Tree fragmentation and prefetching

While looking into the performance of scrub I noticed that a significant amount of time is being used for loading the extent tree and the csum tree. While this is no surprise I did some prototyping on how to improve on it.
The main idea is to load the tree (or parts of it) top-down, order the needed blocks and distribute it over all disks.
To keep you interested, some results first.

a) by tree enumeration with reada=2
reading extent tree: 242s
reading csum tree: 140s

reading both trees: 324s

b) prefetch prototype
reading extent tree: 23.5s
reading csum tree: 20.4s
reading both trees: 25.7s

The test setup consists of a 7 Seagate ES.2 1TB disks filesystem, filled 28%. It is created with the current git tree + the round robin patch and filled with

fs_mark -D 512 -t 16 -n 4096 -F -S0

The 'normal' read is done by enumerating the leaves by btrfs_next_leaf() with path->reada=2. Both trees are being enumerated one after the other. The prototype currently just uses raw bios, does not make use of the page cache and does not enter the read pages into the cache. This will probably add some overhead. It also does not check the crcs.

While it is very promising to implement it for scrub, I think a more general interface which can be used for every enumeration would be beneficial. Use cases that come to mind are rebalance, reflink, deletion of large files, listing of large directories etc..

I'd imagine an interface along the lines of

int btrfs_readahead_init(struct btrfs_reada_ctx *reada); int btrfs_readahead_add(struct btrfs_root *root,
struct btrfs_key *start, struct btrfs_key *end, struct btrfs_reada_ctx *reada);
void btrfs_readahead_wait(struct btrfs_reada_ctx *reada);

to trigger the readahead of parts of a tree. Multiple readahead requests can be given before waiting. This would enable the very beneficial folding seen above for 'reading both trees'.

Also it would be possible to add a cascading readahead, where the content of leaves would trigger readaheads in other trees, maybe by giving a callback for the decisions what to read instead of the fixed start/end range.

For the implementation I'd need an interface which I haven't been able to find yet. Currently I can trigger the read of several pages / tree blocks and wait for the completion of each of them. What I'd need would be an interface that gives me a callback on each completion or a waiting function that wakes up on each completion with the information which pages just completed.
One way to achieve this would be to add a hook, but I gladly take any implementation hints.

--
Arne

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

While looking into the performance of scrub I noticed that a signific=

amount of time is being used for loading the extent tree and the csum tree. While this is no surprise I did some prototyping on how to impr=

on it.
The main idea is to load the tree (or parts of it) top-down, order th=

needed blocks and distribute it over all disks.
To keep you interested, some results first.

a) by tree enumeration with reada=3D2
=C2=A0 reading extent tree: 242s
=C2=A0 reading csum tree: 140s
=C2=A0 reading both trees: 324s

b) prefetch prototype
=C2=A0 reading extent tree: 23.5s
=C2=A0 reading csum tree: 20.4s
=C2=A0 reading both trees: 25.7s

The test setup consists of a 7 Seagate ES.2 1TB disks filesystem, fil=

28%. It is created with the current git tree + the round robin patch =

filled with

fs_mark -D 512 -t 16 -n 4096 -F -S0

The 'normal' read is done by enumerating the leaves by btrfs_next_lea=

with path->reada=3D2. Both trees are being enumerated one after the o=

The prototype currently just uses raw bios, does not make use of the page cache and does not enter the read pages into the cache. This wil=

probably add some overhead. It also does not check the crcs.

While it is very promising to implement it for scrub, I think a more general interface which can be used for every enumeration would be beneficial. Use cases that come to mind are rebalance, reflink, delet=

of large files, listing of large directories etc..

I'd imagine an interface along the lines of

int btrfs_readahead_init(struct btrfs_reada_ctx *reada); int btrfs_readahead_add(struct btrfs_root *root,
=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0=

=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0=

=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0=

void btrfs_readahead_wait(struct btrfs_reada_ctx *reada);

to trigger the readahead of parts of a tree. Multiple readahead requests can be given before waiting. This would enable the very beneficial folding seen above for 'reading both trees'.

Also it would be possible to add a cascading readahead, where the content of leaves would trigger readaheads in other trees, maybe by giving a callback for the decisions what to read instead of the fixed start/end range.

For the implementation I'd need an interface which I haven't been abl=

to find yet. Currently I can trigger the read of several pages / tree blocks and wait for the completion of each of them. What I'd need wou=

be an interface that gives me a callback on each completion or a wait=

function that wakes up on each completion with the information which pages just completed.
One way to achieve this would be to add a hook, but I gladly take any implementation hints.

