RTPS discovery slow

[jwillemsen]

When running the `DevGuidExamples/DCPS/Messenger' with RTPS on Windows it takes mostly 4 to 5 seconds to run this example, on Linux it finished almost directly. Looks RTPS discovery is slow on Windows, looking at the logs it takes around 3 seconds to let the publisher/subscriber find each other.

[jwillemsen]

Attached log files.

publisher.log
subscriber.log

[simpsont-oci]

What's the value of ResendPeriod in the rtps_discovery section? Try setting it to something small?

[jwillemsen]

I am not behind my PC right now but this is the DevGuide messenger example with all defaults, only tcp as data transport

[simpsont-oci]

Yeah, I think the default is 30 seconds or something large like that, and the default "response" time is about 3 seconds, which sounds like it's what you're seeing. If you set it somewhere in the 1-10 seconds range I'm guessing it would feel more reasonable.

[jwillemsen]

I will test a little bit with these settings, but when I have a subscriber running and I start the publisher, I would expect that at the moment the publisher is created the RTPS discovery will run, waiting on a resend period doesn't feel optimal, shouldn't the discovery process be triggered at the moment the publisher is created?

[jwillemsen]

Debugged this a little bit further, it looks on Windows the order in which things are triggered is slightly different than on Linux. I start the subscriber on the console and the publisher in the debugger. At the moment now the datawriter is created in the publisher there is no associated participant at https://github.com/objectcomputing/OpenDDS/blob/045cde5cb9e4a01d8291917ba4a1527feb071e88/dds/DCPS/RTPS/Sedp.cpp#L5126 so the Sedp is not done at that point.

Now, a little bit later the remote participant is discovered and at https://github.com/objectcomputing/OpenDDS/blob/045cde5cb9e4a01d8291917ba4a1527feb071e88/dds/DCPS/RTPS/Spdp.cpp#L701 there is a comment that our announcement is send now so that they have to wait, but this is only a shorten, with the default settings of a ResendPeriod=30 and QuickResendRatio=0.1 this still means that there is a 3 second delay on the first announcement.

On a new participant, shouldn't we directly send out our announcement, the user can tune the settings to a shorter period, but that impacts also the resends, can image that a very short resend is also bad in some scenarios.

[jwillemsen]

Created #2798 as first step, this improves the startup delay already a lot, looks it still can be more faster, but let us first discuss this solution before spending even more time on this

[jwillemsen]

I also see the same slow discovery on Linux, in our fedora opendds build the DevGuideExamples/DCPS/Messenger/run_test.pl takes 0 secons with InfoRepo discovery and 7 seconds with RTPS

[Saurabh996]

It's still in similar state, it seems?

[jwillemsen]

Yes, no changes made yet, this needs more time to analyze and test solutions, feel free to open a PR with a more intelligent solution or consider sponsoring someone to work more on this.

[Saurabh996]

Okay..