btrfs: remove the metadata readahead mechanism
cover btrfs: remove the metadata readahead mechanism
This is originally just my preparation for scrub refactors, but when the readahead is involved, it won't just be a small cleanup.
The metadata readahead code is introduced in 2011 (surprisingly, the commit message even contains changelog), but now only one user for it, and even for the only one user, the readahead mechanism can't provide much help in fact.
Scrub needs readahead for commit root, but the existing one can only do current root readahead.
And the code is at a very bad layer inside btrfs, all metadata are at btrfs logical address space, but the readahead is kinda working at device layer (to manage the in-flight readahead).
Personally speaking, I don't think such "optimization" is really even needed, since we have better way like setting IO priority.
I really prefer to let the professional block layer guys do whatever they are good at (and in fact, those block layer guys rock!). Immature optimization is the cause of bugs, and it has already caused several bugs recently.
Nowadays we have btrfs_path::reada to do the readahead, I doubt if we really need such facility.
So here I purpose to completely remove the old and under utilized metadata readahead system.
Qu Wenruo (2):
btrfs: remove the unnecessary path parameter for scrub_raid56_parity()
btrfs: remove reada mechanism
fs/btrfs/Makefile | 2 +- fs/btrfs/ctree.h | 25 - fs/btrfs/dev-replace.c | 5 - fs/btrfs/disk-io.c | 20 +- fs/btrfs/extent_io.c | 3 - fs/btrfs/reada.c | 1086 ---------------------------------------- fs/btrfs/scrub.c | 64 +-- fs/btrfs/super.c | 1 - fs/btrfs/volumes.c | 7 - fs/btrfs/volumes.h | 7 - 10 files changed, 17 insertions(+), 1203 deletions(-) delete mode 100644 fs/btrfs/reada.c
Signature · 1 line
2.34.1
RFC PATCH 1/2 btrfs: remove the unnecessary path parameter for scrub_raid56_parity()
In function scrub_stripe() we allocated two btrfs_path, one @path for extent tree search and another @ppath for full stripe extent tree search for RAID56.
This is totally uncessary, as the @ppath usage is completely inside scrub_raid56_parity(), thus we can move the path allocation into scrub_raid56_parity() completely.
--- fs/btrfs/scrub.c | 28 ++++++++++++++-------------- 1 file changed, 14 insertions(+), 14 deletions(-)
fs/btrfs/scrub.c +14 -14
@@ -2892,7 +2892,6 @@ static void scrub_parity_put(struct scrub_parity *sparity) static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,struct map_lookup *map,struct btrfs_device *sdev, -struct btrfs_path *path,u64 logic_start,u64 logic_end) {
@@ -2901,6 +2900,7 @@ static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,struct btrfs_root *csum_root;struct btrfs_extent_item *extent;struct btrfs_io_context *bioc =NULL ; +struct btrfs_path *path;u64 flags;int ret;int slot;
@@ -2919,6 +2919,14 @@ static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,int extent_mirror_num;int stop_loop =0 ; + path =btrfs_alloc_path (); +if (!path) { +spin_lock (&sctx->stat_lock); + sctx->stat.malloc_errors++; +spin_unlock (&sctx->stat_lock); +return -ENOMEM ; + } +ASSERT (map->stripe_len <=U32_MAX ); nsectors = map->stripe_len >> fs_info->sectorsize_bits; bitmap_len =scrub_calc_parity_bitmap_len (nsectors);
@@ -2928,6 +2936,7 @@ static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,spin_lock (&sctx->stat_lock); sctx->stat.malloc_errors++;spin_unlock (&sctx->stat_lock); +btrfs_free_path (path);return -ENOMEM ; }
@@ -3117,7 +3126,7 @@ static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,scrub_wr_submit (sctx);mutex_unlock (&sctx->wr_lock); -btrfs_release_path (path); +btrfs_free_path (path);return ret <0 ? ret :0 ; }
@@ -3167,7 +3176,7 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx, int num, u64 base, u64 length, struct btrfs_block_group *cache) { - struct btrfs_path *path, *ppath; + struct btrfs_path *path;struct btrfs_fs_info *fs_info = sctx->fs_info;struct btrfs_root *root;struct btrfs_root *csum_root;
@@ -3229,12 +3238,6 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,if (!path)return -ENOMEM ; - ppath =btrfs_alloc_path (); -if (!ppath) { -btrfs_free_path (path); -return -ENOMEM ; - } -/* * work on commit root. The related disk blocks are static as * long as COW is applied. This means, it is save to rewrite
@@ -3243,8 +3246,6 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx, path->search_commit_root =1 ; path->skip_locking =1 ; - ppath->search_commit_root =1 ; - ppath->skip_locking =1 ;/* * trigger the readahead for extent tree csum tree and wait for * completion. During readahead, the scrub is officially paused
@@ -3347,7 +3348,7 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx, stripe_logical += base; stripe_end = stripe_logical + increment; ret =scrub_raid56_parity (sctx, map, scrub_dev, - ppath, stripe_logical, + stripe_logical, stripe_end);if (ret) goto out;
@@ -3518,7 +3519,7 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx, stripe_end = stripe_logical + increment; ret =scrub_raid56_parity (sctx, - map, scrub_dev, ppath, + map, scrub_dev, stripe_logical, stripe_end);if (ret)
@@ -3565,7 +3566,6 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,blk_finish_plug (&plug);btrfs_free_path (path); -btrfs_free_path (ppath);if (sctx->is_dev_replace && ret >=0 ) {int ret2;
Signature · 1 line
2.34.1
RFC PATCH 2/2 btrfs: remove reada mechanism
In reply to Qu WenruoCurrently there is only one user for btrfs metadata readahead, and that's scrub.
But even for the single user, it's not providing the correct functionality it needs, as scrub needs reada for commit root, which current readahead can't provide. (Although it's pretty easy to add such feature).
Despite this, there are some extra problems related to metadata readahead:
- Duplicated feature with btrfs_path::reada
- Partly duplicated feature of btrfs_fs_info::buffer_radix Btrfs already caches its metadata in buffer_radix, while readahead tries to read the tree block no matter if it's already cached.
- Poor layer separation
To my surprise, metadata readahead works kinda at device level. This is definitely not the correct layer it should be, since metadata is at btrfs logical address space, it should not bother device at all.
This brings extra chance for bugs to sneak in, while brings unnecessary complexity.
- Dead codes
In the very beginning of scrub.c we have #undef DEBUG, rendering all the debug related code useless and unable to test.
Thus here I purpose to remove the metadata readahead mechanism completely.
