Sending metadata across UDP

Nicolas Dufresne nicolas at ndufresne.ca
Fri Oct 4 15:20:59 UTC 2019


Le ven. 4 oct. 2019 06 h 55, priyanka kataria <priyanka.kataria86 at gmail.com>
a écrit :

> Hi Nicolas,
>
> >Looks like packet lost to me. If you need all frames, use a reliable
> protocol.
>
> I am creating program to handle live sources so I will have to use UDP
> protocol only.
>

Was a deviated way to ask if your code tolerates lost, which will happen
whatever you do about it over UDP/RTP.


> Plus I think this behavior occurs because attaching the probe on each
> frame and modifying it, slows down the pipeline a lot.
> Is this correct?
>

Can't say, by themselves, C written pad probes don't slow down much the
flow, but memory operations could. We've done what you are doing on RPI2
over wifi, and it worked, so it makes your case a bit suspicious.


> Thanks,
> Priyanka
>
> On Wed, Oct 2, 2019 at 12:03 AM Nicolas Dufresne <nicolas at ndufresne.ca>
> wrote:
>
>>
>>
>> Le mar. 1 oct. 2019 08 h 10, priyanka kataria <
>> priyanka.kataria86 at gmail.com> a écrit :
>>
>>> Hi Nicolas,
>>>
>>> Thank you again.
>>>
>>> > I'm suspicious you don't check the probe type correctly, since if you add
>>> both LIST and BUFFER flag on your probe,
>>> > it means the probe data can be both. If LIST was effectively create,
>>> you'd likely get 1 list per frame I think.
>>>
>>> I don't think I understand this completely.
>>> Are you pointing towards the flag "GST_PAD_PROBE_TYPE_BUFFER_LIST" used
>>> in send_jpeg.c and "pay_src_probe" function handling both
>>> "GST_PAD_PROBE_TYPE_BUFFER_LIST" and "GST_PAD_PROBE_TYPE_BUFFER"?
>>> If that is the case, I have attached modified source wherein I have
>>> commented code for GST_PAD_PROBE_TYPE_BUFFER, but the behavior is still the
>>> same, around 100 lists per frame.
>>>
>>> > That being said, for "frame base" metadata, you may put that info on all
>>> packet, for redundancy, but then you need a pair of probes on the
>>> > receiver, one remember the last seen value, and the other will appy it somehow
>>> to your reconstructed frames.
>>>
>>> In modified "recv_joeg.c", I have added frame ID as metadata in
>>> "decode_buffer_probe" but the output I get is:
>>> 2222222 Frame id is : 1
>>> 2222222 Frame id is : 2
>>> 2222222 Frame id is : 62
>>> 2222222 Frame id is : 63
>>> 2222222 Frame id is : 64
>>> 2222222 Frame id is : 65
>>> 2222222 Frame id is : 66
>>> 2222222 Frame id is : 67
>>> 2222222 Frame id is : 68
>>>>>>
>>> Many of the frame IDs are not attached on the buffer or they are just
>>> not being printed, I am not sure.
>>> But this is for sure that this method makes the program very slow and
>>> cannot be used for live stream.
>>>
>>> Please correct me if I am wrong.
>>>
>>
>>
>> Looks like packet lost to me. If you need all frames, use a reliable
>> protocol.
>>
>>
>>> Thanks,
>>> Priyanka
>>>
>>> On Mon, Sep 30, 2019 at 9:55 PM Nicolas Dufresne <nicolas at ndufresne.ca>
>>> wrote:
>>>
>>>> Le lundi 30 septembre 2019 à 17:13 +0530, priyanka kataria a écrit :
>>>> > Hi Nicolas,
>>>> >
>>>> > >There is no code for buffer list there, with buffer list you need to
>>>> > use different API. Also, before you modify a buffer, you need to
>>>> ensure
>>>> > it's writable, making it writable will may change the GstBuffer, so
>>>> you
>>>> > have to update that pointer.
>>>> >
>>>> > Thank you for the pointers.
>>>> >
>>>> > I have attached the modified JPEG code which handles BUFFER_LIST.
>>>> > However, for each frame I get around 100 lists, hence, same
>>>> frame_count is printed around 100 times, both in case of send_jpeg.c and
>>>> recv_jpeg.c.
>>>> >
>>>> > And the program is also very slow, which I believe is due to
>>>> attaching probe on each bufferlist.
>>>> >
>>>> > But my pain point is I wanted to propagate a unique frame number
>>>> across UDP to identify each frame and what I have got is around 100 lists
>>>> per frame, which makes this frame number useless for me. How do I make use
>>>> of this frame number which I have got?
>>>> >
>>>> > Is my explanation clear enough to understand the purpose of sending
>>>> frame number across UDP?
>>>>
>>>> I'm suspicious you don't check the probe type correctly, since if you
