fs: don't set *REFERENCED on single use objects
PATCH 1/1 fs: don't set *REFERENCED on single use objects
By default we set DCACHE_REFERENCED and I_REFERENCED on any dentry or inode we create. This is problematic as this means that it takes two trips through the LRU for any of these objects to be reclaimed, regardless of their actual lifetime. With enough pressure from these caches we can easily evict our working set from page cache with single use objects. So instead only set *REFERENCED if we've already been added to the LRU list. This means that we've been touched since the first time we were accessed, and so more likely to need to hang out in cache.
To illustrate this issue I wrote the following scripts
https://github.com/josefbacik/debug-scripts/tree/master/cache-pressure
on my test box. It is a single socket 4 core CPU with 16gib of RAM and I tested on an Intel 2tib NVME drive. The cache-pressure.sh script creates a new file system and creates 2 6.5gib files in order to take up 13gib of the 16gib of ram with pagecache. Then it runs a test program that reads these 2 files in a loop, and keeps track of how often it has to read bytes for each loop. On an ideal system with no pressure we should have to read 0 bytes indefinitely. The second thing this script does is start a fs_mark job that creates a ton of 0 length files, putting pressure on the system with slab only allocations. On exit the script prints out how many bytes were read by the read-file program. The results are as follows
Without patch:/mnt/btrfs-test/reads/file1: total read during loops 27262988288 /mnt/btrfs-test/reads/file2: total read during loops 27262976000
With patch:/mnt/btrfs-test/reads/file2: total read during loops 18640457728 /mnt/btrfs-test/reads/file1: total read during loops 9565376512
This patch results in a 50% reduction of the amount of pages evicted from our working set.
--- fs/dcache.c | 4 ++-- fs/inode.c | 3 ++- 2 files changed, 4 insertions(+), 3 deletions(-)
fs/dcache.c +2 -2
@@ -419,6 +419,8 @@ static void dentry_lru_add(struct dentry *dentry) {if (unlikely (!(dentry->d_flags &DCACHE_LRU_LIST )))d_lru_add (dentry); +else if (unlikely (!(dentry->d_flags &DCACHE_REFERENCED ))) + dentry->d_flags |=DCACHE_REFERENCED ; }/**
@@ -779,8 +781,6 @@ void dput(struct dentry *dentry) goto kill_it; } -if (!(dentry->d_flags &DCACHE_REFERENCED )) - dentry->d_flags |=DCACHE_REFERENCED ;dentry_lru_add (dentry); dentry->d_lockref.count--;
fs/inode.c +2 -1
@@ -405,6 +405,8 @@ static void inode_lru_list_add(struct inode *inode) {if (list_lru_add (&inode->i_sb->s_inode_lru, &inode->i_lru))this_cpu_inc (nr_unused); +else + inode->i_state |=I_REFERENCED ; }/*
@@ -1492,7 +1494,6 @@ static void iput_final(struct inode *inode) drop =generic_drop_inode (inode);if (!drop && (sb->s_flags &MS_ACTIVE )) { - inode->i_state |=I_REFERENCED ;inode_add_lru (inode);spin_unlock (&inode->i_lock);return ;
Signature · 1 line
2.7.4
Jan Kara
reviewerQuoted 30 lines · Josef Bacik
By default we set
DCACHE_REFERENCEDandI_REFERENCEDon any dentry or inode we create. This is problematic as this means that it takes two trips through the LRU for any of these objects to be reclaimed, regardless of their actual lifetime. With enough pressure from these caches we can easily evict our working set from page cache with single use objects. So instead only set *REFERENCED if we've already been added to the LRU list. This means that we've been touched since the first time we were accessed, and so more likely to need to hang out in cache.To illustrate this issue I wrote the following scripts
https://github.com/josefbacik/debug-scripts/tree/master/cache-pressure
on my test box. It is a single socket 4 core CPU with 16gib of RAM and I tested on an Intel 2tib NVME drive. The
cache-pressure.shscript creates a new file system and creates 2 6.5gib files in order to take up 13gib of the 16gib of ram with pagecache. Then it runs a test program that reads these 2 files in a loop, and keeps track of how often it has to read bytes for each loop. On an ideal system with no pressure we should have to read 0 bytes indefinitely. The second thing this script does is start afs_markjob that creates a ton of 0 length files, putting pressure on the system with slab only allocations. On exit the script prints out how many bytes were read by the read-file program. The results are as followsWithout patch:
/mnt/btrfs-test/reads/file1: total read during loops 27262988288/mnt/btrfs-test/reads/file2: total read during loops 27262976000
Quoted 8 lines
With patch:
/mnt/btrfs-test/reads/file2: total read during loops 18640457728/mnt/btrfs-test/reads/file1: total read during loops 9565376512This patch results in a 50% reduction of the amount of pages evicted from our working set.