--
Arne

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs=

the body of a message to [email protected]
More majordomo info at =C2=A0http://vger.kernel.org/majordomo-info.ht=

While looking into the performance of scrub I noticed that a significant amount of time is being used for loading the extent tree and the csum tree. While this is no surprise I did some prototyping on how to improve on it.
The main idea is to load the tree (or parts of it) top-down, order the needed blocks and distribute it over all disks.
To keep you interested, some results first.

a) by tree enumeration with reada=2
reading extent tree: 242s
reading csum tree: 140s
reading both trees: 324s

b) prefetch prototype
reading extent tree: 23.5s
reading csum tree: 20.4s
reading both trees: 25.7s

The test setup consists of a 7 Seagate ES.2 1TB disks filesystem, filled 28%. It is created with the current git tree + the round robin patch and filled with

fs_mark -D 512 -t 16 -n 4096 -F -S0

The 'normal' read is done by enumerating the leaves by btrfs_next_leaf() with path->reada=2. Both trees are being enumerated one after the other. The prototype currently just uses raw bios, does not make use of the page cache and does not enter the read pages into the cache. This will probably add some overhead. It also does not check the crcs.

While it is very promising to implement it for scrub, I think a more general interface which can be used for every enumeration would be beneficial. Use cases that come to mind are rebalance, reflink, deletion of large files, listing of large directories etc..

I'd imagine an interface along the lines of

int btrfs_readahead_init(struct btrfs_reada_ctx *reada); int btrfs_readahead_add(struct btrfs_root *root,
struct btrfs_key *start, struct btrfs_key *end, struct btrfs_reada_ctx *reada);
void btrfs_readahead_wait(struct btrfs_reada_ctx *reada);

to trigger the readahead of parts of a tree. Multiple readahead requests can be given before waiting. This would enable the very beneficial folding seen above for 'reading both trees'.

Also it would be possible to add a cascading readahead, where the content of leaves would trigger readaheads in other trees, maybe by giving a callback for the decisions what to read instead of the fixed start/end range.

For the implementation I'd need an interface which I haven't been able to find yet. Currently I can trigger the read of several pages / tree blocks and wait for the completion of each of them. What I'd need would be an interface that gives me a callback on each completion or a waiting function that wakes up on each completion with the information which pages just completed.
One way to achieve this would be to add a hook, but I gladly take any implementation hints.

Quoted 13 lines · Arne Jansen

While looking into the performance of scrub I noticed that a significant amount of time is being used for loading the extent tree and the csum tree. While this is no surprise I did some prototyping on how to improve on it.
The main idea is to load the tree (or parts of it) top-down, order the needed blocks and distribute it over all disks.
To keep you interested, some results first.

a) by tree enumeration with reada=2
reading extent tree: 242s
reading csum tree: 140s
reading both trees: 324s

Quoted 4 lines

b) prefetch prototype
reading extent tree: 23.5s
reading csum tree: 20.4s
reading both trees: 25.7s

10x speed-up looks indeed impressive. Just for me to be sure, did I get you right in that you attribute this effect specifically to enumerating tree leaves in key address vs. disk addresses when these two are not aligned?

Regards,
Andrey

Quoted 45 lines

The test setup consists of a 7 Seagate ES.2 1TB disks filesystem, filled 28%. It is created with the current git tree + the round robin patch and filled with

fs_mark -D 512 -t 16 -n 4096 -F -S0

The 'normal' read is done by enumerating the leaves by btrfs_next_leaf() with path->reada=2. Both trees are being enumerated one after the other. The prototype currently just uses raw bios, does not make use of the page cache and does not enter the read pages into the cache. This will probably add some overhead. It also does not check the crcs.

While it is very promising to implement it for scrub, I think a more general interface which can be used for every enumeration would be beneficial. Use cases that come to mind are rebalance, reflink, deletion of large files, listing of large directories etc..

I'd imagine an interface along the lines of

int btrfs_readahead_init(struct btrfs_reada_ctx *reada); int btrfs_readahead_add(struct btrfs_root *root,
struct btrfs_key *start, struct btrfs_key *end, struct btrfs_reada_ctx *reada);
void btrfs_readahead_wait(struct btrfs_reada_ctx *reada);

to trigger the readahead of parts of a tree. Multiple readahead requests can be given before waiting. This would enable the very beneficial folding seen above for 'reading both trees'.

Also it would be possible to add a cascading readahead, where the content of leaves would trigger readaheads in other trees, maybe by giving a callback for the decisions what to read instead of the fixed start/end range.