--- fs/btrfs/Makefile | 2 +- fs/btrfs/ctree.h | 25 - fs/btrfs/dev-replace.c | 5 - fs/btrfs/disk-io.c | 20 +- fs/btrfs/extent_io.c | 3 - fs/btrfs/reada.c | 1086 ---------------------------------------- fs/btrfs/scrub.c | 36 -- fs/btrfs/super.c | 1 - fs/btrfs/volumes.c | 7 - fs/btrfs/volumes.h | 7 - 10 files changed, 3 insertions(+), 1189 deletions(-) delete mode 100644 fs/btrfs/reada.c
fs/btrfs/Makefile +1 -1
@@ -27,7 +27,7 @@ btrfs-y += super.o ctree.o extent-tree.o print-tree.o root-tree.o dir-item.o \ extent_io.o volumes.o async-thread.o ioctl.o locking.o orphan.o \ export.o tree-log.o free-space-cache.o zlib.o lzo.o zstd.o \ compression.o delayed-ref.o relocation.o delayed-inode.o scrub.o \ - reada.o backref.o ulist.o qgroup.o send.o dev-replace.o raid56.o \ + backref.o ulist.o qgroup.o send.o dev-replace.o raid56.o \ uuid-tree.o props.o free-space-tree.o tree-checker.o space-info.o \ block-rsv.o delalloc-space.o block-group.o discard.o reflink.o \ subpage.o tree-mod-log.o
fs/btrfs/ctree.h +0 -25
@@ -821,7 +821,6 @@ struct btrfs_fs_info {struct btrfs_workqueue *endio_write_workers;struct btrfs_workqueue *endio_freespace_worker;struct btrfs_workqueue *caching_workers; -struct btrfs_workqueue *readahead_workers;/* * fixup workers take dirty pages that didn't properly go through
@@ -958,13 +957,6 @@ struct btrfs_fs_info {struct btrfs_delayed_root *delayed_root; -/* readahead tree */ -spinlock_t reada_lock; -struct radix_tree_root reada_tree; - -/* readahead works cnt */ -atomic_t reada_works_cnt; -/* Extent buffer radix tree */ spinlock_t buffer_lock;/* Entries are eb->start / sectorsize */
@@ -3836,23 +3828,6 @@ static inline void btrfs_bio_counter_dec(struct btrfs_fs_info *fs_info)btrfs_bio_counter_sub (fs_info,1 ); } -/* reada.c */ -struct reada_control { -struct btrfs_fs_info *fs_info;/* tree to prefetch */ -struct btrfs_key key_start; -struct btrfs_key key_end;/* exclusive */ -atomic_t elems; -struct kref refcnt; -wait_queue_head_t wait; -}; -struct reada_control *btrfs_reada_add (struct btrfs_root *root, -struct btrfs_key *start,struct btrfs_key *end); -int btrfs_reada_wait (void *handle); -void btrfs_reada_detach (void *handle); -int btree_readahead_hook (struct extent_buffer *eb,int err); -void btrfs_reada_remove_dev (struct btrfs_device *dev); -void btrfs_reada_undo_remove_dev (struct btrfs_device *dev); -static inline int is_fstree (u64 rootid) { if (rootid == BTRFS_FS_TREE_OBJECTID ||
fs/btrfs/dev-replace.c +0 -5
@@ -906,9 +906,6 @@ static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info, }btrfs_wait_ordered_roots (fs_info,U64_MAX ,0 , (u64)-1 ); -if (!scrub_ret) -btrfs_reada_remove_dev (src_device); -/* * We have to use this loop approach because at this point src_device * has to be available for transaction commit to complete, yet new
@@ -917,7 +914,6 @@ static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,while (1 ) { trans =btrfs_start_transaction (root,0 );if (IS_ERR (trans)) { -btrfs_reada_undo_remove_dev (src_device);mutex_unlock (&dev_replace->lock_finishing_cancel_unmount);return PTR_ERR (trans); }
@@ -968,7 +964,6 @@ static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,up_write (&dev_replace->rwsem);mutex_unlock (&fs_info->chunk_mutex);mutex_unlock (&fs_info->fs_devices->device_list_mutex); -btrfs_reada_undo_remove_dev (src_device);btrfs_rm_dev_replace_blocked (fs_info);if (tgt_device)btrfs_destroy_dev_replace_tgtdev (tgt_device);
fs/btrfs/disk-io.c +2 -18
@@ -665,9 +665,6 @@ static int validate_subpage_buffer(struct page *page, u64 start, u64 end,if (ret <0 ) goto err; -if (test_and_clear_bit (EXTENT_BUFFER_READAHEAD , &eb->bflags)) -btree_readahead_hook (eb, ret); -set_extent_buffer_uptodate (eb);free_extent_buffer (eb);
@@ -715,10 +712,6 @@ int btrfs_validate_metadata_buffer(struct btrfs_bio *bbio, } ret =validate_extent_buffer (eb); err: -if (reads_done && -test_and_clear_bit (EXTENT_BUFFER_READAHEAD , &eb->bflags)) -btree_readahead_hook (eb, ret); -if (ret) {/* * our io error hook is going to dec the io pages
@@ -2224,7 +2217,6 @@ static void btrfs_stop_all_workers(struct btrfs_fs_info *fs_info)btrfs_destroy_workqueue (fs_info->endio_freespace_worker);btrfs_destroy_workqueue (fs_info->delayed_workers);btrfs_destroy_workqueue (fs_info->caching_workers); -btrfs_destroy_workqueue (fs_info->readahead_workers);btrfs_destroy_workqueue (fs_info->flush_workers);btrfs_destroy_workqueue (fs_info->qgroup_rescan_workers);if (fs_info->discard_ctl.discard_workers)
@@ -2437,9 +2429,6 @@ static int btrfs_init_workqueues(struct btrfs_fs_info *fs_info) fs_info->delayed_workers =btrfs_alloc_workqueue (fs_info,"delayed-meta" , flags, max_active,0 ); - fs_info->readahead_workers = -btrfs_alloc_workqueue (fs_info,"readahead" , flags, - max_active,2 ); fs_info->qgroup_rescan_workers =btrfs_alloc_workqueue (fs_info,"qgroup-rescan" , flags,1 ,0 ); fs_info->discard_ctl.discard_workers =
@@ -2451,9 +2440,8 @@ static int btrfs_init_workqueues(struct btrfs_fs_info *fs_info) fs_info->endio_meta_write_workers && fs_info->endio_write_workers && fs_info->endio_raid56_workers && fs_info->endio_freespace_worker && fs_info->rmw_workers && - fs_info->caching_workers && fs_info->readahead_workers && - fs_info->fixup_workers && fs_info->delayed_workers && - fs_info->qgroup_rescan_workers && + fs_info->caching_workers && fs_info->fixup_workers && + fs_info->delayed_workers && fs_info->qgroup_rescan_workers && fs_info->discard_ctl.discard_workers)) { return -ENOMEM; }
@@ -3083,7 +3071,6 @@ void btrfs_init_fs_info(struct btrfs_fs_info *fs_info)atomic_set (&fs_info->async_delalloc_pages,0 );atomic_set (&fs_info->defrag_running,0 ); -atomic_set (&fs_info->reada_works_cnt,0 );atomic_set (&fs_info->nr_delayed_iputs,0 );atomic64_set (&fs_info->tree_mod_seq,0 ); fs_info->global_root_tree =RB_ROOT ;
