4.1 Design and construction

The overall design and construction of each antenna was carried out by the 54th Research Institute of China Electronics Technology Group Corporation (CETC54). Each antenna was pre-assembled in a large factory facility before being fully dismantled and packed into three shipping containers for delivery to the MRO site.

The antennas are 12 m diameter parabolic reflectors with a quadruped that holds a PAF at the prime focus. They have a focal ratio (f/D) of 0.5 and move on three axes; azimuth, elevation, and roll. The antennas can be driven over a continuous azimuth range of 179° –359° and an elevation range of 15° –89°. The roll axis (or parallactic angle axis) rotates the entire reflector about the optical axis with a range of − 179° to 179° (Forsyth et al. Reference Forsyth, Jackson, Kesteven, Yifan, Zanming, Zhenguo and Minxiang2009). This unique third axis of rotation allows the orientation of the reflecting surface, quadruped and PAF to be kept fixed with respect to the sky and has sometimes been referred to as a ‘sky mount’. This design was preferred over an equatorial mount as it was more economical, easier to construct and provided the significant benefit of keeping all beams stationary with respect to the antenna structure when tracking a source across the sky.

The drive system can move at a maximum rate of 3 deg s^{− 1} in azimuth and polarisation, and 1 deg s^{− 1} in elevation. Due to the finite slew rates, there is a small zone around the zenith through which the telescope cannot track while maintaining a lock on a source. The tracking system has a positional accuracy of 5 arcsec and the antennas are rated to operate in wind conditions with gusts up to 60 km h^{− 1}. The TOS will automatically stow the array if wind conditions exceed safe operational limits.

The roll axis eliminates the computational load that would have been required to track position angle using beamforming and reduces the impact of any systematic errors resulting from non-circular symmetry in the beam side lobes. This allows for a more uniform field of view and has greatly simplified commissioning tests in the early stages of hardware deployment. It also provides a unique opportunity to characterise beam patterns and polarisation characteristics. For instance, an observation could be repeated with the orientation of the antenna set to a different position angle. This will provide information on systematic trends in instrumental polarisation and hopefully improve the polarisation performance of the full ASKAP array.

For a description of the mathematical formalism of radio astronomical polarisation and instrumental calibration techniques, please see Hamaker et al. (Reference Hamaker, Bregman and Sault1996) and Sault et al. (Reference Sault, Hamaker and Bregman1996). In addition, several EVLA MemosFootnote ⁸ (in particular, 131 and 177) have discussed measurements of instrumental polarisation and its impact on the dynamic range of radio interferometers.

4.2 Measured performance

Photogrammetry experiments show that the surface accuracy of the antennas is very good; root-mean-square deviations from the best-fit paraboloid are typically less than 0.5 mm at all elevations, for all antennas. This will allow efficient operation up to 20 GHz if necessary in future.

The 12-m diameter of the primary reflector corresponds to a diffraction-limited beam width at half power of approximately 2° at the lowest observable frequency, decreasing to 0.8° at the high-frequency end of the band. Figure 1 shows a set of beam profiles from one of the BETA antennas, measured using a drift scan through three simultaneous beams formed at 1-degree intervals along the drift axis using the maximum sensitivity algorithm (see Section 5 for more information). Auto-correlations on the Sun were recorded for a single 1 MHz channel at a sky frequency of 860 MHz to estimate the beam profile. This example measurement does not make use of interferometric methods and is intended as a demonstration only. The nature of the formed beams (and indeed, their combined properties over the full field of view) depends heavily on the beam-forming technique used, which is still an active area of research.

Figure 1. Drift-scan cut through a set of three beams formed with maximum sensitivity weights, at a frequency of 860 MHz over a bandwidth of 1 MHz.

It is interesting to note that the off-axis primary beams exhibit an asymmetry in the level of the first side-lobe, which is a side-effect of the maximum sensitivity beam-forming algorithm. The imaging software currently makes the assumption that all beams are identical and deviations from this assumption may result in dynamic range limitations. It is possible to constrain the shape of a beam when forming a weight solution, though this will likely to be at the cost of sensitivity (see Section 5.2). An important use of BETA will be to quantify these trade-offs and establish optimal beam-forming methods tailored for specific astronomical survey strategies. Future publications will describe the results of these ongoing studies.