Signed-off-by Josef Bacik <[email protected]>
To me this makes sense and I was myself wondering why we add dentries and inodes to the lru list with referenced bit set in the past. I would just note that the patch is changing the balance between dentry/inode cache eviction and page cache eviction (so that pages have lower chance of being evicted) and your test measures only the benefit of less pages being evicted. So to be fair you should also complement it with a test where dentries & inodes actually do get eventually reused and how that works out in presence of memory pressure from use-once pages in the page cache.
Honza
Quoted 51 lines
--- fs/dcache.c | 4 ++-- fs/inode.c | 3 ++- 2 files changed, 4 insertions(+), 3 deletions(-)fs/dcache.c +2
-2@@ -419,6 +419,8 @@ static void dentry_lru_add(struct dentry *dentry) {if (unlikely (!(dentry->d_flags &DCACHE_LRU_LIST )))d_lru_add (dentry); +else if (unlikely (!(dentry->d_flags &DCACHE_REFERENCED ))) + dentry->d_flags |=DCACHE_REFERENCED ; }/** @@ -779,8 +781,6 @@ void dput(struct dentry *dentry) goto kill_it; } -if (!(dentry->d_flags &DCACHE_REFERENCED )) - dentry->d_flags |=DCACHE_REFERENCED ;dentry_lru_add (dentry); dentry->d_lockref.count--;fs/inode.c +2
-1@@ -405,6 +405,8 @@ static void inode_lru_list_add(struct inode *inode) {if (list_lru_add (&inode->i_sb->s_inode_lru, &inode->i_lru))this_cpu_inc (nr_unused); +else + inode->i_state |=I_REFERENCED ; }/* @@ -1492,7 +1494,6 @@ static void iput_final(struct inode *inode) drop =generic_drop_inode (inode);if (!drop && (sb->s_flags &MS_ACTIVE )) { - inode->i_state |=I_REFERENCED ;inode_add_lru (inode);spin_unlock (&inode->i_lock);return ;Signature · 1 line
2.7.4
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Re: [PATCH] don't set *REFERENCED on single use objects
Quoted 43 lines · Jan Kara
Quoted 30 lines · Josef Bacik
By default we set
DCACHE_REFERENCEDandI_REFERENCEDon any dentry or inode we create. This is problematic as this means that it takes two trips through the LRU for any of these objects to be reclaimed, regardless of their actual lifetime. With enough pressure from these caches we can easily evict our working set from page cache with single use objects. So instead only set *REFERENCED if we've already been added to the LRU list. This means that we've been touched since the first time we were accessed, and so more likely to need to hang out in cache.To illustrate this issue I wrote the following scripts
https://github.com/josefbacik/debug-scripts/tree/master/cache-pressure
on my test box. It is a single socket 4 core CPU with 16gib of RAM and I tested on an Intel 2tib NVME drive. The
cache-pressure.shscript creates a new file system and creates 2 6.5gib files in order to take up 13gib of the 16gib of ram with pagecache. Then it runs a test program that reads these 2 files in a loop, and keeps track of how often it has to read bytes for each loop. On an ideal system with no pressure we should have to read 0 bytes indefinitely. The second thing this script does is start afs_markjob that creates a ton of 0 length files, putting pressure on the system with slab only allocations. On exit the script prints out how many bytes were read by the read-file program. The results are as followsWithout patch:
/mnt/btrfs-test/reads/file1: total read during loops 27262988288/mnt/btrfs-test/reads/file2: total read during loops 27262976000Quoted 8 lines
With patch:
/mnt/btrfs-test/reads/file2: total read during loops 18640457728/mnt/btrfs-test/reads/file1: total read during loops 9565376512This patch results in a 50% reduction of the amount of pages evicted from our working set.
Signed-off-by Josef Bacik <[email protected]>
Quoted 9 lines
To me this makes sense and I was myself wondering why we add dentries and inodes to the lru list with referenced bit set in the past. I would just note that the patch is changing the balance between dentry/inode cache eviction and page cache eviction (so that pages have lower chance of being evicted) and your test measures only the benefit of less pages being evicted. So to be fair you should also complement it with a test where dentries & inodes actually do get eventually reused and how that works out in presence of memory pressure from use-once pages in the page cache.
