks for testing this, I'll have a look.

Hello,

Here is a second version of my mballoc improvements to avoid spreading allocations with mb_optimize_scan=1. The patches fix the performance regression I was able to reproduce with reaim on my test machine:

mb_optimize_scan=0 mb_optimize_scan=1 patched

Hmean     disk-1       2076.12 (   0.00%)     2099.37 (   1.12%)     2032.52 (  -2.10%)
Hmean     disk-41     92481.20 (   0.00%)    83787.47 *  -9.40%*    90308.37 (  -2.35%)
Hmean     disk-81    155073.39 (   0.00%)   135527.05 * -12.60%*   154285.71 (  -0.51%)
Hmean     disk-121   185109.64 (   0.00%)   166284.93 * -10.17%*   185298.62 (   0.10%)
Hmean     disk-161   229890.53 (   0.00%)   207563.39 *  -9.71%*   232883.32 *   1.30%*
Hmean     disk-201   223333.33 (   0.00%)   203235.59 *  -9.00%*   221446.93 (  -0.84%)
Hmean     disk-241   235735.25 (   0.00%)   217705.51 *  -7.65%*   239483.27 *   1.59%*
Hmean     disk-281   266772.15 (   0.00%)   241132.72 *  -9.61%*   263108.62 (  -1.37%)
Hmean     disk-321   265435.50 (   0.00%)   245412.84 *  -7.54%*   267277.27 (   0.69%)

The changes also significanly reduce spreading of allocations for small / moderately sized files. I'm not able to measure a performance difference resulting from this but on eMMC storage this seems to be the main culprit of reduced performance. Untarring of raspberry-pi archive touches following numbers of groups:

mb_optimize_scan=0 mb_optimize_scan=1 patched
groups 4 22 7

To achieve this I have added two more changes on top of v1 - patches 4 and 5. Patch 4 makes sure we use locality group preallocation even for files that are not likely to grow anymore (previously we have disabled all preallocations for such files, however locality group preallocation still makes a lot of sense for such files). This patch reduced spread of a small file allocations but larger file allocations were still spread significantly because they avoid locality group preallocation and as they are not power-of-two in size, they also immediately start with cr=1 scan. To address that I've changed the data structure for looking up the best block group to allocate from (see patch 5 for details).

Stefan, can you please test whether these patches fix the problem for you as well? Comments & review welcome.

Honza
Previous versions:

Link http://lore.kernel.org/r/[email protected] # v1 open here

Quoted 37 lines · Jan Kara

Hello,

Here is a second version of my mballoc improvements to avoid spreading allocations with mb_optimize_scan=1. The patches fix the performance regression I was able to reproduce with reaim on my test machine:

mb_optimize_scan=0 mb_optimize_scan=1 patched

Hmean     disk-1       2076.12 (   0.00%)     2099.37 (   1.12%)     2032.52 (  -2.10%)
Hmean     disk-41     92481.20 (   0.00%)    83787.47 *  -9.40%*    90308.37 (  -2.35%)
Hmean     disk-81    155073.39 (   0.00%)   135527.05 * -12.60%*   154285.71 (  -0.51%)
Hmean     disk-121   185109.64 (   0.00%)   166284.93 * -10.17%*   185298.62 (   0.10%)
Hmean     disk-161   229890.53 (   0.00%)   207563.39 *  -9.71%*   232883.32 *   1.30%*
Hmean     disk-201   223333.33 (   0.00%)   203235.59 *  -9.00%*   221446.93 (  -0.84%)
Hmean     disk-241   235735.25 (   0.00%)   217705.51 *  -7.65%*   239483.27 *   1.59%*
Hmean     disk-281   266772.15 (   0.00%)   241132.72 *  -9.61%*   263108.62 (  -1.37%)
Hmean     disk-321   265435.50 (   0.00%)   245412.84 *  -7.54%*   267277.27 (   0.69%)

The changes also significanly reduce spreading of allocations for small / moderately sized files. I'm not able to measure a performance difference resulting from this but on eMMC storage this seems to be the main culprit of reduced performance. Untarring of raspberry-pi archive touches following numbers of groups:

mb_optimize_scan=0 mb_optimize_scan=1 patched
groups 4 22 7

To achieve this I have added two more changes on top of v1 - patches 4 and 5. Patch 4 makes sure we use locality group preallocation even for files that are not likely to grow anymore (previously we have disabled all preallocations for such files, however locality group preallocation still makes a lot of sense for such files). This patch reduced spread of a small file allocations but larger file allocations were still spread significantly because they avoid locality group preallocation and as they are not power-of-two in size, they also immediately start with cr=1 scan. To address that I've changed the data structure for looking up the best block group to allocate from (see patch 5 for details).

Quoted 6 lines

Stefan, can you please test whether these patches fix the problem for you as well? Comments & review welcome.

Honza
Previous versions:

Link http://lore.kernel.org/r/[email protected] # v1 open here

Hi Jan,

Thanks for the patch. I tested this series on my raspberry pi and I can confirm that the regression is no longer present with both mb_optimize_scan=0 and =1 taking similar amount of time to untar. The allocation spread I'm seeing is as follows:
mb_optimize_scan=0: 10
mb_optimize_scan=1: 17 (Check graphs for more details)

For mb_optimize_scan=1, I also compared the spread of locality group PA eligible files (<64KB) and inode pa files. The results can be found here:

mb_optimize_scan=0:
https://github.com/OjaswinM/mbopt-bug/blob/master/grpahs/patchv2-mbopt0.png
mb_optimize_scan=1:
https://github.com/OjaswinM/mbopt-bug/blob/master/grpahs/patchv2.png
mb_optimize_scan=1 (lg pa only):
https://github.com/OjaswinM/mbopt-bug/blob/master/grpahs/patchv2-lgs.png
mb_optimize_scan=1 (inode pa only):
https://github.com/OjaswinM/mbopt-bug/blob/master/grpahs/patchv2-i.png

The smaller files are now closer together due to the changes to locality group pa logic. Most of the spread is now coming from mid to large files.

To test this further, I created a tar of 2000 100KB files to see if there is any performance drop due to higher spread of these files and notcied that although the spread is slightly higher(5BGs vs 9), we don't see a performance drop while untarring with mb_optimize_scan=1.

Although we still have some spread, I think we have brought it down to a much more manageable level and that combined with improvements to CR1 allocation have given us a good performance improvement.

Feel free to add:

Tested-by Ojaswin Mujoo <[email protected]>