[ENet-discuss] not enough throughput when packet loss is added

[Lee Salzman]

If you're shuttling that much primarily reliable data back and forth,
you should really just be using TCP. You will get the best performance
that way.

On Fri, Nov 24, 2017 at 1:56 PM, Chris Barrett <cbarrett.colo at gmail.com> wrote:
> I have a streaming uncompressed audio application.  It is a round trip path
> from node A to node B and then back to node A.  A is setup as a client.  B
> is setup as a host.  The links can have up to three 48k/24bit PCM audio
> channels - thus 1.1 Mbit each, plus one setup/control channel with
> negligible bandwidth requirements.  This application cannot have errors, so
> enet packets are all setup as reliable.
>
> I have tried these cases:
>
> (1) One audio channel, no packet drops, no network delay - works great.
> (2) One audio channel, 0.85% packet loss in each direction, 120ms network
> delay in each direction - works great with ~900ms buffer at both A and B
> (3) Two audio channels, no packet drops, no network delay - works great.
> (4) Two audio channels, 0.85% packet loss in each direction, 120ms network
> delay in each direction - fails consistently with not enough throughput.
> Thus, the buffers at either node deplete relatively quickly (~2 seconds) and
> the stream is broken, falls behind.
>
> I have set both enet hosts up for infinite bandwidth.  I believe the network
> bandwidth capacity is very large, > 100 Mbit.  I setup the network
> conditioner (that is simulating the packet loss and network delay) to have
> 60 Mbit bandwidth...so I don't think it is the problem.
>
> At both nodes I have enet_host_service setup in it's own task, so I believe
> enet is always able to perform any necessary work.
>
> Any ideas?  I see there is some sort of throttling algorithm in enet, but it
> appears to be used to limit unreliable communication (?).  Any ideas?
>
> Thanks,
> Chris
>
> _______________________________________________
> ENet-discuss mailing list
> ENet-discuss at cubik.org
> http://lists.cubik.org/mailman/listinfo/enet-discuss
>