This commits contains several fixes to improve a multi-UE scenario.
This is not the end of the story.
Summary of work:
================
1 - improve SR (scheduling requests):
We configured n1pucch == 3 for scheduling requests, for all
UEs. We now use 71 - UE_id/10.
For it to work, it is vital that pucch_nRB_CQI in the configuration
file is set to 0, otherwise the SR will go to an RB used for
PUSCH and uplink decoding will fail whenever an UE does SR.
Note that we will have problems with 20MHz when we use a CCE that
let the UE send the ACK/NACK using a n1pucch allocated for SR,
because when the PDCCH is of size 3, we can have 87 CCEs
and it may lead to an n1pucch colliding with one for SR.
The work done in this patch is a quick solution, seems to work
with 10MHz.
The real solution is to disable the use of those CCEs that would
lead an UE to use a n1pucch colliding with an SR n1pucch. Then
we can use whatever n1pucch we want for SR, as long as the
scheduler protects them.
Impacted files: configuration files
openair2/RRC/LITE/MESSAGES/asn1_msg.c
2 - some fixes for uplink scheduling:
- Do not use PRACH for PUSCH, that leads to too many false
PRACH detection. Plus the PUSCH receiving may fail if one
UE uses the PRACH at the same time.
- Take care of retransmissions. That was not done at all, so
we could allocate one RB to several UEs. The current design
of the code makes it hard to do it properly, so we chose a
quick and dirty solution, which is to increase "first_rb"
to skip any RB used for retransmission. In this process we
may skip a lot of free RBs. A proper solution is needed here.
- Do not allocate the last RB. This one is used for PUCCH.
It was sometimes allocated to PUSCH.
- In the pre-processor we didn't pre-allocate RBs to UEs
with an empty buffer status. We didn't check if the UE
sent an SR. For example in a three UEs scenario, we
could have the third UE never scheduled in the uplink.
- rb_table_index was not decreased properly, so we allocated
too much RBs to some UEs and thus not enough to others.
Impacted files: openair2/LAYER2/MAC/eNB_scheduler_ulsch.c
openair2/LAYER2/MAC/pre_processor.c
openair1/SCHED/phy_procedures_lte_eNb.c
3 - some fixes for downlink scheduling:
- The check on CCE allocation was not correct. We did something
like:
if (cce allocation is possible) {
prepare
}
We should have done:
save current cce allocation
if (cce allocation is possible) {
allocate cce
prepare
}
reset current cce allocation
Basically, when we scheduled several UEs, they were checked
separately, and the totality of them was not checked.
Impacted file: openair2/LAYER2/MAC/eNB_scheduler_dlsch.c
- The retransmissions are probably not handled correctly.
Check in openair2/LAYER2/MAC/pre_processor.c, near
the comment "// control channel or retransmission",
the case "round > 0" was added. It's probably not enough,
even maybe not correct.
- Change SF05_LIMIT with SF0_LIMIT. We accept to use
central blocks in SF 5. The code was also not correct,
vrb_map was not set properly because the loop on j
was wrong for the last RBG (which can have one less
RB than the others).
This is not satisfying. The real solution is to use the
central RBs and check that the MCS used is compatible
with the numbers of resource elements allocated (we don't
want to put too more data bits than what fits).
4 - some fixes in PUCCH decoding:
See: openair1/PHY/LTE_TRANSPORT/pucch.c
Probably not enough. Some more work and analysis is
required for a proper use of the PUCCH. What we see
is that the PUCCH constellation gets wrong when there
are several UEs, meaning the received ACK/NACK is
not properly decoded (this, or something else...).
5 - several fixes/checks added here and there:
- The final allocate_CCEs in eNB_dlsch_ulsch_scheduler
is checked and we brutally exit if it fails.
- We exit in get_num_pdcch_symbols in case of failure
(this should never happen anyway, no big deal normally).
- Some logs added or changed to error/warning instead
of debug.
- In dlsch_scheduler_pre_processor an abort() was added.
The code here looks suspicious.
- In assign_max_mcs_min_rb, rb_table_index was not set
to 2, the value 0 was used. This was not correct.
What remains to be done:
========================
- Correct CCE allocation (take into account SR n1pucch,
check that all the n1pucch that will be used are "compatible").
- Take into account the PHICH when scheduling uplink. As of
today it is very possible to have two UEs use the same PHICH
group and PHICH sequence index. We can use n_DMRS in the DCI
to have uniqueness (see 36.213 table 9.1.2-2). We can drop an
allocation if there is no free PHICH group/sequence index for
a given UE.
- When there is an uplink retransmission in the PRACH, we have
to disable PRACH detection. It is possible that one UE does
PRACH at the same time, but then what to do? We could use
DCI0-based retransmission in this specific case maybe...
- Handle free RBs in uplink in a much better way in case of
a retransmission. We may have a lot of free unused RBs with
the current code.
- Check downlink retransmissions. Not much has been done there.
- Surely more stuff not clear yet. In some situations we don't
have a good behavior of the system. Hard to describe more
precisely at this point.
How to use:
1 - compilation of softmodem:
./build_oai --eNB -t ETHERNET
2 - compilation of mobipass driver:
cd cmake_targets/lte_build_oai/build
make oai_mobipass
ln -sf liboai_mobipass.so liboai_transpro.so
3 - configuration:
edit the configuration file, set "node_timing" to
"synch_to_mobipass_standalone"
that is, have the line:
node_timing = "synch_to_mobipass_standalone";
4 - run:
run as usual: sudo ./lte-softmodem -C <configuration file>
protobuf-c does not compile anymore.
