Performance of Pavel's multicredit i/o SMB3 patches continues to look
good. Additional informal performance results below comparing cifs
mounts with smb3 mounts (vers=3.0) with and without Pavel's patch set.
I plan to do additional testing with large rsize/wsize (default with
Pavel's code is 1MB).
3.16-rc4 (Ubuntu) on client. Server is Windows 8.1. Both VMs on same
host (host disk is fairly fast SSD).
Copy to server performance increased about 20% percent
dd if=/dev/zero of=/mnt/targetfile bs=80M count=25
got similar results with or without conv=fdatasync
1st run copying to empty directory, 2nd run copying over targetfile,
(pattern repeated multiple times) averaging results
New code (with Pavel's patches)
---------------------------------------------
CIFS 167MB/s
SMB3 200MB/s
Existing code (without his patches)
------------------------------------------------
SMB3 166MB/s
CIFS 164.5MB/s
For large file reading SMB3 performance with Pavel's patches increased
76% over existing SMB3 code
dd of=/dev/null if=/mnt/targetfile bs=80M count=25
(mounting and unmounting between attempts to avoid caching effects on
the client)
New code (with Pavel's patches)
---------------------------------------------
CIFS 114MB/s
SMB3 216MB/s
Existing code (without his patches)
------------------------------------------------
SMB3 123MB/s
CIFS 110MB/s
--
Thanks,
Steve
Repeating these large file copy tests with SMB3 mounts with Pavel's
patches and trying larger wsize and rsize of 4MB (and later 5MB) did
not seem to help performance much (over his default with the new
patchset, 1MB). I want to repeat these with real network adapters
(rather than in vmware or kvm) to compare though (to see if the larger
wsize and rsize help)
On Sun, Jul 13, 2014 at 2:23 PM, Steve French <[email protected]> wrote:
> Performance of Pavel's multicredit i/o SMB3 patches continues to look
> good. Additional informal performance results below comparing cifs
> mounts with smb3 mounts (vers=3.0) with and without Pavel's patch set.
> I plan to do additional testing with large rsize/wsize (default with
> Pavel's code is 1MB).
>
> 3.16-rc4 (Ubuntu) on client. Server is Windows 8.1. Both VMs on same
> host (host disk is fairly fast SSD).
>
> Copy to server performance increased about 20% percent
> dd if=/dev/zero of=/mnt/targetfile bs=80M count=25
> got similar results with or without conv=fdatasync
>
> 1st run copying to empty directory, 2nd run copying over targetfile,
> (pattern repeated multiple times) averaging results
>
> New code (with Pavel's patches)
> ---------------------------------------------
> CIFS 167MB/s
> SMB3 200MB/s
>
> Existing code (without his patches)
> ------------------------------------------------
> SMB3 166MB/s
> CIFS 164.5MB/s
>
> For large file reading SMB3 performance with Pavel's patches increased
> 76% over existing SMB3 code
> dd of=/dev/null if=/mnt/targetfile bs=80M count=25
> (mounting and unmounting between attempts to avoid caching effects on
> the client)
>
> New code (with Pavel's patches)
> ---------------------------------------------
> CIFS 114MB/s
> SMB3 216MB/s
>
> Existing code (without his patches)
> ------------------------------------------------
> SMB3 123MB/s
> CIFS 110MB/s
>
> --
> Thanks,
>
> Steve
--
Thanks,
Steve
Impressive!
On Sun, Jul 13, 2014 at 12:23 PM, Steve French <[email protected]> wrote:
> Performance of Pavel's multicredit i/o SMB3 patches continues to look
> good. Additional informal performance results below comparing cifs
> mounts with smb3 mounts (vers=3.0) with and without Pavel's patch set.
> I plan to do additional testing with large rsize/wsize (default with
> Pavel's code is 1MB).
>
> 3.16-rc4 (Ubuntu) on client. Server is Windows 8.1. Both VMs on same
> host (host disk is fairly fast SSD).
