[4.8] btrfs heats my room with lock contention
Simple test. 8GB pmem device on a 16p machine:
# mkfs.btrfs /dev/pmem1 # mount /dev/pmem1 /mnt/scratch # dbench -t 60 -D /mnt/scratch 16
And heat your room with the warm air rising from your CPUs. Top half of the btrfs profile looks like:
36.71% [kernel] [k] _raw_spin_unlock_irqrestore ¿ 32.29% [kernel] [k] native_queued_spin_lock_slowpath ¿ 5.14% [kernel] [k] queued_write_lock_slowpath ¿ 2.46% [kernel] [k] _raw_spin_unlock_irq ¿ 2.15% [kernel] [k] queued_read_lock_slowpath ¿ 1.54% [kernel] [k] _find_next_bit.part.0 ¿ 1.06% [kernel] [k] __crc32c_le ¿ 0.82% [kernel] [k] btrfs_tree_lock ¿ 0.79% [kernel] [k] steal_from_bitmap.part.29 ¿ 0.70% [kernel] [k] __copy_user_nocache ¿ 0.69% [kernel] [k] btrfs_tree_read_lock ¿ 0.69% [kernel] [k] delay_tsc ¿ 0.64% [kernel] [k] btrfs_set_lock_blocking_rw ¿ 0.63% [kernel] [k] copy_user_generic_string ¿ 0.51% [kernel] [k] do_raw_read_unlock ¿ 0.48% [kernel] [k] do_raw_spin_lock ¿ 0.47% [kernel] [k] do_raw_read_lock ¿ 0.46% [kernel] [k] btrfs_clear_lock_blocking_rw ¿ 0.44% [kernel] [k] do_raw_write_lock ¿ 0.41% [kernel] [k] __do_softirq ¿ 0.28% [kernel] [k] __memcpy ¿ 0.24% [kernel] [k] map_private_extent_buffer ¿ 0.23% [kernel] [k] find_next_zero_bit ¿ 0.22% [kernel] [k] btrfs_tree_read_unlock ¿
Performance vs CPu usage is:
nprocs throughput cpu usage
1 440MB/s 50%
2 770MB/s 100%
4 880MB/s 250%
8 690MB/s 450%
16 280MB/s 950%
In comparision, at 8-16 threads ext4 is running at ~2600MB/s and XFS is running at ~3800MB/s. Even if I throw 300-400 processes at ext4 and XFS, they only drop to ~1500-2000MB/s as they hit internal limits.
Cheers,
Dave.
Signature · 2 lines
Dave Chinner [email protected]
Quoted 44 lines · Dave Chinner
Simple test. 8GB pmem device on a 16p machine:
# mkfs.btrfs /dev/pmem1 # mount /dev/pmem1 /mnt/scratch # dbench -t 60 -D /mnt/scratch 16And heat your room with the warm air rising from your CPUs. Top half of the btrfs profile looks like:
36.71% [kernel] [k] _raw_spin_unlock_irqrestore ¿ 32.29% [kernel] [k] native_queued_spin_lock_slowpath ¿ 5.14% [kernel] [k] queued_write_lock_slowpath ¿ 2.46% [kernel] [k] _raw_spin_unlock_irq ¿ 2.15% [kernel] [k] queued_read_lock_slowpath ¿ 1.54% [kernel] [k] _find_next_bit.part.0 ¿ 1.06% [kernel] [k] __crc32c_le ¿ 0.82% [kernel] [k] btrfs_tree_lock ¿ 0.79% [kernel] [k] steal_from_bitmap.part.29 ¿ 0.70% [kernel] [k] __copy_user_nocache ¿ 0.69% [kernel] [k] btrfs_tree_read_lock ¿ 0.69% [kernel] [k] delay_tsc ¿ 0.64% [kernel] [k] btrfs_set_lock_blocking_rw ¿ 0.63% [kernel] [k] copy_user_generic_string ¿ 0.51% [kernel] [k] do_raw_read_unlock ¿ 0.48% [kernel] [k] do_raw_spin_lock ¿ 0.47% [kernel] [k] do_raw_read_lock ¿ 0.46% [kernel] [k] btrfs_clear_lock_blocking_rw ¿ 0.44% [kernel] [k] do_raw_write_lock ¿ 0.41% [kernel] [k] __do_softirq ¿ 0.28% [kernel] [k] __memcpy ¿ 0.24% [kernel] [k] map_private_extent_buffer ¿ 0.23% [kernel] [k] find_next_zero_bit ¿ 0.22% [kernel] [k] btrfs_tree_read_unlock ¿Performance vs CPu usage is:
nprocs throughput cpu usage
1 440MB/s 50%
2 770MB/s 100%
4 880MB/s 250%
8 690MB/s 450%
16 280MB/s 950%
Quoted 5 lines
In comparision, at 8-16 threads ext4 is running at ~2600MB/s and XFS is running at ~3800MB/s. Even if I throw 300-400 processes at ext4 and XFS, they only drop to ~1500-2000MB/s as they hit internal limits.