@@ -3094,9 +3081,6 @@ void btrfs_init_fs_info(struct btrfs_fs_info *fs_info) fs_info->tree_mod_log =RB_ROOT ; fs_info->commit_interval =BTRFS_DEFAULT_COMMIT_INTERVAL ; fs_info->avg_delayed_ref_runtime =NSEC_PER_SEC >>6 ;/* div by 64 */ -/* readahead state */ -INIT_RADIX_TREE (&fs_info->reada_tree,GFP_NOFS & ~__GFP_DIRECT_RECLAIM); -spin_lock_init (&fs_info->reada_lock);btrfs_init_ref_verify (fs_info); fs_info->thread_pool_size = min_t(unsigned long,
fs/btrfs/extent_io.c +0 -3
@@ -3087,9 +3087,6 @@ static void end_bio_extent_readpage(struct bio *bio)set_bit (EXTENT_BUFFER_READ_ERR , &eb->bflags); eb->read_mirror = mirror;atomic_dec (&eb->io_pages); -if (test_and_clear_bit (EXTENT_BUFFER_READAHEAD , - &eb->bflags)) -btree_readahead_hook (eb, -EIO ); } readpage_ok:if (likely (uptodate)) {
fs/btrfs/reada.c deleted +0 -1086
1,086 lines not on this page · raw message
fs/btrfs/scrub.c +0 -36
@@ -3193,10 +3193,7 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,u64 physical_end;u64 generation;int mirror_num; -struct reada_control *reada1; -struct reada_control *reada2;struct btrfs_key key; -struct btrfs_key key_end;u64 increment = map->stripe_len;u64 offset;u64 extent_logical;
@@ -3246,11 +3243,6 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx, path->search_commit_root =1 ; path->skip_locking =1 ; -/* -* trigger the readahead for extent tree csum tree and wait for -* completion. During readahead, the scrub is officially paused -* to not hold off transaction commits -*/ logical = base + offset; physical_end = physical + nstripes * map->stripe_len;if (map->type &BTRFS_BLOCK_GROUP_RAID56_MASK ) {
@@ -3265,36 +3257,8 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,scrub_blocked_if_needed (fs_info); root =btrfs_extent_root (fs_info, logical); - -/* FIXME it might be better to start readahead at commit root */ - key.objectid = logical; - key.type =BTRFS_EXTENT_ITEM_KEY ; - key.offset = (u64 )0 ; - key_end.objectid = logic_end; - key_end.type =BTRFS_METADATA_ITEM_KEY ; - key_end.offset = (u64)-1 ; - reada1 =btrfs_reada_add (root, &key, &key_end); - csum_root =btrfs_csum_root (fs_info, logical); -if (cache->flags &BTRFS_BLOCK_GROUP_DATA ) { - key.objectid =BTRFS_EXTENT_CSUM_OBJECTID ; - key.type =BTRFS_EXTENT_CSUM_KEY ; - key.offset = logical; - key_end.objectid =BTRFS_EXTENT_CSUM_OBJECTID ; - key_end.type =BTRFS_EXTENT_CSUM_KEY ; - key_end.offset = logic_end; - reada2 =btrfs_reada_add (csum_root, &key, &key_end); - }else { - reada2 =NULL ; - } - -if (!IS_ERR (reada1)) -btrfs_reada_wait (reada1); -if (!IS_ERR_OR_NULL (reada2)) -btrfs_reada_wait (reada2); - -/* * collect all data csums for the stripe to avoid seeking during * the scrub. This might currently (crc32) end up to be about 1MB
fs/btrfs/super.c +0 -1
@@ -1842,7 +1842,6 @@ static void btrfs_resize_thread_pool(struct btrfs_fs_info *fs_info,btrfs_workqueue_set_max (fs_info->endio_write_workers, new_pool_size);btrfs_workqueue_set_max (fs_info->endio_freespace_worker, new_pool_size);btrfs_workqueue_set_max (fs_info->delayed_workers, new_pool_size); -btrfs_workqueue_set_max (fs_info->readahead_workers, new_pool_size);btrfs_workqueue_set_max (fs_info->scrub_wr_completion_workers, new_pool_size); }
fs/btrfs/volumes.c +0 -7
@@ -1166,7 +1166,6 @@ static void btrfs_close_one_device(struct btrfs_device *device)ASSERT (!test_bit (BTRFS_DEV_STATE_REPLACE_TGT , &device->dev_state));ASSERT (list_empty (&device->dev_alloc_list));ASSERT (list_empty (&device->post_commit_list)); -ASSERT (atomic_read (&device->reada_in_flight) ==0 ); }static void close_fs_devices (struct btrfs_fs_devices *fs_devices)
@@ -2148,8 +2147,6 @@ int btrfs_rm_device(struct btrfs_fs_info *fs_info, } ret =btrfs_shrink_device (device,0 ); -if (!ret) -btrfs_reada_remove_dev (device);if (ret) goto error_undo;
@@ -2247,7 +2244,6 @@ int btrfs_rm_device(struct btrfs_fs_info *fs_info,return ret; error_undo: -btrfs_reada_undo_remove_dev (device);if (test_bit (BTRFS_DEV_STATE_WRITEABLE , &device->dev_state)) {mutex_lock (&fs_info->chunk_mutex); list_add(&device->dev_alloc_list,
@@ -6978,11 +6974,8 @@ struct btrfs_device *btrfs_alloc_device(struct btrfs_fs_info *fs_info,INIT_LIST_HEAD (&dev->dev_alloc_list);INIT_LIST_HEAD (&dev->post_commit_list); -atomic_set (&dev->reada_in_flight,0 );atomic_set (&dev->dev_stats_ccnt,0 );btrfs_device_data_ordered_init (dev); -INIT_RADIX_TREE (&dev->reada_zones,GFP_NOFS & ~__GFP_DIRECT_RECLAIM); -INIT_RADIX_TREE (&dev->reada_extents,GFP_NOFS & ~__GFP_DIRECT_RECLAIM);extent_io_tree_init (fs_info, &dev->alloc_state,IO_TREE_DEVICE_ALLOC_STATE ,NULL );
fs/btrfs/volumes.h +0 -7
@@ -123,13 +123,6 @@ struct btrfs_device {/* per-device scrub information */ struct scrub_ctx *scrub_ctx; -/* readahead state */ -atomic_t reada_in_flight; -u64 reada_next; -struct reada_zone *reada_curr_zone; -struct radix_tree_root reada_zones; -struct radix_tree_root reada_extents; -/* disk I/O failure stats. For detailed description refer to * enum btrfs_dev_stat_values in ioctl.h */ int dev_stats_valid;
Signature · 1 line
2.34.1
Quoted 10 lines · Qu Wenruo
This is originally just my preparation for scrub refactors, but when the readahead is involved, it won't just be a small cleanup.
The metadata readahead code is introduced in 2011 (surprisingly, the commit message even contains changelog), but now only one user for it, and even for the only one user, the readahead mechanism can't provide much help in fact.
Scrub needs readahead for commit root, but the existing one can only do current root readahead.
If support for the commit root is added, is there a noticeable speedup? Have you tested that?