For the implementation I'd need an interface which I haven't been able to find yet. Currently I can trigger the read of several pages / tree blocks and wait for the completion of each of them. What I'd need would be an interface that gives me a callback on each completion or a waiting function that wakes up on each completion with the information which pages just completed.
One way to achieve this would be to add a hook, but I gladly take any implementation hints.

Quoted 9 lines

--
Arne

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Quoted 2 lines · Andrey Kuzmin

On Wed, Mar 23, 2011 at 4:06 PM, Arne Jansen<[email protected]> =C2=A0=

Quoted 2 lines

Quoted 2 lines

While looking into the performance of scrub I noticed that a signif=

Quoted 1 line

Quoted 1 line

amount of time is being used for loading the extent tree and the cs=

Quoted 1 line

Quoted 1 line

tree. While this is no surprise I did some prototyping on how to im=

Quoted 2 lines

Quoted 2 lines

on it.
The main idea is to load the tree (or parts of it) top-down, order =

Quoted 13 lines

Quoted 8 lines

needed blocks and distribute it over all disks.
To keep you interested, some results first.

a) by tree enumeration with reada=3D2
=C2=A0 reading extent tree: 242s
=C2=A0 reading csum tree: 140s
=C2=A0 reading both trees: 324s

Quoted 4 lines

b) prefetch prototype
=C2=A0 reading extent tree: 23.5s
=C2=A0 reading csum tree: 20.4s
=C2=A0 reading both trees: 25.7s

Quoted 4 lines

10x speed-up looks indeed impressive. Just for me to be sure, did I get you right in that you attribute this effect specifically to enumerating tree leaves in key address vs. disk addresses when these two are not aligned?

Yes. Leaves and the intermediate nodes tend to be quite scattered around the disk with respect to their logical order. Reading them in logical (ascending/descending) order require lots of seeks.

Quoted 7 lines

Regards,
Andrey

Quoted 2 lines

The test setup consists of a 7 Seagate ES.2 1TB disks filesystem, f=

Quoted 1 line

Quoted 1 line

28%. It is created with the current git tree + the round robin patc=

Quoted 5 lines

Quoted 5 lines

filled with

fs_mark -D 512 -t 16 -n 4096 -F -S0

The 'normal' read is done by enumerating the leaves by btrfs_next_l=

Quoted 1 line

Quoted 1 line

with path->reada=3D2. Both trees are being enumerated one after the=

Quoted 1 line

Quoted 1 line

The prototype currently just uses raw bios, does not make use of th=

Quoted 1 line

Quoted 1 line

page cache and does not enter the read pages into the cache. This w=

Quoted 3 lines

Quoted 3 lines

probably add some overhead. It also does not check the crcs.

While it is very promising to implement it for scrub, I think a mor=

Quoted 2 lines

Quoted 2 lines

general interface which can be used for every enumeration would be beneficial. Use cases that come to mind are rebalance, reflink, del=

Quoted 7 lines

Quoted 7 lines

of large files, listing of large directories etc..

I'd imagine an interface along the lines of

int btrfs_readahead_init(struct btrfs_reada_ctx *reada); int btrfs_readahead_add(struct btrfs_root *root,
=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0=

Quoted 1 line

Quoted 1 line

=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0=

Quoted 1 line

Quoted 1 line

=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0=

Quoted 9 lines

Quoted 9 lines

void btrfs_readahead_wait(struct btrfs_reada_ctx *reada);

to trigger the readahead of parts of a tree. Multiple readahead requests can be given before waiting. This would enable the very beneficial folding seen above for 'reading both trees'.

Also it would be possible to add a cascading readahead, where the content of leaves would trigger readaheads in other trees, maybe by giving a callback for the decisions what to read instead of the fix=

Quoted 3 lines

Quoted 3 lines

start/end range.

For the implementation I'd need an interface which I haven't been a=

Quoted 1 line

Quoted 1 line

to find yet. Currently I can trigger the read of several pages / tr=

Quoted 1 line

Quoted 1 line

blocks and wait for the completion of each of them. What I'd need w=

Quoted 1 line

Quoted 1 line

be an interface that gives me a callback on each completion or a wa=

Quoted 1 line

Quoted 1 line

function that wakes up on each completion with the information whic=

Quoted 2 lines

Quoted 2 lines

pages just completed.
One way to achieve this would be to add a hook, but I gladly take a=

Quoted 8 lines

Quoted 8 lines

implementation hints.