4.3 Array layout

The six BETA antennas were chosen to provide a mix of relatively short baselines with small fringe rates for ease of commissioning, and slightly longer baselines to provide spatial resolution. Figure 2 shows the relative position of each antenna on a scale diagram. Table 2 shows the distance between each pair of antennas in the array and Figure 3 shows a histogram of the baseline distribution. The longest baseline is 916 m (from antenna 6 to 15) and the shortest is 37 m (from antenna 1 to 3).

Figure 2. Scale diagram of the relative geographic positions of the BETA antennas, with selected baseline distances marked.

Figure 3. Histogram showing the distribution of baseline lengths, with ten bins of 100 m width.

Table 2. Baseline lengths for all pairs of antennas in metres.

This layout favours East-West orientations, but the relative elongation of the array in the East-West direction is only a factor of approximately two, as compared to the North-South direction.

4.4 Expected imaging performance

Simulations of a typical 8 h observation at a declination of − 60 ° were used to calculate the expected point spread function (PSF) of a single boresight beam (see Figure 4). In addition, Table 3 shows predicted survey speeds for BETA, based on the calculations presented in Johnston et al. (Reference Johnston2007).

Figure 4. Point spread function for a naturally-weighted synthesised beam covering 1.1–1.4 GHz assuming an 8 h observation at a declination of − 60 °. Contours represent a linear scale in 10% increments, with solid lines being positive and dashed lines negative.

Table 3. Predicted parameters for three different survey modes, split into two columns for the upper and lower parts of the BETA observing band, assuming nine simultaneous dual-polarisation beams. Sensitivity for the first two rows is expressed as the 1σ threshold over the stated bandwidth. The surface brightness survey is characterised by the rate required to reach a 1σ limit of 1 K over the specified bandwidth, assuming the stated spatial resolution which was obtained using a maximum baseline of 400 m and only five of the BETA antennas (antenna 15 excluded due to its relatively long baseline). The final row gives the time required to reach a 1σ limit of 1 mJy beam^{− 1} for a point source in the field centre.

5.1 Physical and electrical properties

Each BETA antenna is equipped with an ASKAP Mk I chequerboard phased array feed (Hay Reference Hay2010; Hay et al. Reference Hay, O’Sullivan, Kot and Granet2006; Hay & O’Sullivan Reference Hay and O’Sullivan2008; Hay, O’Sullivan, & Mittra Reference Hay, O’Sullivan and Mittra2011). These consist of 188 active elements (94 for each of two orthogonal polarisations) arranged in a regular grid (see Figure 5) that covers one square metre of the focal plane. Low noise active baluns (Shaw, Hay, & Ranga Reference Shaw, Hay and Ranga2012) connect between adjacent corners of the chequerboard elements to amplify the signals they receive. As second order intermodulation of strong signals is possible in subsequent amplifiers, selectable sub-octave band filters are included after the first stage. The amplified and filtered signals are then transported by coaxial cable to the antenna pedestal through cable wraps.

Figure 5. Diagram showing the layout of antenna elements on the surface of the ASKAP Mk I PAF. Low noise active baluns are connected at the points of adjacent diamonds (black). The solid aluminium ground plane is shown in outline and the foam spacing between the chequerboard and the ground plane is shaded grey.

Voltages, currents and temperatures in the PAF are monitored, but care is needed to minimise the radio frequency (RF) interference that such a monitoring system may produce. This is done by using an optical connection for monitor and control, and an implementation of the serial peripheral interface (SPI) communications protocol, which is only clocked by a master hub when data communications are required.

The feed is housed in a sealed enclosure designed to keep the internal environmental conditions as stable as possible. This enclosure is water-tight and kept at a slight positive pressure by a supply of dry air from the antenna pedestal. Because the unit is sealed, heat generated by the electronics cannot escape directly into the surrounding air. Instead, a water-cooled heat exchange system is used to keep the air within the PAF at a steady temperature of 25 °C. Control is achieved by varying the flow rate of cold water through the unit.