Quoted 7 lines
And the code is at a very bad layer inside btrfs, all metadata are at btrfs logical address space, but the readahead is kinda working at device layer (to manage the in-flight readahead).
Personally speaking, I don't think such "optimization" is really even needed, since we have better way like setting IO priority.
Have you done any benchmarks?
How? On physical machines or VMs?
Please include such details in the changelogs.
Quoted 8 lines
I really prefer to let the professional block layer guys do whatever they are good at (and in fact, those block layer guys rock!). Immature optimization is the cause of bugs, and it has already caused several bugs recently.
Nowadays we have
btrfs_path::readato do the readahead, I doubt if we really need such facility.
btrfs_path:reada is important and it makes a difference.
I recently changed send to use it, and benchmarks can be found in the changelogs.
There are also other places where it makes a difference, such as when reading a large chunk tree during mount or when reading a large directory.
It's all about reading other leaves/nodes in the background that will be needed in the near future while the task is doing something else. Even if the nodes/leaves are not physically contiguous on disk (that's the main reason why the mechanism exists).
Quoted 24 lines
So here I purpose to completely remove the old and under utilized metadata readahead system.
Qu Wenruo (2):
btrfs: remove the unnecessary path parameter forscrub_raid56_parity()
btrfs: remove reada mechanismfs/btrfs/Makefile | 2 +- fs/btrfs/ctree.h | 25 - fs/btrfs/dev-replace.c | 5 - fs/btrfs/disk-io.c | 20 +- fs/btrfs/extent_io.c | 3 - fs/btrfs/reada.c | 1086 ---------------------------------------- fs/btrfs/scrub.c | 64 +-- fs/btrfs/super.c | 1 - fs/btrfs/volumes.c | 7 - fs/btrfs/volumes.h | 7 - 10 files changed, 17 insertions(+), 1203 deletions(-) delete mode 100644 fs/btrfs/reada.cSignature · 1 line
2.34.1
Quoted 9 lines · Filipe Manana
Quoted 8 lines · Qu Wenruo
This is originally just my preparation for scrub refactors, but when the readahead is involved, it won't just be a small cleanup.
The metadata readahead code is introduced in 2011 (surprisingly, the commit message even contains changelog), but now only one user for it, and even for the only one user, the readahead mechanism can't provide much help in fact.
Quoted 6 lines
Quoted 2 lines
Scrub needs readahead for commit root, but the existing one can only do current root readahead.
If support for the commit root is added, is there a noticeable speedup? Have you tested that?
Will craft a benchmark for that.
Although I don't have any HDD available for benchmark, thus would only have result from SATA SSD.
Quoted 19 lines
Quoted 7 lines
And the code is at a very bad layer inside btrfs, all metadata are at btrfs logical address space, but the readahead is kinda working at device layer (to manage the in-flight readahead).
Personally speaking, I don't think such "optimization" is really even needed, since we have better way like setting IO priority.
Have you done any benchmarks?
How? On physical machines or VMs?Please include such details in the changelogs.
Quoted 6 lines
I really prefer to let the professional block layer guys do whatever they are good at (and in fact, those block layer guys rock!). Immature optimization is the cause of bugs, and it has already caused several bugs recently.
Quoted 6 lines
Quoted 2 lines
Nowadays we have
btrfs_path::readato do the readahead, I doubt if we really need such facility.btrfs_path:reada is important and it makes a difference.
I recently changed send to use it, and benchmarks can be found in the changelogs.
For the "such facility" I mean the btrfs_reada_add() facility, not the btrfs_path::reada one.
Quoted 8 lines
There are also other places where it makes a difference, such as when reading a large chunk tree during mount or when reading a large directory.
It's all about reading other leaves/nodes in the background that will be needed in the near future while the task is doing something else. Even if the nodes/leaves are not physically contiguous on disk (that's the main reason why the mechanism exists).
Unfortunately, not really in the background.
For scrub usage, it kicks readahead and wait for it, not really doing it in the background.
(Nor btrfs_path::reada either though, btrfs_path reada also happens at tree search time, and it's synchronous).
Another reason why the existing btrfs_reada_add() facility is not suitable for scrub is, our default tree block size is way larger than the scrub data length.
The current data length is 16 pages (64K), while even one 16K leaf can contain at least csum for 8M (CRC32) or 1M (SHA256).
This means for most readahead, it doesn't make much sense as it won't cross leaf boundaries that frequently.
(BTW, in this particular case, btrfs_path::reada may perform better than the start/end based reada, as that would really do some readahead)
Anyway, only benchmark can prove whether I'm correct or wrong.
Thanks,
Qu
Quoted 25 lines
Quoted 25 lines
So here I purpose to completely remove the old and under utilized metadata readahead system.
Qu Wenruo (2):
btrfs: remove the unnecessary path parameter forscrub_raid56_parity()
btrfs: remove reada mechanismfs/btrfs/Makefile | 2 +- fs/btrfs/ctree.h | 25 - fs/btrfs/dev-replace.c | 5 - fs/btrfs/disk-io.c | 20 +- fs/btrfs/extent_io.c | 3 - fs/btrfs/reada.c | 1086 ---------------------------------------- fs/btrfs/scrub.c | 64 +-- fs/btrfs/super.c | 1 - fs/btrfs/volumes.c | 7 - fs/btrfs/volumes.h | 7 - 10 files changed, 17 insertions(+), 1203 deletions(-)delete mode 100644
fs/btrfs/reada.c--
2.34.1
Quoted 62 lines · Qu Wenruo
Quoted 9 lines · Filipe Manana
Quoted 8 lines · Qu Wenruo
This is originally just my preparation for scrub refactors, but when the readahead is involved, it won't just be a small cleanup.
The metadata readahead code is introduced in 2011 (surprisingly, the commit message even contains changelog), but now only one user for it, and even for the only one user, the readahead mechanism can't provide much help in fact.
Quoted 6 lines
Quoted 2 lines
Scrub needs readahead for commit root, but the existing one can only do current root readahead.
If support for the commit root is added, is there a noticeable speedup? Have you tested that?
Will craft a benchmark for that.
Although I don't have any HDD available for benchmark, thus would only have result from SATA SSD.
Quoted 19 lines
Quoted 7 lines
And the code is at a very bad layer inside btrfs, all metadata are at btrfs logical address space, but the readahead is kinda working at device layer (to manage the in-flight readahead).
Personally speaking, I don't think such "optimization" is really even needed, since we have better way like setting IO priority.
Have you done any benchmarks?
How? On physical machines or VMs?Please include such details in the changelogs.
Quoted 6 lines
I really prefer to let the professional block layer guys do whatever they are good at (and in fact, those block layer guys rock!). Immature optimization is the cause of bugs, and it has already caused several bugs recently.
Quoted 6 lines
Quoted 2 lines
Nowadays we have
btrfs_path::readato do the readahead, I doubt if we really need such facility.btrfs_path:reada is important and it makes a difference.
I recently changed send to use it, and benchmarks can be found in the changelogs.For the "such facility" I mean the
btrfs_reada_add()facility, not thebtrfs_path::readaone.Quoted 8 lines
There are also other places where it makes a difference, such as when reading a large chunk tree during mount or when reading a large directory.
