[Samba] Performance of samba in linux vs windows

eric roseme eroseme at emonster.rose.hp.com
Fri Oct 1 23:01:40 GMT 2004


Hi Tim,

Just as a sanity check.....

I did some testing earlier this year to characterize performance 
differences btw 2.2.8a and 3.0.2a.  I tested simple copies of one .5 GB 
file, and also a directory with 5000 files (with very long names 
including upper and lower case).  As long as I was testing the version 
deltas, I also compared the tests to a Windows 2003 Server.

I do not want to specify the exact results and hardware (since I work 
for a vendor), but for the single-big-file test Windows 2003 ftp was 
slower than Samba by a factor of 3.  ftp on HP-UX was just slightly 
faster than Samba.  For the 5000-files test, reading from the server was 
about the same for all SMB server platforms (XP from W2003, 2.2.8a, 
3.0.2a).  For the 5000-files test, writing to the server was 
significantly slower on Samba versus Windows.  This is well-known 
behavior for large directories due to name mangling and case sensitivity.

I also tested extensively versus NFS (but this was on 2.0.6 - quite a 
while ago) and the total throughput numbers (MB/s) were almost the same 
for SMB vs NFS.  These were 8-way 4-GbE boxes, though.

I cannot claim these results as "benchmarks" - maybe someday if we get a 
CIFS benchmark like SPEC then we'll have a level playing field.  The 
point is, that results vary all over the place by environment.  (also - 
turn off strict locking and test again).

Go Mustangs! (c/o '80 & '88)

Eric Roseme
Hewlett-Packard

Tim Harvey wrote:

> I'm doing some performance tests on a samba NAS server and I've found some
> interesting statistics:
> 
> I'm doing my performance tests in linux using:
>   # time dd if=somelargefileovershare of=/dev/null bs=1M count=100
> Then calculating the bandwidth
> 
> For windows I'm low-tech: stopwatch plus drag-n-drop of a large file (any
> recommendations on a 'simple' windows program that will tell you how long it
> took to copy a file, or even calc the BW for you?)
> 
> Here are my bandwidth results:
> 
> nfs via linux: 10MB/s
> smb via linux: 5MB/s
> smb via win: 8MB/s
> 
> Questions:
>   - why would I be getting half the performance via nfs vs smb?  Is there a
> lot more overhead with smb vs nfs?
>   - why the large difference between using smb from a linux box vs smb from
> windows?  The windows transfers are much faster... almost 2X
> 
> I'm just trying to understand my results better.  The samba server I'm
> mounting to is running on a 1.2GHz Celeron, 256MB SDRAM, using a raid5 array
> with an XFS filesystem on ATA drives with a 100mbps nic.  The bottleneck
> here is the 100mbps nic, which theoretically will give me a max throughput
> from the server of 12.5MB/sec, so I'm fairly satisfied to see 10MB/sec from
> the nfs test.
> 
> Thanks for any assistance in understanding these results,
> 
> Tim
> 


More information about the samba mailing list