The feed was designed to cover a frequency range of 0.7–1.8 GHz, with a noise temperature of 50 K, but the prototype chequerboard suffers from an impedance mismatch between the antenna elements and the low-noise amplifiers that results in an elevated noise temperature in the upper part of the band (see Section 5.3).

Electromagnetic simulations are critical to the PAF design process and the design of the Mk I PAF has allowed improvements in the models used, such that the simulated performance is now a good match to the measured performance. Successive generations of chequerboard currently under development have been verified to meet the 50 K system temperature specification across the entire band (Shaw et al. Reference Shaw, Hay and Ranga2012). These new Mk II systems will be installed on all ASKAP antennas, eventually replacing the BETA Mk I PAF systems.

In order to fully characterise the performance of the PAF on the sky, it is necessary to form a beam that efficiently illuminates the antenna. Practicalities associated with this are discussed in the next section.

5.2 Beamforming and calibration methods

ASKAP belongs to a new class of PAF imaging array instruments where the wide field of view depends on the ability to electronically form many simultaneous, well-shaped, and stable beams on the sky. Achieving these goals poses (among other issues), two new engineering challenges not encountered in conventional synthesis imaging arrays: digital beamformer weight calculation and beamformer calibration. The latter is distinct from traditional imaging array calibration.

To keep patterns stable, beamformer weights must be recomputed regularly over intervals ranging from a few days to months in order to adjust for relative per-receptor gain and phase drift, variability in the electronics, and intentional changes in system parameters such as receiver chain gain settings. Much more frequent updates are required if maximum sensitivity is to be achieved for each observation by optimising with respect to the local direction dependent noise field, or for spatial nulling of moving or transient RFI (Landon et al. Reference Landon, Elmer, Jones, Stemmons, Jeffs, Warnick, Fisher and Norrod2010; Landon, Jeffs, & Warnick Reference Landon, Jeffs and Warnick2012).

To date, all weights used for real data processing on ASKAP and other published PAF results have been from the statistically optimal maximum signal-to-noise ratio algorithm (max SNR, also known as maximum sensitivity) (Applebaum Reference Applebaum1976)

\begin{eqnarray*} y_k [i] & = & {\bf w}_k^T {\bf x}[i] \nonumber \\ {\bf w}_k & = & \hat{\bf R}^{-1}_n \hat{\bf v}_k \nonumber \end{eqnarray*}

where x[i] is the vector of complex array voltages for a single frequency channel at time sample i, y_k [i] is the kth beam time series output, w _k is the corresponding vector of complex weights, $\hat{\bf R}_n$ is the integrated sample estimate for the local PAF array covariance matrix (ACM), and $\hat{\bf v}_k$ is the estimated array response vector for a unit magnitude plane wave signal arriving from the direction of the kth beam’s mainlobe peak. This approach is optimum with respect to a given noise covariance model $\hat{\bf R}_n$ , but beam pattern detail is not directly controlled. This can lead to pattern variation across beams and over time.

Since the spatial structures of spillover noise, sky noise, and other components can depend on elevation pointing and other variables, best results are obtained using contemporary $\hat{\bf R}_n$ estimates from a nearby off-source pointing at the same elevation. However, it can be challenging to obtain frequent, effectively source-free noise field estimates. So a single typical $\hat{\bf R}_n$ obtained during a major calibration cycle is often used for a period of several days in computing weights for all beams and dish pointing directions. The result is a compromise set of weights which do not truly yield maximum sensitivity for the current observation, but which produce beam patterns that vary little over time.

Other algorithms have been proposed which deterministically control the beam pattern. These can constrain fixed mainlobe or sidelobe shapes (e.g. Gaussian beam or equiripple sidelobes), improve pattern stability, or provide a parametric hybrid trade-off between statistically optimum and controlled shape beamforming (Elmer et al. Reference Elmer, Jeffs, Warnick, Fisher and Norrod2012; Willis Reference Willis2009). These advantages come at the cost of reduced sensitivity. Also, such methods require a denser, and thus more time consuming beamformer calibration grid observation, which is discussed next. Evaluation of the relative effectiveness of these candidate beam-forming algorithms, particularly as regards to final imaging quality metrics, is a major goal of the ASKAP BETA project.