It's all about reading other leaves/nodes in the background that will be needed in the near future while the task is doing something else. Even if the nodes/leaves are not physically contiguous on disk (that's the main reason why the mechanism exists).
Quoted 7 lines
Unfortunately, not really in the background.
For scrub usage, it kicks readahead and wait for it, not really doing it in the background.
(Nor
btrfs_path::readaeither though,btrfs_pathreada also happens at tree search time, and it's synchronous).
No, the btrfs_path::reada mechanism is not synchronous - it does not wait for the reads (IO) to complete.
btrfs_readahead_node_child() triggers a read for the extent buffer's pages but does not wait for the reads to complete. I.e. we end up calling:
read_extent_buffer_pages(eb, WAIT_NONE, 0);
So it does not wait on the read bios to complete.
Otherwise that would be pointless.
Quoted 10 lines
Another reason why the existing
btrfs_reada_add()facility is not suitable for scrub is, our default tree block size is way larger than the scrub data length.The current data length is 16 pages (64K), while even one 16K leaf can contain at least csum for 8M (CRC32) or 1M (SHA256).
This means for most readahead, it doesn't make much sense as it won't cross leaf boundaries that frequently.
Quoted 4 lines
(BTW, in this particular case,
btrfs_path::readamay perform better than the start/end based reada, as that would really do some readahead)Anyway, only benchmark can prove whether I'm correct or wrong.
Yep, and preferably on a spinning disk and bare metal (no VM).
Thanks.
Quoted 28 lines
Thanks,
QuQuoted 25 lines
Quoted 25 lines
So here I purpose to completely remove the old and under utilized metadata readahead system.
Qu Wenruo (2):
btrfs: remove the unnecessary path parameter forscrub_raid56_parity()
btrfs: remove reada mechanismfs/btrfs/Makefile | 2 +- fs/btrfs/ctree.h | 25 - fs/btrfs/dev-replace.c | 5 - fs/btrfs/disk-io.c | 20 +- fs/btrfs/extent_io.c | 3 - fs/btrfs/reada.c | 1086 ---------------------------------------- fs/btrfs/scrub.c | 64 +-- fs/btrfs/super.c | 1 - fs/btrfs/volumes.c | 7 - fs/btrfs/volumes.h | 7 - 10 files changed, 17 insertions(+), 1203 deletions(-)delete mode 100644
fs/btrfs/reada.c--
2.34.1
Quoted 74 lines · Filipe Manana
Quoted 62 lines · Qu Wenruo
Quoted 9 lines · Filipe Manana
Quoted 8 lines · Qu Wenruo
This is originally just my preparation for scrub refactors, but when the readahead is involved, it won't just be a small cleanup.
The metadata readahead code is introduced in 2011 (surprisingly, the commit message even contains changelog), but now only one user for it, and even for the only one user, the readahead mechanism can't provide much help in fact.
Quoted 6 lines
Quoted 2 lines
Scrub needs readahead for commit root, but the existing one can only do current root readahead.
If support for the commit root is added, is there a noticeable speedup? Have you tested that?
Will craft a benchmark for that.
Although I don't have any HDD available for benchmark, thus would only have result from SATA SSD.
Quoted 19 lines
Quoted 7 lines
And the code is at a very bad layer inside btrfs, all metadata are at btrfs logical address space, but the readahead is kinda working at device layer (to manage the in-flight readahead).
Personally speaking, I don't think such "optimization" is really even needed, since we have better way like setting IO priority.
Have you done any benchmarks?
How? On physical machines or VMs?Please include such details in the changelogs.
Quoted 6 lines
I really prefer to let the professional block layer guys do whatever they are good at (and in fact, those block layer guys rock!). Immature optimization is the cause of bugs, and it has already caused several bugs recently.
Quoted 6 lines
Quoted 2 lines
Nowadays we have
btrfs_path::readato do the readahead, I doubt if we really need such facility.btrfs_path:reada is important and it makes a difference.
I recently changed send to use it, and benchmarks can be found in the changelogs.For the "such facility" I mean the
btrfs_reada_add()facility, not thebtrfs_path::readaone.Quoted 8 lines
There are also other places where it makes a difference, such as when reading a large chunk tree during mount or when reading a large directory.
It's all about reading other leaves/nodes in the background that will be needed in the near future while the task is doing something else. Even if the nodes/leaves are not physically contiguous on disk (that's the main reason why the mechanism exists).
Quoted 7 lines
Unfortunately, not really in the background.
For scrub usage, it kicks readahead and wait for it, not really doing it in the background.
(Nor
btrfs_path::readaeither though,btrfs_pathreada also happens at tree search time, and it's synchronous).No, the
btrfs_path::readamechanism is not synchronous - it does not wait for the reads (IO) to complete.
Quoted 7 lines
btrfs_readahead_node_child()triggers a read for the extent buffer's pages but does not wait for the reads to complete. I.e. we end up calling:read_extent_buffer_pages(eb, WAIT_NONE, 0);So it does not wait on the read bios to complete.
Otherwise that would be pointless.
Oh, forgot the WAIT_NONE part...
Then to be honest, there is even less meaning to have btrfs_reada_add() facility, the path reada is better than that.
Quoted 11 lines
Quoted 11 lines
Another reason why the existing
btrfs_reada_add()facility is not suitable for scrub is, our default tree block size is way larger than the scrub data length.The current data length is 16 pages (64K), while even one 16K leaf can contain at least csum for 8M (CRC32) or 1M (SHA256).
This means for most readahead, it doesn't make much sense as it won't cross leaf boundaries that frequently.
Quoted 6 lines
Quoted 4 lines
(BTW, in this particular case,
btrfs_path::readamay perform better than the start/end based reada, as that would really do some readahead)Anyway, only benchmark can prove whether I'm correct or wrong.
Yep, and preferably on a spinning disk and bare metal (no VM).
Unfortunately, that's what I don't have...
Physical machines either is the main ITX desktop I use for development and VM testing, or laptops without SATA interface for HDD.
And even worse, I don't have any HDD available for testing yet. (The only HDDs are utilized by my NAS)
I'll go for VM with SATA SSD (directly passed to VM) first, then try to get a physical machine (maybe NUC with 2.5in HDD then).
Thanks,
Qu
Quoted 32 lines
Thanks.
Quoted 28 lines
Thanks,
QuQuoted 25 lines
Quoted 25 lines
So here I purpose to completely remove the old and under utilized metadata readahead system.
