RFsim: don't shell out to improve kernel socket buffer sizes, teach the user instead
A later commit in this series adds performance tuning tips, which includes the call made in RFsim until this commit.
This commit is contained in:
@@ -174,3 +174,18 @@ To successfully establish a connection with such a GEO satellite channel, both g
|
||||
# Caveats
|
||||
|
||||
There are issues in power control: txgain/rxgain setting is not supported.
|
||||
|
||||
# How to improve performance
|
||||
|
||||
Most importantly, note that the RFsimulator is not designed to be as performant
|
||||
as possible, nor is it designed to run close to real-time. It might run faster
|
||||
or slower than realtime, depending on CPU, and by design, as this allows to
|
||||
stop the entire system for inspection, e.g., using a debugger.
|
||||
|
||||
In order to improve performance, you can modify the radio parameters of the gNB
|
||||
to reduce the amount of transported samples:
|
||||
|
||||
- Use option `-E` for three-quarter sampling (also to be done on the UE-side!)
|
||||
- Prefer smaller cell bandwidths
|
||||
|
||||
A possible, unimplemented optimization would be to compress samples.
|
||||
|
||||
Reference in New Issue
Block a user