--
Arne

--
To unsubscribe from this list: send the line "unsubscribe linux-btr=

Quoted 2 lines

Quoted 2 lines

the body of a message to [email protected]
More majordomo info at =C2=A0http://vger.kernel.org/majordomo-info.=

Quoted 3 lines

Quoted 1 line

--
To unsubscribe from this list: send the line "unsubscribe linux-btrf=

Quoted 2 lines

the body of a message to [email protected]
More majordomo info at =C2=A0http://vger.kernel.org/majordomo-info.h=

Quoted 19 lines · Andrey Kuzmin

Quoted 13 lines · Arne Jansen

While looking into the performance of scrub I noticed that a significant amount of time is being used for loading the extent tree and the csum tree. While this is no surprise I did some prototyping on how to improve on it.
The main idea is to load the tree (or parts of it) top-down, order the needed blocks and distribute it over all disks.
To keep you interested, some results first.

a) by tree enumeration with reada=2
reading extent tree: 242s
reading csum tree: 140s
reading both trees: 324s

Quoted 4 lines

b) prefetch prototype
reading extent tree: 23.5s
reading csum tree: 20.4s
reading both trees: 25.7s

Quoted 4 lines

10x speed-up looks indeed impressive. Just for me to be sure, did I get you right in that you attribute this effect specifically to enumerating tree leaves in key address vs. disk addresses when these two are not aligned?

Yes. Leaves and the intermediate nodes tend to be quite scattered around the disk with respect to their logical order. Reading them in logical (ascending/descending) order require lots of seeks.

Quoted 63 lines

Regards,
Andrey

Quoted 54 lines

The test setup consists of a 7 Seagate ES.2 1TB disks filesystem, filled 28%. It is created with the current git tree + the round robin patch and filled with

fs_mark -D 512 -t 16 -n 4096 -F -S0

The 'normal' read is done by enumerating the leaves by btrfs_next_leaf() with path->reada=2. Both trees are being enumerated one after the other. The prototype currently just uses raw bios, does not make use of the page cache and does not enter the read pages into the cache. This will probably add some overhead. It also does not check the crcs.

While it is very promising to implement it for scrub, I think a more general interface which can be used for every enumeration would be beneficial. Use cases that come to mind are rebalance, reflink, deletion of large files, listing of large directories etc..

I'd imagine an interface along the lines of

int btrfs_readahead_init(struct btrfs_reada_ctx *reada); int btrfs_readahead_add(struct btrfs_root *root,
struct btrfs_key *start, struct btrfs_key *end, struct btrfs_reada_ctx *reada);
void btrfs_readahead_wait(struct btrfs_reada_ctx *reada);

to trigger the readahead of parts of a tree. Multiple readahead requests can be given before waiting. This would enable the very beneficial folding seen above for 'reading both trees'.

Also it would be possible to add a cascading readahead, where the content of leaves would trigger readaheads in other trees, maybe by giving a callback for the decisions what to read instead of the fixed start/end range.

For the implementation I'd need an interface which I haven't been able to find yet. Currently I can trigger the read of several pages / tree blocks and wait for the completion of each of them. What I'd need would be an interface that gives me a callback on each completion or a waiting function that wakes up on each completion with the information which pages just completed.
One way to achieve this would be to add a hook, but I gladly take any implementation hints.

Signature · 7 lines

Arne


-- 
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Signature · 3 lines

To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Signature · 3 lines

To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Quoted 21 lines · Arne Jansen

Quoted 19 lines · Andrey Kuzmin

Quoted 9 lines · Arne Jansen

The main idea is to load the tree (or parts of it) top-down, order the needed blocks and distribute it over all disks.
To keep you interested, some results first.

a) by tree enumeration with reada=2
reading extent tree: 242s
reading csum tree: 140s
reading both trees: 324s

Quoted 4 lines

b) prefetch prototype
reading extent tree: 23.5s
reading csum tree: 20.4s
reading both trees: 25.7s

10x speed-up looks indeed impressive. Just for me to be sure, did I get you right in that you attribute this effect specifically to enumerating tree leaves in key address vs. disk addresses when these two are not aligned?

Quoted 4 lines

Yes. Leaves and the intermediate nodes tend to be quite scattered around the disk with respect to their logical order. Reading them in logical (ascending/descending) order require lots of seeks.

I'm also dealing with tree fragmentation problem, I try to store the leaves which have the same parent closely.

