[LU-12158] Don't have large stripe_width and stride to prevent unexpected performance regression

Related — Wed, 6 May 2020 19:29:41 +0000

stripe width and stride was calculated based on optimal_io_size that kernel got from backend storage devices. If storage controller returns large optimal io size (e.g. 16MB) to kernel, stripe_width and stride are are large value (e.g. optimal_io_size/blcok size) at mkfs.

[root@es18k-vm1 ~]# dumpe2fs -h /dev/sda  | grep RAID
dumpe2fs 1.44.3.wc1 (23-July-2018)
RAID stride:              4096
RAID stripe width:        4096
[root@es18k-vm1 ~]# dumpe2fs -h /dev/sdi  | grep RAID
dumpe2fs 1.44.3.wc1 (23-July-2018)
RAID stride:              512
RAID stripe width:        512

[root@es18k-vm1 ~]# cat /sys/block/sda/queue/optimal_io_size 
16777216
[root@es18k-vm1 ~]# cat /sys/block/sdi/queue/optimal_io_size 
2097152

However, such large stripe_width and stride size causes performance regression. because ext4 takes more costs to find stripe_width-sized free chunks.

	2MB Chunk size		256K Chunk size
	Write(MB/s)	Read(MB/s)	Write(MB/s)	Read(MB/s)
stripe_width=512,stride=512	10,810	10,124	10,492	6,923
stripe_width=1024,stride=1024	10,793	10,064	10,431	6,921
stripe_width=2048,stride=2048	8,047	10,080	6,629	7,381
stripe_width=4096,stride=4096	7,350	10,089	6,505	7,282

Also, stripe_width and stride are tunable (e.g. -E stripe_width= 4096,stride=4096) option to mkfs. If administrator formats OSTs with wrong stripe_width and stride, it may cause unexpected performance regressions.

mkfs.lustre should avoid such large stripe_width/stride and prints a warning message if it's trying to configure.

Whamcloud Community JIRA

[LU-12158] Don't have large stripe_width and stride to prevent unexpected performance regression