[REPORT] Performance Test Samba vs Windows

Garming Sam garming at catalyst.net.nz
Mon May 14 03:40:12 UTC 2018

On 14/05/18 15:23, William Brown wrote:
>> 2) Looking at median packet response times, we generally spend twice
>> as
>> long on processing a packet. However, the difference between our 95
>> percentile values is often much smaller than this and some of the
>> mean
>> values are actually at parity.
>> 3) With small to moderate loads, Samba handles a similar amount of
>> operations per second to Windows. For the values shown, it actually
>> seems to get more throughput (although not incredibly so) for a good
>> chunk of the table shown.
> What about "large" loads? 

Joe's original email seems to have been eaten by the list for now. There
was a table he had, which I'll see if I can get him to resend. There's
two axis of load that we generate, more operations along the same
connections + using the same client vs having more clients. As you ramp
up more operations per client, we start falling behind to Windows up to
the 50% number I mentioned. On the other hand, increasing the amount of
clients doesn't seem to have the same behaviour. Windows doesn't seem to
handle more clients (+ more open connections) much better than we do and
the gap between us and Windows does not appear to be quite so bad. After
a certain point, adding more simultaneous clients degrades the maximum
performance. It would be interesting to see what a ridiculous amount of
clients would compare between Samba and Windows.



>> It's actually kind of surprising that Samba doesn't perform all that
>> much worse than Windows does. It's certainly not off by an order of
>> magnitude (or two), although this testing has been run in the prefork
>> mode. Our previous analyses showed much smaller numbers and that's
>> likely due to the limitations of being stuck with a single-process.
>> Cheers,
>> Garming
>> On 14/05/18 14:45, joeg at catalyst.net.nz wrote:
>>> Hi Team,
>>> I've been trying to make the Samba performance test tool work
>>> against
>>> Windows.
>>> After a lot of effort, Garming and I fixed all the errors for
>>> different packets, and finally it's working(fixes are merged into
>>> master).
>>> Base on that, we can send traffic to Samba and Windows, get the
>>> maximum load, and compare the performance.
>>> An example of the command I was using:
>>> script/traffic_replay -U Administrator%PASSWORD  --realm
>>> krb.samba.site --workgroup KRB --fixed-password FIXED-PASSWORD -r 4
>>> -S
>>> 4 traffic-sample-1-model.txt dc1.krb.samba.site 2>&1 >
>>> traffic_replay_stats/dc1_r4_S4.txt
>>> By changing the combination of different -r and -S, we can get test
>>> results and save to files, then parse files to get summary tables
>>> as
>>> below.
>>> The number displayed is "*Successful operations per s**econd*", the
>>> 0
>>> ones are combinations we think not important and skipped.
>>>       Test Result for Samba (master code on 2018.05.11):
>>>       Test Result for Windows Server 2012r2:
>>> I am running the test with 4 servers in Catalyst Cloud which is
>>> Openstack:
>>> All data are attached, hope this can help the team to understand
>>> how
>>> is Samba performing with compare to Windows.
>>> -- 
>>> Joe Guo
>>> joeg at catalyst.net.nz
>>> Catalyst IT

More information about the samba-technical mailing list