rsize/wsize and Samba performance

Jeremy Allison jra at samba.org
Tue May 22 17:43:04 UTC 2018


On Mon, May 21, 2018 at 08:28:49PM -0500, Steve French wrote:
> Any idea why Samba performance is worse with 8MB i/os than 1MB i/os for
> sequential
> read/write multiple process i/os?   Performance was good ... but it surprised
> me that 1MB was better than 8MB

I'm guessing multiple 8MB IO's is a little too much for the buffer cache
to handle.

> The fio job file I was using was the normal fio-seq-RW
> # cat ~/fio/fio-seq-RW.job
> [global]
> name=fio-seq-RW
> filename=fio-seq-RW
> rw=rw
> rwmixread=60
> rwmixwrite=40
> bs=256K
> direct=0
> numjobs=4
> time_based=1
> runtime=100
> 
> [file1]
> size=10G
> ioengine=libaio
> iodepth=16
> 
> 
> SMB3.02 mounts to Samba
> 
> rsize/wsize default (ie 1MB on Linux client)
>    READ: bw=3098MiB/s (3249MB/s), 773MiB/s-776MiB/s (810MB/s-814MB/s), io=
> 303GiB (325GB), run=100001-100002msec
>   WRITE: bw=2071MiB/s (2172MB/s), 518MiB/s-518MiB/s (543MB/s-543MB/s), io=
> 202GiB (217GB), run=100001-100002msec
> 
> ---------------------------------------------
> 
> rsize/wsize=8MB mount on Linux client
>    READ: bw=2779MiB/s (2913MB/s), 692MiB/s-696MiB/s (726MB/s-730MB/s), io=
> 271GiB (291GB), run=100001-100001msec
>   WRITE: bw=1858MiB/s (1948MB/s), 464MiB/s-465MiB/s (487MB/s-487MB/s), io=
> 181GiB (195GB), run=100001-100001msec
> 
> NFSv3 and NFSv4.1 as expected (to kernel NFS server) were much slower than SMB3
> for both read and write (presumably Samba is handling more I/Os and doing
> better job async - i/o sizes are now the same 1MB by default)
> --
> Thanks,
> 
> Steve



More information about the samba-technical mailing list