Yup that's fair. I wrote up a different script to exercise the warm icache/dcache vs once use pages, the scripts are in the same place as my original script
https://github.com/josefbacik/debug-scripts/tree/master/cache-pressure
Basically create 3million empty nodes, create a 100gib sparse file, stat in a loop all the files while cat'ing the sparse file to /dev/null. There's a bcc script in there to count how many times the stat'ing programs allocate a new inode, and with and without my patch the test exited with those processes creating 0 nodes. This makes sense as we will create the files and they will not have REFERENCED, and then as soon as the files are read in again they will be marked with REFERENCED, so they become just as hard to remove as they were originally. Thanks,
Josef
Re: [PATCH] don't set *REFERENCED on single use objects
Quoted 60 lines · Josef Bacik
Quoted 43 lines · Jan Kara
Quoted 30 lines · Josef Bacik
By default we set
DCACHE_REFERENCEDandI_REFERENCEDon any dentry or inode we create. This is problematic as this means that it takes two trips through the LRU for any of these objects to be reclaimed, regardless of their actual lifetime. With enough pressure from these caches we can easily evict our working set from page cache with single use objects. So instead only set *REFERENCED if we've already been added to the LRU list. This means that we've been touched since the first time we were accessed, and so more likely to need to hang out in cache.To illustrate this issue I wrote the following scripts
https://github.com/josefbacik/debug-scripts/tree/master/cache-pressure
on my test box. It is a single socket 4 core CPU with 16gib of RAM and I tested on an Intel 2tib NVME drive. The
cache-pressure.shscript creates a new file system and creates 2 6.5gib files in order to take up 13gib of the 16gib of ram with pagecache. Then it runs a test program that reads these 2 files in a loop, and keeps track of how often it has to read bytes for each loop. On an ideal system with no pressure we should have to read 0 bytes indefinitely. The second thing this script does is start afs_markjob that creates a ton of 0 length files, putting pressure on the system with slab only allocations. On exit the script prints out how many bytes were read by the read-file program. The results are as followsWithout patch:
/mnt/btrfs-test/reads/file1: total read during loops 27262988288/mnt/btrfs-test/reads/file2: total read during loops 27262976000Quoted 8 lines
With patch:
/mnt/btrfs-test/reads/file2: total read during loops 18640457728/mnt/btrfs-test/reads/file1: total read during loops 9565376512This patch results in a 50% reduction of the amount of pages evicted from our working set.
Signed-off-by Josef Bacik <[email protected]>Quoted 10 lines
To me this makes sense and I was myself wondering why we add dentries and inodes to the lru list with referenced bit set in the past. I would just note that the patch is changing the balance between dentry/inode cache eviction and page cache eviction (so that pages have lower chance of being evicted) and your test measures only the benefit of less pages being evicted. So to be fair you should also complement it with a test where dentries & inodes actually do get eventually reused and how that works out in presence of memory pressure from use-once pages in the page cache.
Yup that's fair. I wrote up a different script to exercise the warm icache/dcache vs once use pages, the scripts are in the same place as my original script
https://github.com/josefbacik/debug-scripts/tree/master/cache-pressure
Quoted 8 lines
Basically create 3million empty nodes, create a 100gib sparse file, stat in a loop all the files while cat'ing the sparse file to
/dev/null. There's a bcc script in there to count how many times the stat'ing programs allocate a new inode, and with and without my patch the test exited with those processes creating 0 nodes. This makes sense as we will create the files and they will not have REFERENCED, and then as soon as the files are read in again they will be marked with REFERENCED, so they become just as hard to remove as they were originally.
OK, sounds good. Feel free to add:
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
Quoted 30 lines · Josef Bacik
By default we set
DCACHE_REFERENCEDandI_REFERENCEDon any dentry or inode we create. This is problematic as this means that it takes two trips through the LRU for any of these objects to be reclaimed, regardless of their actual lifetime. With enough pressure from these caches we can easily evict our working set from page cache with single use objects. So instead only set *REFERENCED if we've already been added to the LRU list. This means that we've been touched since the first time we were accessed, and so more likely to need to hang out in cache.To illustrate this issue I wrote the following scripts
https://github.com/josefbacik/debug-scripts/tree/master/cache-pressure
on my test box. It is a single socket 4 core CPU with 16gib of RAM and I tested on an Intel 2tib NVME drive. The
cache-pressure.shscript creates a new file system and creates 2 6.5gib files in order to take up 13gib of the 16gib of ram with pagecache. Then it runs a test program that reads these 2 files in a loop, and keeps track of how often it has to read bytes for each loop. On an ideal system with no pressure we should have to read 0 bytes indefinitely. The second thing this script does is start afs_markjob that creates a ton of 0 length files, putting pressure on the system with slab only allocations. On exit the script prints out how many bytes were read by the read-file program. The results are as followsWithout patch:
/mnt/btrfs-test/reads/file1: total read during loops 27262988288/mnt/btrfs-test/reads/file2: total read during loops 27262976000
Quoted 8 lines
With patch:
/mnt/btrfs-test/reads/file2: total read during loops 18640457728/mnt/btrfs-test/reads/file1: total read during loops 9565376512This patch results in a 50% reduction of the amount of pages evicted from our working set.
Signed-off-by Josef Bacik <[email protected]>
Hello Al,
Can we get this included in the upcoming merge window? Thanks,
Josef