Space-Time Processing—Linux Style

Developing a new generation of wireless communications means you need FPGA development tools, a cluster for simulation and an embedded OS for prototype devices. Linux fits the need.

Here at Tait Electronics Group Research in Christchurch, New Zealand, we quietly have been building an advanced wireless networking concept called space-time (ST) processing. ST is predicted by many to drive the next big-step improvement in wireless technology after 3G. The nature of space-time is that the mathematics behind it and, consequently, its implementation, are extremely complex. Many academic researchers are working on this, but we have completed two world-first practical implementations on our ST array research (STAR) platform. These are time-reversal space-time block coding (TR-STBC) and the tongue-twisting single carrier, adaptive multivariate decision feedback equalised multiple-in, multiple-out (SC-AMV-DFE-MIMO) coding scheme. Without a doubt, these achievements were possible largely thanks to the performance and adaptability of Linux.

This article describes how Linux was key in enabling our mathematical simulations and was adopted for our embedded processing and runtime control. We cover such diverse aspects as message passing interface (MPI), cluster computing, embedded Linux, PHP, shell scripts, network filesystem (NFS), SMB (server message block) and kernel modules on ARM, Alpha and Intel architecture computers. We skip development details, but describe the system as it operates today, highlighting some of the real advantages we gained by using Linux.

In its basic configuration, each STAR platform consists of one multichannel transmitter unit and one multichannel receiver unit, both connected to a shared LAN and transmitting anything up to around 200Mb/s at microwave frequencies. There actually are 23 printed circuit boards in each unit, which took 15 people a year to design and build.

Figure 1. A bidirectional RF link uses one multichannel transmitter and one multichannel receiver on each STAR platform.

Figure 1 shows the system wired to a shared Ethernet as a bidirectional RF link. The setup shown is for experimental purposes; an actual product would not have a shared Ethernet between transmitter and receiver.

All of the clever radio processing is done on the digital board, not by the ARM processor but by a dedicated field programmable gate array (FPGA) and digital signal processor (DSP). We call FPGA code firmware because we can't decide whether it's software or hardware; this way we don't have to comply with either the software coding standards or the hardware design standards of the company. In the middle of last year, when we designed the system, we used the biggest, fastest and most expensive FPGA and DSP we could obtain, and we since have added two more big FPGAs. The ARM is not used for low-level processing because we require over 50,000MIPS for ST processing. With this complexity, even the fastest components alone are not enough, so we decided early on to build a multiprocessor-capable system.

STAR units are extensible through high-speed low-voltage differential signaling (LVDS) connections that run at up to 1Gb/s. Each board has two sets of five-channel bidirectional LVDS interconnects to link two neighbouring boards. At the same time, each board is wired to Ethernet. High-speed data travels over LVDS, and control data uses Ethernet.


Most processing occurs in the FPGA, coded in VHDL. A growing number of VHDL tools are available for Linux (see sidebar), and we used Quartus from Altera. In our system, we developed algorithms first with GNU-Octave and then ported them to VHDL. Octave and the mostly compatible MATLAB are available for Linux and Microsoft Windows, but Octave has an MPI-enabled version that runs on clusters. We compiled this for a cluster of old DEC Alphas we call zion. MPI-Octave really screams on zion despite the fact that the fastest CPU is only 500MHz.

For our digital board, we used an ARM Linux toolchain from to compile freshly patched 2.4.18 kernel sources. Russell King began ARM Linux when working on a distribution for Acorn computers in the early 1990s. Now, ARM is one of the best-supported processors under Linux and is a real joy to us. Three days was enough to port Linux to our custom hardware, although the Ethernet driver took a couple more days to complete. ARM is easily the world's number one processor by sales, with a huge amount of support for running Linux. This led Tait Electronics to commit to ARM processors in our future processor road map.

With ARM Linux booting, it was time for a RAM disk. Having 16MB of SDRAM and 2MB of Flash memory on the ARM, we opted for a maximum compressed kernel and RAM disk size of 1,024K each, although the uncompressed RAM disk is 4MB in size. Both are stored in Flash and allow booting without external interaction.

We opted to use BusyBox for common tools, including ls, cd, mount, insmod and ping, and TinyLogin for login and password support. Other onboard utilities handle memory-mapped peripherals and Flash, and netkit-base provides a telnet dæmon that uses TinyLogin functions. Tait Electronics bought a range of MAC addresses from the IEEE to support its Linux-on-ARM development programme, and the MAC address for each board is held in Flash memory.

Figure 2. Each unit has both a Flash-based and a RAM-based filesystem.

We developed many small applications and even got the Abyss Web server running, but these all can't fit into Flash. We therefore NFS-mount a directory from the zion mainframe to each ARM, accessing many gigabytes of extra space. Figure 2 shows the filesystem arrangement.

We also set up SMB mounts to users' shared drives from zion, accessible to the ARM boards over NFS. If a Web server is run on the ARM boards, we end up with a highly interlinked system. Windows users can access part of their own local drivespace through the Web served by way of an embedded ARM board linked over NFS to a remote cluster server.

Figure 3. The Filesystem Tree as Seen from the Unit