Efficient handling of RTP frames?

Efficient handling of RTP frames?

[Sam Ravnborg]

Hi all.

We (my company) are looking into a solution for transporting a high amount
of RTP packets across a single CPU. A 82xx, presumeably a 8270.
With high amount of RTP packets we are talking about 10 to 20.000 packets
per second with a payload size of ~64 bytes each (on top of UDP header).

The RTP packets needs to be sent between an external device connected
to one of the FCC's (needs to write our own driver here) and the Ethernet
port sitting on FCC1.

Does this sound like a dead-end doing this in Linux?

Can anyone point in a direction where to find more information
about how to realise this in the most efficiet way?

Google did not bring up anything useful.

Thanks in advance,

	Sam

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: Efficient handling of RTP frames?

[Dan Malek]

On May 10, 2004, at 4:58 PM, Sam Ravnborg wrote:


> Does this sound like a dead-end doing this in Linux?

No problem.  It will work fine.  Depending upon the processing
you are doing, you could probably do it with an 8xx.

> Can anyone point in a direction where to find more information
> about how to realise this in the most efficiet way?

Without any more knowledge of your protocol details, you'll
probably want to design the software to reduce the amount of
interrupts.  There are buffer management concerns, such as
alignment and location, but all of this depends on all of your
system requirements outside of this processing.  Where is
the RTP processing being done, on the 8xxx or the external device?

All of this is well within the capabilities of the processor and Linux.


	-- Dan


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: Efficient handling of RTP frames?

[Sam Ravnborg]

On Mon, May 10, 2004 at 05:16:40PM -0400, Dan Malek wrote:
>
> >Can anyone point in a direction where to find more information
> >about how to realise this in the most efficiet way?
>
> Without any more knowledge of your protocol details, you'll
> probably want to design the software to reduce the amount of
> interrupts.  There are buffer management concerns, such as
> alignment and location, but all of this depends on all of your
> system requirements outside of this processing.  Where is
> the RTP processing being done, on the 8xxx or the external device?

All (or most) procesing will be done outside the 82xx. The main
purpose of the 82xx will be to do a simple conversion to/from
RTP - since the external DSP will not use this format.

The alignment concerns you mention I expect can be addressed
we anyway have to develop a driver for the DSP connected to the FCC.
The buffer management needs a second thought.

But I still do not see how the two drivers can 'communicate',
or what other requirements is put on the driver that communicate
with the DSP.

	Sam

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: Efficient handling of RTP frames?

[Dan Malek]

On May 10, 2004, at 5:33 PM, Sam Ravnborg wrote:

> But I still do not see how the two drivers can 'communicate',
> or what other requirements is put on the driver that communicate
> with the DSP.

Since you are using RTP over UDP, you are going to use the
network stack, so the drivers aren't going to "communicate"
directly.  You can use some FCC/TDM interface to format the
data for the DSP, then use an application to do the rest.
The driver for the DSP will have to be something custom, then
you can configure the FCC Ethernet to handle small frames by
increasing the number of ring buffers (but don't go too high or
that overhead will get too high) and use the NAPI to reduce the
interrupt overhead.  Depending upon the performance requirements,
the custom DSP driver could be implemented to share buffers
directly with the application.


	-- Dan


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: Efficient handling of RTP frames?

[Eugene Surovegin]

On Mon, May 10, 2004 at 10:58:08PM +0200, Sam Ravnborg wrote:
> We (my company) are looking into a solution for transporting a high amount
> of RTP packets across a single CPU. A 82xx, presumeably a 8270.
> With high amount of RTP packets we are talking about 10 to 20.000 packets
> per second with a payload size of ~64 bytes each (on top of UDP header).

Hmm, I'm a little surprised by this number - 64 bytes. Most common is 20ms
RTP payload and for and ordinary voice traffic it yields 160 byte payload, but
probably your application is somewhat special :).

For high packet rates you might be interested in NAPI enabled network driver.

Eugene.

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: Efficient handling of RTP frames?

[Matevz Langus]

Hi,

I did the following in advance to all others said before:
left UDP CRC field empty, so you don't need to calculated CRC. VoIP
systems do not complain about that.

rgds,
  Matevz

On Mon, 2004-05-10 at 22:58, Sam Ravnborg wrote:
> Hi all.
>
> We (my company) are looking into a solution for transporting a high amount
> of RTP packets across a single CPU. A 82xx, presumeably a 8270.
> With high amount of RTP packets we are talking about 10 to 20.000 packets
> per second with a payload size of ~64 bytes each (on top of UDP header).
>
> The RTP packets needs to be sent between an external device connected
> to one of the FCC's (needs to write our own driver here) and the Ethernet
> port sitting on FCC1.
>
> Does this sound like a dead-end doing this in Linux?
>
> Can anyone point in a direction where to find more information
> about how to realise this in the most efficiet way?
>
> Google did not bring up anything useful.
>
> Thanks in advance,
>
> 	Sam
>


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: Efficient handling of RTP frames?

[sam]

Tir, 11 May 2004 10:28:02 +0200 skrev Matevz Langus <matevy@arhitektova-delavnica.com> :

>Hi,
>
>I did the following in advance to all others said before:
>left UDP CRC field empty, so you don't need to calculated CRC. VoIP
>systems do not complain about that.
Thanks for this tip, unfortunately we already use this 'feature'.

We are operating with 5 ms samples so we need to be very careful about any overhead (present implementation is running in a VxWorks environment).

   Sam

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/