Uplink Multi-User MIMO Testbed Implementation in OpenAirInterface ^††thanks: The authors acknowledge the financial support by the Federal Ministry for Research, Technology and Space (BMFTR) in Germany in the programme of “Souverän. Digital. Vernetzt.,” joint project xG-RIC, project identification numbers: 16KIS2429K and 16KIS243, and by the 6G-MIRAI project, which has received funding from the Smart Networks and Services Joint Undertaking (SNS JU) under the European Union’s Horizon Europe research and innovation program under Grant Agreement No 10119236. Views and opinions expressed are however those of the author(s) only, and they do not necessarily reflect those of the European Union or the SNS JU (granting authority). Neither the European Union nor the granting authority can be held responsible for them. The authors thank Ralf Lindstedt, Angelo Athanassopoulos, and Andreas Schiller of Fraunhofer Heinrich Hertz Institute (HHI) for their valuable technical assistance during the building and testing of this system.

The MIMO system model with a flat-fading channel with $K$ transmit and $M$ receive antennas is given by

where $\bm{y}\in\mathbb{C}^{M}$ is the signal received from the receive (Rx) array, $\bm{x}\in\mathbb{C}^{K}$ is the transmit signal vector, $\bm{H}\in\mathbb{C}^{M\times K}$ is the MIMO channel matrix, and $\bm{n}\sim\mathcal{C}\mathcal{N}\left(\bm{0},\sigma^{2}\bm{I}_{M}\right)$ is the additive white Gaussian noise vector. Each variable corresponds to one OFDM subcarrier. The subcarrier indices are dropped for the sake of brevity. For the case of MU-MIMO, $\bm{x}$ is a column vector where each element is the transmit symbol of a separate user. The receiver needs to estimate $\bm{x}$ by applying a linear operation to $\bm{y}$ as

where $\hat{\bm{x}}$ is the estimated transmit vector and $\bm{V}^{\mathsf{H}}\in\mathbb{C}^{K\times M}$ is the MIMO combiner matrix. We consider two common MIMO detection algorithms; namely, Maximum Ratio Combining (MRC) and Regularized Zero Forcing (RZF), $\bm{V}^{\mathsf{H}}_{\mathrm{MRC}}=\bm{H}^{\mathsf{H}}$, and $\bm{V}^{\mathsf{H}}_{\mathrm{RZF}}=\left(\bm{H}^{\mathsf{H}}\bm{H}+\lambda\bm{I}\right)^{-1}\bm{H}^{\mathsf{H}}$, where $\bm{I}$ is the identity matrix and $\lambda$, a nonnegative real number, is the regularization parameter.

The model in (1) assumes that the receiver has perfect channel state information. The residual carrier frequency offset between the transmitter and the receiver becomes an important factor when periodic pilots independent of the user data are used to calculate $\bm{V}$. Due to the time difference $\tau_{k}$ between the periodic pilot transmission of the user $k$ and the data symbol transmission, the channel vector of each user undergoes a phase rotation $\theta_{k}$, proportional to the frequency offset and $\tau_{k}$. Here we assume that the channel has a coherence time longer than the maximum $\tau_{k}$, and the channel has not changed much except the phase rotation. We can represent the phase rotation with the diagonal matrix $\bm{\Theta}=\operatorname{diag}(e^{j\theta_{1}},\ldots,e^{j\theta_{K}})$ from the individual phases. As $\bm{H}$ represents the channel state in the instance of channel estimation, the model during data transmission becomes

Applying RZF gives us

which is the user signal vector with canceled interference, but with additional phase rotation. The noise vector is transformed as $\tilde{\bm{n}}=\bm{V}^{\mathsf{H}}\bm{n}\sim\mathcal{C}\mathcal{N}\left(\bm{0},\sigma^{2}\bm{V}^{\mathsf{H}}\bm{V}\right)$. The remaining phase rotation can be estimated using a single-layer channel estimation from the demodulation reference signal (DMRS) pilots embedded inside the data symbols and equalized by rotating the symbols back accordingly.

The software implementation of this project is built as a private feature branch of the Open Air Interface [oai2014] project. Our implementation is based on the 2024.w45 version of the official repository. Before detailing the PHY implementation of the MU-MIMO processing, we will provide details on how OAI handles PHY Rx processing. As summarized in Algorithm 1, PHY functions for the gNB receiver run in their own thread. The main worker function of this thread, phy_procedures_gNB_uespec_RX is called for each uplink slot and processes every physical channel scheduled to that slot. It is implemented as three consecutive for-loops. The first loop iterates over all Physical Uplink Control Channel (PUCCH) instances expected to arrive in that slot. The same operation is performed for Physical Uplink Shared Channel (PUSCH), which carries the user data. Lastly, SRS is processed, whose function will be detailed in the next section.

The DecodePUSCH procedure performs channel estimation and equalization using PUSCH DMRS, soft quadrature-amplitude-demodulation, descrambling, and channel decoding to recover MAC data payload. Similar operations are performed for the control channel in DecodePUCCH routine. As the name suggests, the SRSChannelEstimation routine estimates the wireless channel from SRS pilots.

The SRS is a physical channel designed to transmit pilot symbols from users to the base station to perform uplink channel measurements. It can be configured to be sent periodically from each user in a wide bandwidth. SRS plays an important role for our system because we use it to estimate channel state information to compute the MU-MIMO combiner ($\bm{H}$ and $\bm{V}^{\mathsf{H}}$ from Section III).