Qu Wenruo (2):
btrfs: remove the unnecessary path parameter forscrub_raid56_parity()
btrfs: remove reada mechanismfs/btrfs/Makefile | 2 +- fs/btrfs/ctree.h | 25 - fs/btrfs/dev-replace.c | 5 - fs/btrfs/disk-io.c | 20 +- fs/btrfs/extent_io.c | 3 - fs/btrfs/reada.c | 1086 ---------------------------------------- fs/btrfs/scrub.c | 64 +-- fs/btrfs/super.c | 1 - fs/btrfs/volumes.c | 7 - fs/btrfs/volumes.h | 7 - 10 files changed, 17 insertions(+), 1203 deletions(-)delete mode 100644
fs/btrfs/reada.c--
2.34.1
Quoted 15 lines · Qu Wenruo
Quoted 14 lines · Filipe Manana
Quoted 13 lines · Qu Wenruo
Quoted 12 lines · Filipe Manana
Quoted 10 lines · Qu Wenruo
This is originally just my preparation for scrub refactors, but when the readahead is involved, it won't just be a small cleanup.
The metadata readahead code is introduced in 2011 (surprisingly, the commit message even contains changelog), but now only one user for it, and even for the only one user, the readahead mechanism can't provide much help in fact.
Scrub needs readahead for commit root, but the existing one can only do current root readahead.
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 2 lines
If support for the commit root is added, is there a noticeable speedup? Have you tested that?
Will craft a benchmark for that.
Although I don't have any HDD available for benchmark, thus would only have result from SATA SSD.
I'm doing some tests, in a VM on a dedicated HDD.
Quoted 19 lines · David Sterba
Quoted 18 lines · Qu Wenruo
Quoted 17 lines · Filipe Manana
Quoted 16 lines · Qu Wenruo
Quoted 14 lines · Filipe Manana
Quoted 10 lines · Qu Wenruo
This is originally just my preparation for scrub refactors, but when the readahead is involved, it won't just be a small cleanup.
The metadata readahead code is introduced in 2011 (surprisingly, the commit message even contains changelog), but now only one user for it, and even for the only one user, the readahead mechanism can't provide much help in fact.
Scrub needs readahead for commit root, but the existing one can only do current root readahead.
If support for the commit root is added, is there a noticeable speedup? Have you tested that?
Quoted 6 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Will craft a benchmark for that.
Although I don't have any HDD available for benchmark, thus would only have result from SATA SSD.
I'm doing some tests, in a VM on a dedicated HDD.
There's some measurable difference:
With readahead:
Duration: 0:00:20 Total to scrub: 7.02GiB Rate: 236.92MiB/s
Duration: 0:00:48 Total to scrub: 12.02GiB Rate: 198.02MiB/s
Without readahead:
Duration: 0:00:22 Total to scrub: 7.02GiB Rate: 215.10MiB/s
Duration: 0:00:50 Total to scrub: 12.02GiB Rate: 190.66MiB/s
The setup is: data/single, metadata/dup, no-holes, free-space-tree, there are 8 backing devices but all reside on one HDD.
Data generated by fio like
fio --rw=randrw --randrepeat=1 --size=3000m \ --bsrange=512b-64k --bs_unaligned \ --ioengine=libaio --fsync=1024 \ --name=job0 --name=job1 \
and scrub starts right away this. VM has 4G or memory and 4 CPUs.
The difference is 2 seconds, roughly 4% but the sample is not large enough to be conclusive.
Quoted 53 lines · David Sterba
Quoted 19 lines · David Sterba
Quoted 18 lines · Qu Wenruo
Quoted 17 lines · Filipe Manana
Quoted 16 lines · Qu Wenruo
Quoted 14 lines · Filipe Manana
Quoted 10 lines · Qu Wenruo
This is originally just my preparation for scrub refactors, but when the readahead is involved, it won't just be a small cleanup.
The metadata readahead code is introduced in 2011 (surprisingly, the commit message even contains changelog), but now only one user for it, and even for the only one user, the readahead mechanism can't provide much help in fact.
Scrub needs readahead for commit root, but the existing one can only do current root readahead.
If support for the commit root is added, is there a noticeable speedup? Have you tested that?
Quoted 6 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Will craft a benchmark for that.
Although I don't have any HDD available for benchmark, thus would only have result from SATA SSD.
I'm doing some tests, in a VM on a dedicated HDD.
There's some measurable difference:
With readahead:
Duration: 0:00:20 Total to scrub: 7.02GiB Rate: 236.92MiB/sDuration: 0:00:48 Total to scrub: 12.02GiB Rate: 198.02MiB/sWithout readahead:
Duration: 0:00:22 Total to scrub: 7.02GiB Rate: 215.10MiB/sDuration: 0:00:50 Total to scrub: 12.02GiB Rate: 190.66MiB/sThe setup is: data/single, metadata/dup, no-holes, free-space-tree, there are 8 backing devices but all reside on one HDD.
Quoted 8 lines
Data generated by fio like
fio --rw=randrw --randrepeat=1 --size=3000m \ --bsrange=512b-64k --bs_unaligned \ --ioengine=libaio --fsync=1024 \ --name=job0 --name=job1 \and scrub starts right away this. VM has 4G or memory and 4 CPUs.
How about using bare metal? And was it a debug kernel, or a default kernel config from a distro?
Those details often make all the difference (either for the best or for the worse).
I'm curious to see as well the results when:
1) The reada.c code is changed to work with commit roots;
2) The standard btree readahead (struct btrfs_path::reada) is used instead of the reada.c code.
Quoted 3 lines
The difference is 2 seconds, roughly 4% but the sample is not large enough to be conclusive.
A bit too small.
Thanks.
Quoted 63 lines · Filipe Manana
Quoted 53 lines · David Sterba
Quoted 19 lines · David Sterba
Quoted 18 lines · Qu Wenruo
Quoted 17 lines · Filipe Manana
Quoted 16 lines · Qu Wenruo
Quoted 14 lines · Filipe Manana
Quoted 10 lines · Qu Wenruo
This is originally just my preparation for scrub refactors, but when the readahead is involved, it won't just be a small cleanup.
The metadata readahead code is introduced in 2011 (surprisingly, the commit message even contains changelog), but now only one user for it, and even for the only one user, the readahead mechanism can't provide much help in fact.
Scrub needs readahead for commit root, but the existing one can only do current root readahead.
If support for the commit root is added, is there a noticeable speedup? Have you tested that?
Quoted 6 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Will craft a benchmark for that.
Although I don't have any HDD available for benchmark, thus would only have result from SATA SSD.
I'm doing some tests, in a VM on a dedicated HDD.
There's some measurable difference:
With readahead:
Duration: 0:00:20 Total to scrub: 7.02GiB Rate: 236.92MiB/sDuration: 0:00:48 Total to scrub: 12.02GiB Rate: 198.02MiB/sWithout readahead:
Duration: 0:00:22 Total to scrub: 7.02GiB Rate: 215.10MiB/sDuration: 0:00:50 Total to scrub: 12.02GiB Rate: 190.66MiB/sThe setup is: data/single, metadata/dup, no-holes, free-space-tree, there are 8 backing devices but all reside on one HDD.
Quoted 8 lines
Data generated by fio like
fio --rw=randrw --randrepeat=1 --size=3000m \ --bsrange=512b-64k --bs_unaligned \ --ioengine=libaio --fsync=1024 \ --name=job0 --name=job1 \and scrub starts right away this. VM has 4G or memory and 4 CPUs.
Quoted 8 lines
How about using bare metal? And was it a debug kernel, or a default kernel config from a distro?