Let's handle this a bit better.
We now install protobuf and protobuf-c only for the
flexran agent. That is, if you want to use the flexran
agent, you need to install protobuf/protobuf-c and
you do it this way:
./build_oai -I -a
(you add -a)
Other targets don't need protobuf nor protobuf-c, so
it's not installed by the -I command of build_oai,
unless you pass -a with -I.
Also, we now use protobuf 3.3.0, not 2.6.1. The code
has been adapted, a quick test seems to indicate that
the system works, but it has not been intensively tested.
several problems were present:
- there was no comma after "ue0_trx_write_ns_missing" in the array
eurecomVariablesNames;
comma was put, and also commas for the last element in the array,
which doesn't hurt and will prevent future problems
- bad order of values in eurecomVariablesNames, which
was different from the enum vcd_signal_dump_variables;
order was checked and fixed
- strange/wrong use of VCD_SIGNAL_DUMPER_MODULE_END/LAST;
the whole logic was removed/simplified
Summary of changes:
- UE: new thread idx
- UE: Pdcch optim
- UE: slot0 slot1 parallelization
- bugfixes for RLC AM (see gitlab issue 250)
- fix bug "unknown UE_id for rnti"
Note: due to the new UE threading architecture,
oaisim and the phy simulators may not work properly
anymore. Adaptation in the code has been done, automatic
tests seem to pass, but it may not be enough.
See merge request !215
With the new threading architecture of the UE dlsim
(and others) does not work properly anymore.
When looking at the scope, you see a difference
in PDSCH LLR display. The end is always 0 where
in the current develop branch (tag 2017.w25) it's not.
This commit attempts to fix it.
We still don't have the same behavior as in 2017.w25.
I disabled channel simulation (so that UE RX = eNB TX)
and I have one error where in 2017.w25 I have zero.
For example, here comes the output of a run of "./dlsim":
**********************SNR = 0.000000 dB (tx_lev 51.000000)**************************
Errors (1(0)/1000 0/1 0/0 0/0), Pe = (1.000000e-03,0.000000e+00,-nan,-nan), dci_errors 0/1001, Pe = 0.000000e+00 => effective rate 99.900100, normalized delay 0.001472 (1.001000)
And in 2017.w25 we have (with the same hack to disable
channel simulation):
**********************SNR = 0.000000 dB (tx_lev 51.000000)**************************
Errors (0(0)/1000 0/0 0/0 0/0), Pe = (0.000000e+00,-nan,-nan,-nan), dci_errors 0/1000, Pe = 0.000000e+00 => effective rate 100.000000, normalized delay 0.001471 (1.000000)
There may be a problem somewhere. Or there was one before and we should
have had one error and the new UE architecture fixed things and now
it's as it has to be. Hard to say at this point...
When looking at the scope we quickly see some zeros for the PDSCH
LLR, at the begining this time, not at the end. This is just when
the GUI appears and then all is fine, so this seems to be for the
first frame only. In 2017.w25 this does not happen.
With the current implementation of oaisim
(rxdata and channel simulation), we cannot
call trx_read_func on a dummy buffer. The
code will actually modify the rxdata buffers
of the UE.
This is has to be rewritten properly. In the
meantime, let's introduce a simple hack. The
idea of the read at this point is to wait for
the synch to finish and not lose samples from
the hardware in the real UE. In the simulator,
as it is today, we can simply sleep until the
synch code has finished its work.
In case of oaisim, dl_phy_sync_success has to be called
by initial_sync, as it used to be.
We introduce an #if OAISIM, this is not elegant, but will
do it for the moment.
This bug happens when we detect uplink failure for one UE.
In this case, a DCI format 1A is sent to the UE to ask it
to do random acces.
The way this DCI is generated was not compatible with how
the software is organized. It was expected that the DCI are
added (with add_ue_spec_dci and add_common_dci) in a very
specific order: first all DCIs in common space are added
(with add_common_dci) then all DCIs in UE specific space
are added (with add_ue_spec_dci).
The problem was that the DCI format 1A DCI sent to the UE
for it to do random access is added (with add_ue_spec_dci)
before the DCIs in common space.
That totally messed up the logic in add_common_dci and
add_ue_spec_dci.
The solution is to get rid of Num_common_dci and Num_ue_spec_dci,
replace those two counters by only one (Num_dci) and add
"search_space" in the dci_alloc structure to be used later by
the function "allocate_CCEs" when calling "get_nCCE_offset".
The software had to be adapted to the new variables, everywhere.
I am not sure that the simulators work. It seems that some
of them didn't use Num_common_dci and Num_ue_spec_dci to
decide on what space (common or UE specific) to put the DCI,
but relied on the rnti (comparing with SI_RNTI). To be tested
properly.
The modified simulators are:
- openair1/SIMULATION/LTE_PHY/dlsim.c
- openair1/SIMULATION/LTE_PHY/dlsim_tm4.c
- openair1/SIMULATION/LTE_PHY/dlsim_tm7.c
- openair1/SIMULATION/LTE_PHY/framegen.c
- openair1/SIMULATION/LTE_PHY/pdcchsim.c
- openair1/SIMULATION/LTE_PHY/syncsim.c
Cedric Roux