Yes, with dbench btrfs does much much better if you make a subvol per dbench dir. The difference is pretty dramatic. I'm working on it this month, but focusing more on database workloads right now.
-chris
Quoted 12 lines · Chris Mason
Quoted 9 lines · Dave Chinner
Simple test. 8GB pmem device on a 16p machine:
# mkfs.btrfs /dev/pmem1 # mount /dev/pmem1 /mnt/scratch # dbench -t 60 -D /mnt/scratch 16And heat your room with the warm air rising from your CPUs. Top half of the btrfs profile looks like:
.....
Quoted 9 lines
Quoted 9 lines
Performance vs CPu usage is:
nprocs throughput cpu usage
1 440MB/s 50%
2 770MB/s 100%
4 880MB/s 250%
8 690MB/s 450%
16 280MB/s 950%
Quoted 8 lines
Quoted 5 lines
In comparision, at 8-16 threads ext4 is running at ~2600MB/s and XFS is running at ~3800MB/s. Even if I throw 300-400 processes at ext4 and XFS, they only drop to ~1500-2000MB/s as they hit internal limits.
Yes, with dbench btrfs does much much better if you make a subvol per dbench dir. The difference is pretty dramatic. I'm working on it this month, but focusing more on database workloads right now.
You've been giving this answer to lock contention reports for the past 6-7 years, Chris. I really don't care about getting big benchmark numbers with contrived setups - the "use multiple subvolumes" solution is simply not practical for users or their workloads. The default config should behave sanely and not not contribute to global warming like this.
Cheers,
Dave.
Signature · 2 lines
Dave Chinner [email protected]
Quoted 32 lines · Dave Chinner
Quoted 12 lines · Chris Mason
Quoted 9 lines · Dave Chinner
Simple test. 8GB pmem device on a 16p machine:
# mkfs.btrfs /dev/pmem1 # mount /dev/pmem1 /mnt/scratch # dbench -t 60 -D /mnt/scratch 16And heat your room with the warm air rising from your CPUs. Top half of the btrfs profile looks like:
.....
Quoted 9 lines
Quoted 9 lines
Performance vs CPu usage is:
nprocs throughput cpu usage
1 440MB/s 50%
2 770MB/s 100%
4 880MB/s 250%
8 690MB/s 450%
16 280MB/s 950%Quoted 8 lines
Quoted 5 lines
In comparision, at 8-16 threads ext4 is running at ~2600MB/s and XFS is running at ~3800MB/s. Even if I throw 300-400 processes at ext4 and XFS, they only drop to ~1500-2000MB/s as they hit internal limits.
Yes, with dbench btrfs does much much better if you make a subvol per dbench dir. The difference is pretty dramatic. I'm working on it this month, but focusing more on database workloads right now.
Quoted 7 lines
You've been giving this answer to lock contention reports for the past 6-7 years, Chris. I really don't care about getting big benchmark numbers with contrived setups - the "use multiple subvolumes" solution is simply not practical for users or their workloads. The default config should behave sanely and not not contribute to global warming like this.
The btree setup that makes lock contention here makes some other benchmarks faster. Needing to create subvolumes in order to fix performance problems on dbench is far from ideal, but in production here the tradeoffs have been worth it.
Basically this one definitely comes up during dbench and fs_mark and much less often elsewhere. For the workloads that hit this lock contention, splitting things out into subvolumes hugely reduces metadata fragmentation on reads. So it's not just CPU we're helping with subvolumes but spindle time too.
It's true I haven't invested time into guessing when the admin wants to split on a per-subvolume basis. Still, I do love the polar bears, so I'll take another shot at the btree lock.
-chris