Beamformer calibration is the process of estimating array response vectors $\hat{\bf v}_k$ for all beam mainlobe peak directions, and for all points where deterministic pattern constraints or response samples are desired. For example, constraints could be placed along each beam’s -3 dB contour to enforce symmetry and size, or in the near sidelobes to set a maximum allowed level.

Calibration requires a sufficiently compact and bright sky source to excite PAF receptors at high signal-to-noise ratio so the corresponding array response can be easily extracted from the total signal plus noise array covariance $\hat{\bf R}_{s+n}$ . Array signature $\hat{\bf v}_k$ is estimated as the dominant eigenvector of the generalised eigen equation for $\hat{\bf R}_{s+n}$ and $\hat{\bf R}_{n}$ (Landon et al. Reference Landon, Elmer, Jones, Stemmons, Jeffs, Warnick, Fisher and Norrod2010). The dish is steered in a grid pattern to place the source at each desired calibration angle relative to dish bore sight. A few seconds of integration (depending on source brightness) are collected at each point. For dense calibration, grids needed to map out full beam response patterns, or with many deterministic beam pattern constraints, this process can be time consuming, on the order of hours. However, currently for BETA max SNR weight calculations, only nine pointings are needed for the nine real-time beamformer main lobe peaks. These are recalibrated every few days to account for instrumental drift.

A significant challenge for ASKAP beamformer calibration is that in the southern sky there are few sufficiently bright sources to serve as reliable calibrators for the relatively small 12 m dishes. Two distinct calibration approaches have been used successfully at the Parkes ASKAP test platform to address this problem. First, we have exploited the large aperture of the co-located Parkes 64 m telescope to facilitate calibration with weaker sources in the range of a few tens of Janskys. For this interferometric calibration, the Parkes 64 m telescope tracks the source while the 12 m is steered through its grid pattern. The 64 m telescope’s L-band receiver signal is wired into spare ports of the PAF ACM correlator, so cross-correlations are computed with all PAF array elements. Due to the high gain of the 64 m telescope, the dominant eigenvector of this extended $\hat{\bf R}_{s+n}$ matrix yields the desired cal vector $\hat{\bf v}_k$ .

Second, since the Parkes 64 m telescope is not always available, and at the MRO there is no large dish, a stand-alone calibration method is needed for 12 m dishes. Because of its finite size and intrinsic variability, the Sun is not an obvious candidate for a beam-forming reference, but we have found that it is the only source available in the southern sky with sufficient radio flux for the maximum signal-to-noise method to work reliably on a single 12 m antenna and 1 MHz frequency coarse channel width.

To confirm suitability of the Sun as a calibrator, we included a model of its average intensity distribution over the disk and corona regions (Kuzmin Reference Kuzmin1966) in a detailed electromagnetic simulation of dish, supports, and array. The simulated PAF was the BYU/NRAO 19 element array, which though smaller has similar calibration performance for the central beams as the ASKAP BETA PAF. Receiver, sky, and spillover noise, along with element mutual coupling effects were carefully represented in the simulation which produced the observed sample array covariance matrices $\hat{\bf R}_{s+n}$ and true point-source array response vectors v _k as output products. Estimates $\hat{\bf v}_k$ were then computed as described above from the simulated $\hat{\bf R}_{s+n}$ , and the deterministic vector correlation coefficients $\rho = (\hat{\bf v}_k^T {\bf v}_k) / ( || \hat{\bf v}_k || \, || {\bf v}_k ||)$ between estimated and true calibration vectors were computed as a function of dish diameter (see Figure 6). A ρ value approaching 1.0 indicates no corruption of the calibration due to the extended non-point-like structure of the Sun. For the 12 m dish, ρ = 0.999 4, and no discernible distortion was found in beam patterns using the Sun calibration vectors for dishes as large as 30 m. This is because the 32.1 arcmin diameter Sun is unresolved in the beamwidth of these smaller dishes. Average intensity variation is not a problem since eigenvectors are arbitrarily normalised, and variation across disc is also unresolved in the natural dish beam.