Those details often make all the difference (either for the best or for the worse).I'm curious to see as well the results when:
1) The
reada.ccode is changed to work with commit roots;
In fact I have a pretty simple patch for that already.
It's pushed to reada_commit_root branch.
(just misc-next with one patch)
Quoted 3 lines
2) The standard btree readahead (struct
btrfs_path::reada) is used instead of thereada.ccode.
It's already used for csum lookup, see btrfs_lookup_csums_range().
The missing extent tree part is pretty easy to do.
This is pushed to remove_reada branch.
(the patchset with one new one-line patch).
Thanks,
Qu
Quoted 9 lines
Quoted 4 lines
The difference is 2 seconds, roughly 4% but the sample is not large enough to be conclusive.
A bit too small.
Thanks.
Quoted 34 lines · Filipe Manana
Quoted 34 lines
Quoted 1 line
I'm doing some tests, in a VM on a dedicated HDD.
There's some measurable difference:
With readahead:
Duration: 0:00:20 Total to scrub: 7.02GiB Rate: 236.92MiB/sDuration: 0:00:48 Total to scrub: 12.02GiB Rate: 198.02MiB/sWithout readahead:
Duration: 0:00:22 Total to scrub: 7.02GiB Rate: 215.10MiB/sDuration: 0:00:50 Total to scrub: 12.02GiB Rate: 190.66MiB/sThe setup is: data/single, metadata/dup, no-holes, free-space-tree, there are 8 backing devices but all reside on one HDD.
Data generated by fio like
fio --rw=randrw --randrepeat=1 --size=3000m \ --bsrange=512b-64k --bs_unaligned \ --ioengine=libaio --fsync=1024 \ --name=job0 --name=job1 \
Quoted 4 lines
Quoted 1 line
and scrub starts right away this. VM has 4G or memory and 4 CPUs.
How about using bare metal? And was it a debug kernel, or a default kernel config from a distro?
It was the debug config I use for normal testing, I'll try to redo it on another physical box.
Quoted 7 lines
Those details often make all the difference (either for the best or for the worse).
I'm curious to see as well the results when:
1) The
reada.ccode is changed to work with commit roots;
Quoted 8 lines
2) The standard btree readahead (struct
btrfs_path::reada) is used instead of thereada.ccode.Quoted 3 lines
The difference is 2 seconds, roughly 4% but the sample is not large enough to be conclusive.
A bit too small.
What's worse, I did a few more rounds and the results were too unstable, from 44 seconds to 25 seconds (all on the removed readahead branch), but the machine was not quiescent.
Quoted 54 lines · David Sterba
Quoted 34 lines · Filipe Manana
Quoted 34 lines
Quoted 1 line
I'm doing some tests, in a VM on a dedicated HDD.
There's some measurable difference:
With readahead:
Duration: 0:00:20 Total to scrub: 7.02GiB Rate: 236.92MiB/sDuration: 0:00:48 Total to scrub: 12.02GiB Rate: 198.02MiB/sWithout readahead:
Duration: 0:00:22 Total to scrub: 7.02GiB Rate: 215.10MiB/sDuration: 0:00:50 Total to scrub: 12.02GiB Rate: 190.66MiB/sThe setup is: data/single, metadata/dup, no-holes, free-space-tree, there are 8 backing devices but all reside on one HDD.
Data generated by fio like
fio --rw=randrw --randrepeat=1 --size=3000m \ --bsrange=512b-64k --bs_unaligned \ --ioengine=libaio --fsync=1024 \ --name=job0 --name=job1 \Quoted 4 lines
Quoted 1 line
and scrub starts right away this. VM has 4G or memory and 4 CPUs.
How about using bare metal? And was it a debug kernel, or a default kernel config from a distro?
It was the debug config I use for normal testing, I'll try to redo it on another physical box.
Quoted 11 lines
Those details often make all the difference (either for the best or for the worse).
I'm curious to see as well the results when:
1) The
reada.ccode is changed to work with commit roots;2) The standard btree readahead (struct
btrfs_path::reada) is used instead of thereada.ccode.Quoted 1 line
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
The difference is 2 seconds, roughly 4% but the sample is not large enough to be conclusive.
A bit too small.
What's worse, I did a few more rounds and the results were too unstable, from 44 seconds to 25 seconds (all on the removed readahead branch), but the machine was not quiescent.
I get such huge variations too when using a debug kernel and virtualized disks for any tests, even for single threaded tests.
That's why I use a default, non-debug, kernel config from a popular distro and without any virtualization (or at least have qemu use a raw device, not a file backed disk on top of another filesystem) when measuring performance.
Quoted 64 lines · Filipe Manana
Quoted 54 lines · David Sterba
Quoted 34 lines · Filipe Manana
Quoted 34 lines
Quoted 1 line
I'm doing some tests, in a VM on a dedicated HDD.
There's some measurable difference:
With readahead:
Duration: 0:00:20 Total to scrub: 7.02GiB Rate: 236.92MiB/sDuration: 0:00:48 Total to scrub: 12.02GiB Rate: 198.02MiB/sWithout readahead:
Duration: 0:00:22 Total to scrub: 7.02GiB Rate: 215.10MiB/sDuration: 0:00:50 Total to scrub: 12.02GiB Rate: 190.66MiB/sThe setup is: data/single, metadata/dup, no-holes, free-space-tree, there are 8 backing devices but all reside on one HDD.
Data generated by fio like
fio --rw=randrw --randrepeat=1 --size=3000m \ --bsrange=512b-64k --bs_unaligned \ --ioengine=libaio --fsync=1024 \ --name=job0 --name=job1 \Quoted 4 lines
Quoted 1 line
and scrub starts right away this. VM has 4G or memory and 4 CPUs.
How about using bare metal? And was it a debug kernel, or a default kernel config from a distro?
It was the debug config I use for normal testing, I'll try to redo it on another physical box.
Quoted 11 lines
Those details often make all the difference (either for the best or for the worse).
I'm curious to see as well the results when:
1) The
reada.ccode is changed to work with commit roots;2) The standard btree readahead (struct
btrfs_path::reada) is used instead of thereada.ccode.Quoted 1 line
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
The difference is 2 seconds, roughly 4% but the sample is not large enough to be conclusive.
A bit too small.
What's worse, I did a few more rounds and the results were too unstable, from 44 seconds to 25 seconds (all on the removed readahead branch), but the machine was not quiescent.
Quoted 7 lines
I get such huge variations too when using a debug kernel and virtualized disks for any tests, even for single threaded tests.
That's why I use a default, non-debug, kernel config from a popular distro and without any virtualization (or at least have qemu use a raw device, not a file backed disk on top of another filesystem) when measuring performance.
I got my 2.5' HDD installed and tested.
[CONCLUSION]
There is a small but very consistent performance drop for HDD.
