[RFC] Performance improvements

Martin Schwenke martin at meltin.net
Sat Jun 16 02:14:21 UTC 2018


Hi Swen,

On Fri, 15 Jun 2018 09:38:08 +0200, Swen Schillig via samba-technical
<samba-technical at lists.samba.org> wrote:

> Over the past few months I was creating a good few patches with the
> goal to improve the Sambas performance in a clustered environment.
> 
> The patchset is modifying CTDB-code only but please keep reading...
> 
> The result of the changes is a huge reduction of the variation whether
> it is execution time or maximum number of operations at a time and
> an overall performance improvement of >10%.
> 
> Attached are a few graphs showing the results of tests performed on our
> internal cluster with an LDX appliance.
> The first three graphs are the results of one test stressing the
> directory listing performance and the two SMB2_TP graphs were performed
> in a separate test using a 4 cluster node with 3000 client connections
> per node (so in total 12000).
> 
> The improvements are achieved by optimizing the memory management,
> e.g. using memory pools or being a more precise in the memory
> requirement prediction and using a less iterative approach for
> calculation. In addition I tried to re-use allocated memory where
> possible instead of free'ing and re-allocating the same.
> 
> Before I dump the code changes on the community I'd like to know
> if there is any other test I need to perform and if the described
> results are good enough for you to consider such non-functional
> changes.
> 
> Thanks in advance for your comments and support.

The LDX appliance also generates spreadsheets containing numerical
data. These include average response times for various SMB operations
(read/write/create/close/session-setup) and overall throughput (in MB).
Sometimes the graphs are strange and the numbers tell us something
different.

For example, your "SMB2 Response Time" graph seems to show a nice flat
average response time of something like 4ms while the response time
across the other graph is usually lower than that. That's surprising!
However, I won't be even more surprised if the numerical data tells us
something more positive.

It seems good that the graphs show a huge reduction in variation but
I'd sure like to see the numerical data from the spreadsheets...

Do you have the numerical data handy?

Thanks...

peace & happiness,
martin



More information about the samba-technical mailing list