>>>> add both LIST and BUFFER flag on your probe, it means the probe data
>>>> can be both. If LIST was effectively create, you'd likely get 1 list
>>>> per frame I think.
>>>>
>>>> That being said, for "frame base" metadata, you may put that info on
>>>> all packet, for redundancy, but then you need a pair of probes on the
>>>> receiver, one remember the last seen value, and the other will appy it
>>>> somehow to your reconstructed frames.
>>>>
>>>> >
>>>> > Thanks,
>>>> > Priyanka Kataria
>>>> >
>>>> > On Fri, Sep 27, 2019 at 7:03 PM Nicolas Dufresne <
>>>> nicolas at ndufresne.ca> wrote:
>>>> > > Le vendredi 27 septembre 2019 à 10:49 +0530, priyanka kataria a
>>>> écrit :
>>>> > > > Hi Nicolas,
>>>> > > >
>>>> > > > Thank you for quick reply.
>>>> > > >
>>>> > > > > Consider adding an rtpjitterbuffer on the receiver, in future
>>>> GST version, you'll be able to use rtpsrc/rtpsink (or the rist variant) in
>>>> order to get a full feature RTP stream without complex pipeline
>>>> construction.
>>>> > > >
>>>> > > > The actual pipeline which I use in the program has
>>>> "rtpjitterbuffer", I shared a simple test pipeline.
>>>> > > >
>>>> > > > >For jpeg, consider configuring max-bitrate property on udpsink.
>>>> As frames are spread out on a lot more packet, it tend to become bursty and
>>>> may saturate the link or exhaust the udpsrc socket buffer-size.
>>>> > > >
>>>> > > > Point noted, will make the changes.
>>>> > > >
>>>> > > > >That is the method I would have used. It should work with any
>>>> RTP packet, so you likely have or hit a bug.
>>>> > > > Here, I am attaching the source code for option 3 I tried.
>>>> > > >
>>>> > > > "send_h264.c" and "recv_h264.c", work successfully and the frame
>>>> number is appended to RTP buffer. Prints in both the probes output the
>>>> correct value.
>>>> > > >
>>>> > > > However, the probe function (pay_src_probe) in "send_jpeg.c"
>>>> never gets called.
>>>> > > > When I change the probe type from "GST_PAD_PROBE_TYPE_BUFFER" to
>>>> "GST_PAD_PROBE_TYPE_BUFFER_LIST", then it gets called, and the frame
>>>> numbers appended are wrong. i.e. for each frame around 100 times this
>>>> function gets called and program slows down like a sloth.
>>>> > >
>>>> > > There is no code for buffer list there, with buffer list you need to
>>>> > > use different API. Also, before you modify a buffer, you need to
>>>> ensure
>>>> > > it's writable, making it writable will may change the GstBuffer, so
>>>> you
>>>> > > have to update that pointer.
>>>> > >
>>>> > > >
>>>> > > > I checked the source code for "rtph264pay" and rtpjpegpay"
>>>> element, both of them are creating buffer lists which I guess is to push
>>>> multiple RTP packets at one go, to the next element in pipeline.
>>>> > > > But strangely H264 works fine and JPEG fails.
>>>> > > >
>>>> > > > Please check if my code has some bug.
>>>> > > >
>>>> > > > And do you have any suggestions on KLV metadata approach?
>>>> > > >
>>>> > > > Thanks,
>>>> > > > Priyanka
>>>> > > >
>>>> > > > On Thu, Sep 26, 2019 at 7:11 PM Nicolas Dufresne <
>>>> nicolas at ndufresne.ca> wrote:
>>>> > > > >
>>>> > > > > Le jeu. 26 sept. 2019 05 h 25, priyanka kataria <
>>>> priyanka.kataria86 at gmail.com> a écrit :
>>>> > > > > > Hello,
>>>> > > > > >
>>>> > > > > > I have an interesting problem:
>>>> > > > > > Need to transfer some kind of metadata (say frame number)
>>>> with each frame over UDP. Receiver on the other hand, extracts he frame
>>>> numebr from each frame and maintains it for some other work.
>>>> > > > > >
>>>> > > > > > Sample sender and receiver pipelines:
>>>> > > > > > Sender: gst-launch-1.0 -v filesrc location=file.h264  !
>>>> h264parse ! rtph264pay ! udpsink port=5001
>>>> > > > > > Receiver: gst-launch-1.0 -v udpsrc port=5001
>>>> caps="application/x-rtp, media=(string)video, clock-rate=(int)90000,
>>>> encoding-name=(string)H264" ! rtph264depay ! decodebin ! autovideosink
>>>> > > > >
>>>> > > > > Consider adding an rtpjitterbuffer on the receiver, in future
>>>> GST version, you'll be able to use rtpsrc/rtpsink (or the rist variant) in
>>>> order to get a full feature RTP stream without complex pipeline
>>>> construction.
>>>> > > > >