Without patchset: average rate = 106.46 MiB/s
With patchset: average rate = 100.74 MiB/s
Diff = -5.67%
[TEST ENV]
HDD: 2T 2.5 inch HDD, 5400rpm device-managed SMR
(WDC WD20SPZX-22UA7T0) HOST: CPU: AMD RYZEN 5900X MEM: 32G DDR4 3200, no ECC No obvious CPU/IO during the test duration
VM: CPU: 16 vcore
MEM: 4G CACHE: none (as even writeback will cause read to be cached)
Although I'm still using VM, the whole disk is passed to VM directly, and has cache=none option.
The initial fs is using 1 device RAID0, as this will cause more stripe based scrub, thus more small metadata readahead triggered.
The initial content for the fs is created by the following fio job first:
[scrub-populate]
directory=/mnt/btrfs
nrfiles=16384
openfiles=16
filesize=2k-512k
readwrite=randwrite
ioengine=libaio
fallocate=none
numjobs=4
Then removed 1/16th (4096) files randomly to create enough gaps in extent tree.
Then run scrub on the fs using both original code, and the patchset with reada enabled for both extent tree (one new one-line patch) and csum tree (already enabled in btrfs_lookup_csums_range()).
Both cases get 8 scrubs run each, between each run, all caches are dropped, and fs get unmounted and re-mounted.
(Yes, this is the perfect situation for the original code, as the fs is not changed, thus current node is the same as commit root)
Each scrub runs shows every small variants, all the duration difference is within 1 second.
The result shows results benefit the original code, while with btrfs_reada_add() removed, the difference is not that large:
[POSSIBLE REASON]
- Synchronous readahead
Maybe this makes readahead less interruptive for data read? As with btrfs_reada_add() removed, path reada is alwasy asynchronous.
- Dedicated readahead thread io priority
Unlike path reada, the readahead thread has dedicated io priority.
I can definitely rework the framework to make it more modern but still keeps above two features.
Or is the 5% performance drop acceptable?
Raw scrub test result attached.
Thanks,
Qu
scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97
Scrub started: Thu Dec 9 20:22:40 2021 Status: finished Duration: 0:02:26 Total to scrub: 18.02GiB Rate: 106.00MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:25:06 2021 Status: finished Duration: 0:02:25 Total to scrub: 18.02GiB Rate: 106.73MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:27:32 2021 Status: finished Duration: 0:02:25 Total to scrub: 18.02GiB Rate: 106.73MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:29:57 2021 Status: finished Duration: 0:02:26 Total to scrub: 18.02GiB Rate: 106.00MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:32:23 2021 Status: finished Duration: 0:02:25 Total to scrub: 18.02GiB Rate: 106.73MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:34:49 2021 Status: finished Duration: 0:02:25 Total to scrub: 18.02GiB Rate: 106.73MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:37:14 2021 Status: finished Duration: 0:02:26 Total to scrub: 18.02GiB Rate: 106.00MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:39:40 2021 Status: finished Duration: 0:02:25 Total to scrub: 18.02GiB Rate: 106.73MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:43:04 2021 Status: finished Duration: 0:02:34 Total to scrub: 18.02GiB Rate: 100.50MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:45:39 2021 Status: finished Duration: 0:02:33 Total to scrub: 18.02GiB Rate: 101.15MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:48:13 2021 Status: finished Duration: 0:02:34 Total to scrub: 18.02GiB Rate: 100.50MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:50:48 2021 Status: finished Duration: 0:02:33 Total to scrub: 18.02GiB Rate: 101.15MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:53:22 2021 Status: finished Duration: 0:02:33 Total to scrub: 18.02GiB Rate: 101.15MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:55:56 2021 Status: finished Duration: 0:02:34 Total to scrub: 18.02GiB Rate: 100.50MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 20:58:30 2021 Status: finished Duration: 0:02:34 Total to scrub: 18.02GiB Rate: 100.50MiB/s Error summary: no errors found scrub done for 16ecd3f9-5466-4f99-854b-2a50a4369a97 Scrub started: Thu Dec 9 21:01:04 2021 Status: finished Duration: 0:02:34 Total to scrub: 18.02GiB Rate: 100.50MiB/s Error summary: no errors found
Quoted 76 lines · Qu Wenruo
Quoted 64 lines · Filipe Manana
Quoted 54 lines · David Sterba
Quoted 34 lines · Filipe Manana
Quoted 34 lines
Quoted 1 line
I'm doing some tests, in a VM on a dedicated HDD.
There's some measurable difference:
With readahead:
Duration: 0:00:20 Total to scrub: 7.02GiB Rate: 236.92MiB/sDuration: 0:00:48 Total to scrub: 12.02GiB Rate: 198.02MiB/sWithout readahead:
Duration: 0:00:22 Total to scrub: 7.02GiB Rate: 215.10MiB/sDuration: 0:00:50 Total to scrub: 12.02GiB Rate: 190.66MiB/sThe setup is: data/single, metadata/dup, no-holes, free-space-tree, there are 8 backing devices but all reside on one HDD.
Data generated by fio like
fio --rw=randrw --randrepeat=1 --size=3000m \ --bsrange=512b-64k --bs_unaligned \ --ioengine=libaio --fsync=1024 \ --name=job0 --name=job1 \Quoted 4 lines
Quoted 1 line
and scrub starts right away this. VM has 4G or memory and 4 CPUs.
How about using bare metal? And was it a debug kernel, or a default kernel config from a distro?
It was the debug config I use for normal testing, I'll try to redo it on another physical box.
Quoted 11 lines
Those details often make all the difference (either for the best or for the worse).
I'm curious to see as well the results when:
1) The
reada.ccode is changed to work with commit roots;2) The standard btree readahead (struct
btrfs_path::reada) is used instead of thereada.ccode.Quoted 1 line
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
The difference is 2 seconds, roughly 4% but the sample is not large enough to be conclusive.
A bit too small.
What's worse, I did a few more rounds and the results were too unstable, from 44 seconds to 25 seconds (all on the removed readahead branch), but the machine was not quiescent.
Quoted 7 lines
I get such huge variations too when using a debug kernel and virtualized disks for any tests, even for single threaded tests.
That's why I use a default, non-debug, kernel config from a popular distro and without any virtualization (or at least have qemu use a raw device, not a file backed disk on top of another filesystem) when measuring performance.
I got my 2.5' HDD installed and tested.
Quoted 9 lines
[CONCLUSION]
There is a small but very consistent performance drop for HDD.
Without patchset: average rate = 106.46 MiB/s
With patchset: average rate = 100.74 MiB/sDiff = -5.67%
5% isn't that big of a deal to get rid of an infrastructure we don't really use and that is in the way.
Additionally I'm going to be reworking scrub quite a bit to work with the new extent tree v2 stuff, the readahead stuff may not fit into the new scheme very well.
Scrub isn't meant to be performant, it's meant to be correct. In fact, many admins would prefer to throttle it and keep it from overwhelming the main workload, so a 5% drop isn't likely to upset anybody.
In my mind we're getting rid of a lot of code for very little cost, which I'm a fan of. I think this is a reasonable trade off, we have more important projects to tackle. In the future if we want to improve things with a readahead framework we can tackle the problem as a whole and make it useful for all of btrfs, not just scrub. Thanks,
Josef