>
> Copy to server performance increased about 20% percent
> dd if=/dev/zero of=/mnt/targetfile bs=80M count=25
> got similar results with or without conv=fdatasync
>
> 1st run copying to empty directory, 2nd run copying over targetfile,
> (pattern repeated multiple times) averaging results
>
> New code (with Pavel's patches)
> ---------------------------------------------
> CIFS 167MB/s
> SMB3 200MB/s
>
> Existing code (without his patches)
> ------------------------------------------------
> SMB3 166MB/s
> CIFS 164.5MB/s
>
> For large file reading SMB3 performance with Pavel's patches increased
> 76% over existing SMB3 code
> dd of=/dev/null if=/mnt/targetfile bs=80M count=25
> (mounting and unmounting between attempts to avoid caching effects on
> the client)
>
> New code (with Pavel's patches)
> ---------------------------------------------
> CIFS 114MB/s
> SMB3 216MB/s
>
> Existing code (without his patches)
> ------------------------------------------------
> SMB3 123MB/s
> CIFS 110MB/s
>
> --
> Thanks,
>
> Steve
Continuing testing of Pavel's newest SMB3 multicredit patch series,
which significantly improves large file read/write speeds to Samba and
Windows from Linux. For this workload LInux to Linux - SMB3 seems
faster than alternatives for read (copying from the server) but about
the same as NFS for write (copy to server). Some additional data:
client Ubuntu with 3.16-rc4 and Pavel's patch series, server Fedora 20
(3.14.9 kernel, and Samba server version 4.1.9)
dd if=/mnt/testfile of=/dev/null bs=50M count=30
testfile is 1.5GB existing file, unmount/mount inbetween each large
file copy to avoid any caching effect on client (although server will
have cached it)
SMB3 averaged 199MB/sec reads (copy from server)
CIFS averaged 170MB/sec reads (copy from server)
NFSv3 averaged 116MB/sec (copy from server)
NFSv4 and v4.1 averaged 110MB/sec (copy from server)
Write speeds (doing dd if=/dev/zero of=/mnt/testfile bs=60M count=25)
were more widely varied but averaged similar speeds for copy to server
for both NFSv3/v4/v4.1 and SMB3 (about 175MB/s)
On Sun, Jul 13, 2014 at 2:23 PM, Steve French <[email protected]> wrote:
> Performance of Pavel's multicredit i/o SMB3 patches continues to look
> good. Additional informal performance results below comparing cifs
> mounts with smb3 mounts (vers=3.0) with and without Pavel's patch set.
> I plan to do additional testing with large rsize/wsize (default with
> Pavel's code is 1MB).
>
> 3.16-rc4 (Ubuntu) on client. Server is Windows 8.1. Both VMs on same
> host (host disk is fairly fast SSD).
>
> Copy to server performance increased about 20% percent
> dd if=/dev/zero of=/mnt/targetfile bs=80M count=25
> got similar results with or without conv=fdatasync
>
> 1st run copying to empty directory, 2nd run copying over targetfile,
> (pattern repeated multiple times) averaging results
>
> New code (with Pavel's patches)
> ---------------------------------------------
> CIFS 167MB/s
> SMB3 200MB/s
>
> Existing code (without his patches)
> ------------------------------------------------
> SMB3 166MB/s
> CIFS 164.5MB/s
>
> For large file reading SMB3 performance with Pavel's patches increased
> 76% over existing SMB3 code
> dd of=/dev/null if=/mnt/targetfile bs=80M count=25
> (mounting and unmounting between attempts to avoid caching effects on
> the client)
>
> New code (with Pavel's patches)
> ---------------------------------------------
> CIFS 114MB/s
> SMB3 216MB/s
>
> Existing code (without his patches)
> ------------------------------------------------
> SMB3 123MB/s
> CIFS 110MB/s
>
> --
> Thanks,
>
> Steve
--
Thanks,
Steve