Regards
Miao

Quoted 27 lines · Miao Xie

Quoted 25 lines · Arne Jansen

Quoted 15 lines · Andrey Kuzmin

Quoted 9 lines · Arne Jansen

The main idea is to load the tree (or parts of it) top-down, order the needed blocks and distribute it over all disks.
To keep you interested, some results first.

a) by tree enumeration with reada=2
reading extent tree: 242s
reading csum tree: 140s
reading both trees: 324s

Quoted 4 lines

b) prefetch prototype
reading extent tree: 23.5s
reading csum tree: 20.4s
reading both trees: 25.7s

Quoted 4 lines

10x speed-up looks indeed impressive. Just for me to be sure, did I get you right in that you attribute this effect specifically to enumerating tree leaves in key address vs. disk addresses when these two are not aligned?

Yes. Leaves and the intermediate nodes tend to be quite scattered around the disk with respect to their logical order. Reading them in logical (ascending/descending) order require lots of seeks.

Quoted 2 lines

I'm also dealing with tree fragmentation problem, I try to store the leaves which have the same parent closely.

That's good to hear. Do you have already anything I can repeat the test with?

-Arne

Quoted 4 lines

Regards
Miao

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Excerpts from Arne Jansen's message of 2011-03-23 09:06:02 -0400:

Quoted 9 lines

For the implementation I'd need an interface which I haven't been able to find yet. Currently I can trigger the read of several pages / tree blocks and wait for the completion of each of them. What I'd need would be an interface that gives me a callback on each completion or a waiting function that wakes up on each completion with the information which pages just completed.
One way to achieve this would be to add a hook, but I gladly take any implementation hints.

We have the bio endio call backs for this, I think that's the only thing you can use.

Quoted 11 lines

Excerpts from Arne Jansen's message of 2011-03-23 09:06:02 -0400:

Quoted 9 lines

For the implementation I'd need an interface which I haven't been able to find yet. Currently I can trigger the read of several pages / tree blocks and wait for the completion of each of them. What I'd need would be an interface that gives me a callback on each completion or a waiting function that wakes up on each completion with the information which pages just completed.
One way to achieve this would be to add a hook, but I gladly take any implementation hints.

Quoted 3 lines

We have the bio endio call backs for this, I think that's the only thing you can use.

ok, so I'll add a new extent state bit EXTENT_READAHEAD and test for it in btree_readpage_end_io_hook.

Excerpts from Arne Jansen's message of 2011-03-25 16:14:35 -0400:

Quoted 16 lines · Chris Mason

Quoted 14 lines

Excerpts from Arne Jansen's message of 2011-03-23 09:06:02 -0400:

Quoted 9 lines

For the implementation I'd need an interface which I haven't been able to find yet. Currently I can trigger the read of several pages / tree blocks and wait for the completion of each of them. What I'd need would be an interface that gives me a callback on each completion or a waiting function that wakes up on each completion with the information which pages just completed.
One way to achieve this would be to add a hook, but I gladly take any implementation hints.

We have the bio endio call backs for this, I think that's the only thing you can use.

Quoted 2 lines

ok, so I'll add a new extent state bit EXTENT_READAHEAD and test for it in btree_readpage_end_io_hook.

It's also common to use a chain of endio handlers. If you're allocating any state for the RA, you just save the original endio handler in your new struct, and then use your own endio handler that does the readahead smarts.

Quoted 17 lines

Excerpts from Arne Jansen's message of 2011-03-25 16:14:35 -0400:

Quoted 16 lines · Chris Mason

Quoted 14 lines

Excerpts from Arne Jansen's message of 2011-03-23 09:06:02 -0400:

Quoted 9 lines

For the implementation I'd need an interface which I haven't been able to find yet. Currently I can trigger the read of several pages / tree blocks and wait for the completion of each of them. What I'd need would be an interface that gives me a callback on each completion or a waiting function that wakes up on each completion with the information which pages just completed.
One way to achieve this would be to add a hook, but I gladly take any implementation hints.

We have the bio endio call backs for this, I think that's the only thing you can use.

Quoted 8 lines

Quoted 2 lines

ok, so I'll add a new extent state bit EXTENT_READAHEAD and test for it in btree_readpage_end_io_hook.

It's also common to use a chain of endio handlers. If you're allocating any state for the RA, you just save the original endio handler in your new struct, and then use your own endio handler that does the readahead smarts.

Do you mean replacing the bio end_io handler or the readpage_end_io_hook? As I want the pages to end up in the page cache, I'd like to use as much of the existing infrastructure as possible. To intercept the bio deep down in the chain would mean to duplicate some code on the way down and on the way up again.