Figure 6. Correlation coefficient, ρ between array response $\hat{\bf v}$ estimated with the Sun as calibrator, and the true point-source response.

Optimising the calibration processes and developing methods for extending interferometric calibration to use several co-phased 12 m dishes as the high gain reference will be major goals of the BETA project.

5.3 Beamformed system temperature

The PAFs used on BETA are the first of their kind and the first ever to be used as part of a synthesis telescope. They are not the final design that will be used for ASKAP, so the measured performance shown here does not reflect the final system configuration. The BETA platform is an engineering prototype and the lessons learnt during its design and construction have already fed back into the design of a Mk II PAF that meets the system temperature requirements for ASKAP across the entire observing band. Here, we describe the as-built measured parameters of the BETA system, knowing that it exhibits a higher than intended system temperature in the upper part of the band.

Aperture array measurements (where the PAF is operated on the ground, pointing upwards at the cold sky or a warm absorber load held overhead) were used to measure the Y-factor of the PAF, giving its system temperature. On-dish measurements (Figure 7) give the combined T_sys/η (a metric suitable for comparing feed systems independently of collecting area) which by comparison to aperture array measurements suggests that the efficiency factor of a formed beam is approximately 0.7. A boresight beam was formed on the Sun and then used to observe either Taurus A or Virgo A for flux calibration.

Figure 7. Plot of the measured beamformed system temperature of the prototype chequerboard PAF as a function of frequency, with 1 MHz resolution. Data were recorded using antenna six at the MRO. Uncertainty estimates represent 1σ limits over four independent trials. Square and diamond markers represent two orthogonal polarisations, recorded simultaneously. A Mk II chequerboard PAF is currently under construction and has already demonstrated significantly better performance.

These measurements use max SNR beamforming on a single dish, performed independently on each 1 MHz channel, using the Sun as a reference source. The uncertainty in the measurement over several trials suggests that the efficiency of a beam can vary slightly from one beam-forming attempt to the next. Given that the process of maximum S/N beamforming relies on experimentally-determined information and imposes no constraint on the shape of the beam, this is not surprising. Single-dish beam-forming methods are expected to perform poorly compared to methods that use correlations between the antennas, which we are currently developing. The subject of beam optimisation is too large to be covered here and will be the topic of future publications. It will be possible to trade sensitivity for beam symmetry, frequency independence or other desirable properties. Individual science goals will likely require different beam-forming optimisations.

Figure 7 also shows several channels with unusually large scatter. The worst of these fall between 1540 and 1560 MHz, corresponding to the service downlink band used by Thuraya 3, a geosynchronous communications satellite that provides mobile telecommunications to the Asia-Pacific region. It is likely that the beam-forming process preferentially steers towards this artificial source when its signal is present in the field of view. We are actively researching methods (such as subspace projection) that allow a spatial null to be placed on known sources of interference when beamforming. This method of interference mitigation is a unique capability of phased array systems.

6.1 Analogue stage

Within the PAF, the signal from each active balun passes through a band-pass filter that selects frequencies in the range 700–1800 MHz. The band-limited signal is further amplified and then passed through a switchable band-select filter. Two overlapping ranges are available, 700–1300 MHz and 1000–1800 MHz. The output from each filter is amplified again and transmitted at the sky frequency via coaxial cable, through a series of mechanical twister mechanisms that ensure safe passage through the rotating structure of the antenna, to the input side of the shielded cabinet in the pedestal.

In the analogue cabinet all the RF signals undergo a dual-heterodyne down conversion to an IF band of 422–726 MHz. This is suitable for the second Nyquist zone of the analogue to digital converter, which is clocked at 768 MS/s. IF power levels are monitored, as well as voltage, current, and temperature for key components. The IF signals pass through variable attenuators with 1 dB steps and 15 dB range, which are used to ensure an appropriate input level for the ADCs.