There are several options to multiplex SRS from multiple users. In this project, we use a purely TDD approach for simplicity of implementation. Each user is scheduled to transmit a wideband pilot sequence for one OFDM symbol in a different slot. The pilots are transmitted twice per radio frame (every $5\text{\,}\mathrm{ms}\text{/}$) and each user transmits their pilots in adjacent slots. Thi results in a time offset of approximately $0.5\text{\,}\mathrm{ms}\text{/}$ between the channel measurement instances of the first and second user.

As we now use SRS channel estimates for data decoding, it is important to use fresh channel state information as soon as it is available. To achieve this, we reversed the order of the PUSCH and SRS loops in the receiver function. As soon as new channel information arrives, the RZF combiner matrix is recomputed with up-to-date channel state information.

We implemented the multi-user interference suppression as a pre-processing step to the regular PUSCH decoding routine. Once the MU-MIMO combiner is applied to the receive vector, the rest of the processing can be performed as if the signal were coming from a single antenna without interference. The modified PHY Rx algorithm is given in Algorithm 2. With this approach, the OAI codebase for PUSCH processing remains mostly unmodified, while adding MU-MIMO interference cancellation capability. As DMRS channel estimation and equalization are part of this process, the phase rotation explained in Section III-B is corrected.

The design choice of using SRS channel estimates for computing the combiner instead of PUSCH-DMRS can be explained as follows: The next milestone for this system is to perform downlink MU-MIMO precoding by leveraging the uplink-downlink channel reciprocity in a TDD system. A parallel project is underway to develop a calibration step to equalize the mismatch between the transmit and receive radio-frequency frontends. In order to achieve this, periodic updates on the channel state information are necessary to keep the downlink precoder matrix up-to-date. This would not be possible if the uplink pilots were only sent when there is uplink data to transmit. Our work shows that SRS can be configured and used to decode data symbols on the uplink. This was a necessary validation step before using SRS-based channel estimates to compute the downlink precoders.

The DSP operations of OAI’s PHY use fixed-point arithmetic. The reason for this design choice is real-time performance, as processors perform integer operations faster than floating-point operations. Fixed-point operations can also be parallelized more efficiently with single-input-multiple-data operations. However, implementing complex algorithms with integer arithmetic is cumbersome. The designer must ensure that all the signal levels are appropriately scaled to minimize overflows and underflows. Inverting a matrix is especially challenging in this case, as division by a small determinant might easily cause an overflow. Because of that, we implemented our functions using floating point arithmetic. After converting the SRS channel estimates and the received IQ samples to floating point representation, we computed the RZF combiner matrix and applied it to the received signal. We finally converted the equalized signal back to integers as OAI’s PUSCH processing function expects an integer signal as input.

Our experiments with network benchmarking tools showed that this approach was fast enough to keep up with the real-time processing constraints. However, it is possible to further optimize this routine to balance speed and numerical accuracy. The idea behind this strategy lies in the following observation: The combiner is computed only when new channel state information arrives, while every received data symbol must be equalized before it can be decoded. This means that the combiner matrix is used much more often than it is computed. The strategy is to keep the combiner computation in float, but to perform matrix vector multiplication between the combiner matrix and the receive vector using integers. In this way, the step that is prone to errors with integers remains float, but the step that uses most of the computer cycles is optimized with integers. We leave this idea for the next step of development in future work.

In this project, we experimented with commercial UE as well as OAI software UE with USRPs. MU-MIMO scheduling and the processing done by the gNB is transparent to UEs. From a UE’s perspective, a radio resource control (RRC) command configures SRS transmission and when the user has uplink data to transmit, a PUSCH grant is issued to them in a certain time and frequency resource. The UE does not know if another user is also transmitting in the same set of resources. As a result, there is no need to modify the UE to enable uplink MU-MIMO.

The main difference between a commercial UE and OAI UE was the possibility of using a reference signal for frequency stability. We observed that OAI software UEs are more stable when their local oscillators were conditioned with an external $10\text{\,}\mathrm{MHz}\text{/}$ reference signal or with Global Positioning System (GPS). Commercial UEs do not have the possibility to accept an external reference signal for frequency synthesis. Although they could also connect to the cell, they lost synchronization and disconnected when high traffic was sent.

SRS is usually used to measure long-term channel statistics. Because of that, its period is typically configured to be relatively long. In OAI, the default period of SRS is 8 radio frames ($80\text{\,}\mathrm{ms}\text{/}$). To track the channel for multi-user interference management, we modified the SRS scheduler so that it configures UEs to transmit SRS pilots twice per radio frame ($5\text{\,}\mathrm{ms}\text{/}$ period).

There are several options to multiplex SRS from multiple users; such as TDD, frequency interlacing with a comb pattern, and a mixture of these two approaches. For simplicity, we choose a TDD approach where each user transmits SRS in neighboring slots.

The bandwidth of SRS is represented in terms of PRBs. Due to the way RRC signaling for SRS configuration is standardized by 3GPP, it is usually not possible to configure SRS to the exact system bandwidth. For example, 20 PRBs is the maximum SRS bandwidth in a 24 PRB system. The missing PRBs at the edge do not get pilots to compute MU-MIMO combiners. We deactivated the PRBs without SRS during PUSCH scheduling, so no data is scheduled for those PRBs. A simplified representation of the uplink OFDM scheduling can be seen in Fig. 2.

To enable nonorthogonal scheduling of PUSCH to multiple users, we modified OAI’s PUSCH scheduler, which is implemented in the pf_ul function inside gNB_scheduler_ulsch.c file. Instead of marking a PRB as occupied when one user is scheduled to it, PRB allocation is tracked separately for both MU-MIMO layers. This leads to reusing time-frequency resources when scheduling both users and increasing the link capacity.