High latency, low throughput

[Juan HM]

Hi everyone,

Is there a reason in a 5G access network end-to-end Simulation, why the latency for small UDP packet size (20KB/s) is experiencing latency of 107ms, while the same escenario but with 4G LTE access network is only 21 ms?

Could this be some header or power implication?

I am simulating 7 sites and 3 sectors each and 30 users per sector. EPC core with 1 PGW and 1 SGW.

BR.

[Gabriel Ferreira]

We need more details to better especulate the issue.

[Juan HM]

Hi Gabriel,

This is a TDD communication with fram distribution DDDSUUDDDD, non- ovelapping case, Numerology 1 SCS 30kHz. Kindly Let me know if you need further details.

On a second point, I see also high latency, how could I enable the delay visualization, I have enabled the logging, but is only telling me the delay in each packet.

BR

[Gabriel Ferreira]

For visualization you need to write your own script to plot it, or use Netsimulyzer (separate module).

Low traffic, high latency. That much additional latency, I'd guess some RLC queueing. If it was HARQ, it would be a lot less.
Possibly because the transmit power is too low? What is the TBLER?

If that is not it, then it could be the number of users per cell and the scheduler.
It can result in starvation in some cases, with proposed fixes meant for NR 4.0 here https://gitlab.com/cttc-lena/nr/-/merge_requests/225
It is easy to check, if the distribution of delays between different users is not uniform.

[Kent Huns]

Hi Juan and Gabriel,

>> it could be the number of users per cell and the scheduler.

If this is the cause, there should be only one UE achieves high throughput & low latency. And the other 29UEs are in bad condition.

[Gabriel Ferreira]

>  If this is the cause, there should be only one UE achieves high throughput & low latency. And the other 29UEs are in bad condition.

Hi Kent. Not really. The issue you correctly pointed out in memory causes TDMA to allocate, e.g. for round-robin, one symbol per UE. Since memory is lost between slots, in the next slot, same UEs get one symbol each if #UEs > #symbols. For OFDMA it can be  just as bad, but at least you have multiple RBGs to further divide allocation, so in theory should happen less frequently. But then you also have the symbols per beam allocation, which can cause a similar issue. I've implemented two alternative symbols per beam allocation, based on round-robin (with memory), and proportional-fair policy.