Hello,
v2 -> v3
1. Switched to suggested approach from Jan to make the approach general
for all file writes rather than only for DAX.
(So as of now both DAX & DIO should benefit from this as both uses the same
iomap path. Although note that I only tested performance improvement for DAX)
Gave a run on xfstests with -g quick,dax and didn't observe any new
issues with this patch.
In case of file writes, currently we start a journal txn irrespective of whether
it's an overwrite or not. In case of an overwrite we don't need to start a
jbd2 txn since the blocks are already allocated.
So this patch optimizes away the txn start in case of file (DAX/DIO) overwrites.
This could significantly boost performance for multi-threaded writes
specially random writes (overwrite).
Fio script used to collect perf numbers is mentioned below.
Below numbers were calculated on a QEMU setup on ppc64 box with simulated
pmem (fsdax) device.
Didn't observe any new failures with this patch in xfstests "-g quick,dax"
Performance numbers with different threads - (~10x improvement)
==========================================
vanilla_kernel(kIOPS) (randomwrite)
60 +-+------+-------+--------+--------+--------+-------+------+-+
| + + + +** + + |
55 +-+ ** +-+
| ** ** |
| ** ** |
50 +-+ ** ** +-+
| ** ** |
45 +-+ ** ** +-+
| ** ** |
| ** ** |
40 +-+ ** ** +-+
| ** ** |
35 +-+ ** ** ** +-+
| ** ** ** ** |
| ** ** ** ** ** |
30 +-+ ** ** ** ** ** ** +-+
| ** +** +** +** ** +** |
25 +-+------**------+**------+**------+**------**------+**----+-+
1 2 4 8 12 16
Threads
patched_kernel(kIOPS) (randomwrite)
600 +-+-----+--------+--------+-------+--------+-------+------+-+
| + + + + + +** |
| ** |
500 +-+ ** +-+
| ** |
| ** ** |
400 +-+ ** ** +-+
| ** ** |
300 +-+ ** ** ** +-+
| ** ** ** |
| ** ** ** |
200 +-+ ** ** ** +-+
| ** ** ** ** |
| ** ** ** ** |
100 +-+ ** ** ** ** ** +-+
| ** ** ** ** ** |
| +** +** ** +** +** +** |
0 +-+-----+**------+**------**------+**------+**-----+**----+-+
1 2 4 8 12 16
Threads
fio script
==========
[global]
rw=randwrite
norandommap=1
invalidate=0
bs=4k
numjobs=16 --> changed this for different thread options
time_based=1
ramp_time=30
runtime=60
group_reporting=1
ioengine=psync
direct=1
size=16G
filename=file1.0.0:file1.0.1:file1.0.2:file1.0.3:file1.0.4:file1.0.5:file1.0.6:file1.0.7:file1.0.8:file1.0.9:file1.0.10:file1.0.11:file1.0.12:file1.0.13:file1.0.14:file1.0.15:file1.0.16:file1.0.17:file1.0.18:file1.0.19:file1.0.20:file1.0.21:file1.0.22:file1.0.23:file1.0.24:file1.0.25:file1.0.26:file1.0.27:file1.0.28:file1.0.29:file1.0.30:file1.0.31
file_service_type=random
nrfiles=32
directory=/mnt/
[name]
directory=/mnt/
direct=1
NOTE:
======
1. Looking at ~10x perf delta, I probed a bit deeper to understand what's causing
this scalability problem. It seems when we are starting a jbd2 txn then slab
alloc code is observing some serious contention around spinlock.
I think that the spinlock contention problem in slab alloc path could be optimized
on PPC in general, will look into it seperately. But I could still see the
perf improvement of close to ~2x on QEMU setup on x86 with simulated pmem device
with the patched_kernel v/s vanilla_kernel with same fio workload.
perf report from vanilla_kernel (this is not seen with patched kernel) (ppc64)
=======================================================================
47.86% fio [kernel.vmlinux] [k] do_raw_spin_lock
|
---do_raw_spin_lock
|
|--19.43%--_raw_spin_lock
| |
| --19.31%--0
| |
| |--9.77%--deactivate_slab.isra.61
| | ___slab_alloc
| | __slab_alloc
| | kmem_cache_alloc
| | jbd2__journal_start
| | __ext4_journal_start_sb
<...>
2. This problem was reported by Dan Williams at [1]
Links
======
[1]: https://lore.kernel.org/linux-ext4/20190802144304.GP25064@xxxxxxxxxxxxxx/T/
[v2]: https://lkml.org/lkml/2020/8/22/123
Ritesh Harjani (1):
ext4: Optimize file overwrites
fs/ext4/inode.c | 18 +++++++++++++++---
1 file changed, 15 insertions(+), 3 deletions(-)
--
2.26.2