>>>> > > > > For jpeg, consider configuring max-bitrate property on udpsink.
>>>> As frames are spread out on a lot more packet, it tend to become bursty and
>>>> may saturate the link or exhaust the udpsrc socket buffer-size.
>>>> > > > >
>>>> > > > >
>>>> > > > > > Things I have already tried (I am still a beginner, so some
>>>> of the below things may look stupid):
>>>> > > > > > 1. In Sender pipeline, attaching a probe on "h264parse"
>>>> element and assigning incremental values to "GST_BUFFER_OFFSET".
>>>> > > > > > But the set offset value is not reflected in the next element
>>>> in the same pipeline only.
>>>> > > > > >
>>>> > > > > > 2. In Sender pipeline, attaching a probe on "h264parse"
>>>> element and assigning incremental values to "GST_BUFFER_PTS".
>>>> > > > > > The set PTS value is reflected in the next elements in the
>>>> same pipeline, but gets lost across UDP.
>>>> > > > > > I checked this by attaching a probe on "rtph264depay" element
>>>> (src pad).
>>>> > > > > >
>>>> > > > > > 3. Using "gst_rtp_buffer_add_extension_twobytes_header()".
>>>> > > > > > This method works for H264 files, but fails with MJPEG files,
>>>> and my solution needs to be generic.
>>>> > > > > > Here, I can provide more details with code if required.
>>>> > > > >
>>>> > > > > That is the method I would have used. It should work with any
>>>> RTP packet, so you likely have or hit a bug.
>>>> > > > >
>>>> > > > > > 4. Last thing I am trying is to mux KLV metadata into stream
>>>> and send it across UDP.
>>>> > > > > > I refer the following link:
>>>> https://www.aeronetworks.ca/2018/05/mpeg-2-transport-streams.html.
>>>> > > > > > This doesn't work though as written in the article but gave
>>>> me an overview on how to use the pipeline.
>>>> > > > > > Now I want to create my custom my KLV metadata file which
>>>> contains only frame numbers and try to mux it.
>>>> > > > > >
>>>> > > > > > Please help me in creating such file.
>>>> > > > > >
>>>> > > > > > Also please share if there are any other working approaches I
>>>> should try to append metadata in each frame buffer.
>>>> > > > > >
>>>> > > > > > Thanks,
>>>> > > > > > Priyanka
>>>> > > > > > _______________________________________________
>>>> > > > > > gstreamer-devel mailing list
>>>> > > > > > gstreamer-devel at lists.freedesktop.org
>>>> > > > > >
>>>> https://lists.freedesktop.org/mailman/listinfo/gstreamer-devel
>>>> > > > >
>>>> > > > > _______________________________________________
>>>> > > > > gstreamer-devel mailing list
>>>> > > > > gstreamer-devel at lists.freedesktop.org
>>>> > > > > https://lists.freedesktop.org/mailman/listinfo/gstreamer-devel
>>>> > > >
>>>> > > > _______________________________________________
>>>> > > > gstreamer-devel mailing list
>>>> > > > gstreamer-devel at lists.freedesktop.org
>>>> > > > https://lists.freedesktop.org/mailman/listinfo/gstreamer-devel
>>>> > >
>>>> > > _______________________________________________
>>>> > > gstreamer-devel mailing list
>>>> > > gstreamer-devel at lists.freedesktop.org
>>>> > > https://lists.freedesktop.org/mailman/listinfo/gstreamer-devel
>>>> >
>>>> > _______________________________________________
>>>> > gstreamer-devel mailing list
>>>> > gstreamer-devel at lists.freedesktop.org
>>>> > https://lists.freedesktop.org/mailman/listinfo/gstreamer-devel
>>>> _______________________________________________
>>>> gstreamer-devel mailing list
>>>> gstreamer-devel at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/gstreamer-devel
>>>
>>> _______________________________________________
>>> gstreamer-devel mailing list
>>> gstreamer-devel at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/gstreamer-devel
>>
>> _______________________________________________
>> gstreamer-devel mailing list
>> gstreamer-devel at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/gstreamer-devel
>
> _______________________________________________
> gstreamer-devel mailing list
> gstreamer-devel at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/gstreamer-devel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/gstreamer-devel/attachments/20191004/bbbe6edc/attachment-0001.html>


More information about the gstreamer-devel mailing list