[ovs-dev] [PATCH v3 00/17] Daemon mode for ovn-nbctl

Jakub Sitnicki jkbs at redhat.com
Tue Jul 17 12:41:22 UTC 2018


Hi Mark,

On Mon, 16 Jul 2018 14:11:19 -0400
Mark Michelson <mmichels at redhat.com> wrote:

> I've had a look through again and this series addresses the findings I 
> previously had, so in short:
> 
> Acked-by: Mark Michelson <mmichels at redhat.com>
> 
> I decided to do some testing to see what sort of performance gain we see 
> with this patch. I'm attaching four scripts that I used for testing.
> 
> In all of the tests, I create a logical router and 159 logical switches 
> connected to the router. I then create 92 logical switch ports on each 
> logical switch (a total of 14628 ports). On alternating logical switch 
> port additions, I create an address set. Every set of two logical switch 
> ports gets added to the most recently created address set. For each 
> logical switch port, I also create two ACLs. Note that the script does 
> not perform any synchronization or port binding. The goal of the script 
> is to isolate the performance of ovn-nbctl at scale rather than as an 
> overall view of the performance of OVN.
> 
> The scale.sh script is a straightforward version that uses individual 
> calls to ovn-nbctl at each step. The scale_batch.sh script batches 
> multiple operations into a single ovn-nbctl call. This way, the switch 
> port, address set, and ACLs are created in one ovn-nbctl invocation. I 
> ran these scripts about a month ago in a sandbox against OVS master. The 
> scale.sh script took ~13 hours to complete. The scale_batch.sh script 
> took ~3 hours to complete.
> 
> I then modified each of these scripts to work with ovn-nbctl in daemon 
> mode. scale.sh was modified into scale-daemon.sh, and scale_batch.sh was 
> modified into scale_batch-daemon.sh. I applied the patch series to the 
> current OVS master and ran the daemon versions of the scripts in a 
> sandbox. scale-daemon.sh completed in 8 minutes 54 seconds. 
> scale_batch-daemon.sh completed in 3 minutes 52 seconds. In both cases, 
> this is a 99.9% decrease in the amount of time to complete.

Thanks a lot for the review feedback and for doing proper benchmarking
of these changes. Great to have a confirmation that there is the
expected speed-up.

Thanks,
Jakub


More information about the dev mailing list