The local oscillators used for down conversion are derived from a 16 MHz reference signal supplied from the central building. Presently, the reference signal is derived from a Rubidium clock, which can be locked to a neighbouring hydrogen maser for additional phase stability. On the receiving end, integer-N phase locked loops (PLLs) with purely integer frequency multiplication factors are used in each pedestal to keep the local oscillators in phase with the reference signal. This ensures that phase noise is uncorrelated between each individual antenna. As the PLL is limited to integer multiples of 16 MHz, setting of the band centre frequency is in 16 MHz steps. Limiting the PLL to integer multiples of 16 MHz also ensures all antennas have the same phase relation after a restart.

6.2 Digital data processing

The signals pass from the analogue cabinet into the digital cabinet, which holds four racks of CSIRO-designed DragonFly-2 circuit boards (Bunton et al. Reference Bunton, Hampson, Brown, Pathikulangara, Tuthill, Souza, Joseph, Bateman and Neuhold2011). Each DragonFly-2 board has four ADCs, a field programmable gate array (FPGA) for signal processing and interfacing and four 10 Gb s^{− 1} single mode optical links to transport data to the central building. The sampler clock is obtained from the LO subsystem and in addition, the 16 MHz reference and a precise measure of coordinated universal time (UTC) encoded as a digital signal (referred to as binary atomic time, BAT) are supplied to each DragonFly-2. This allows a unique time to be assigned to any sample from the ADC. The signals from the digitisers then pass through an oversampled polyphase filterbank (Bunton & Navarro Reference Bunton and Navarro2004) with a frequency channel spacing of 1 MHz and an oversampling ratio of 32/27. The desired bandwidth for each channel is 1 MHz and oversampling is used to reduce aliasing (Schafer & Rabiner Reference Schafer and Rabiner1973; Crochiere Reference Crochiere1980; Wilbur, Davidson, & Reilly Reference Wilbur, Davidson and Reilly2004; Tuthill et al. Reference Tuthill, Hamspon, Bunton, Brown, Neuhold, Bateman, de Souza and Joseph2012). The 1 MHz channel width was chosen to ensure that a narrow-band (weight-and-sum) beamformer (Van Veen & Buckley Reference Van Veen and Buckley1988) can be used without smearing the beams.

The digitiser continuously calculates sampling statistics (the probability distribution function of the output) and makes this information available to the central control system. In the output of the polyphase filterbank, the power levels in each channel are also monitored. Attenuators in the analogue conversion system provide a 15 dB programmable range that is automatically set to allow for on-line scaling of the ADC input. This is necessary to balance the conflicting requirements of maintaining adequate signal-to-noise ratio in the quantised astronomical signals while reducing the possibility of ADC saturation due to strong interference.

Four inputs to the down conversion sub-system (in addition to the 188 inputs from the PAF) are reserved for PAF calibration signals and future developments; one possible example being a reference antenna for active interference mitigation. The PAF calibration signal consists of low-power broad-band noise radiated from the surface of the dish. A copy of the calibration signal is sent directly to the receiver and correlated with the PAF outputs, allowing real-time monitoring of the relative gain and phase for each element. This information can in turn be used to adjust beam weights for maximum pattern stability (Hayman et al. Reference Hayman, Chippendale, Qiao, Bunton, Beresford, Roberts and Axtens2010). The power of this noise source is kept to an absolute minimum to reduce the possibility of it correlating between adjacent antennas. It is also possible to remotely control the noise source so that brief measurements may be taken over longer intervals of time, further reducing the impact on the astronomical signal.

7.1 Digital beamformer

The beamformer system operates on each 1 MHz channel, producing nine dual-polarised beams on the sky for each frequency channel. Future expansion to 36 dual-polarised beams is planned for ASKAP. Since there are 304 separate 1 MHz channels, there are also 304 instances of the beamformer computational unit which are physically distributed across 64 FPGAs on the Redback-2 boards. In the 9-beam system, all 188 PAF ports are used to form the beam. This reduces to an average of 63 in the 36-beam implementation. The data on the links from the digital receivers at the antenna arrive ordered with one quarter of the frequency channels for four ports, whereas each beamformer unit requires all ports for a single frequency channel. The large cross-connect needed to supply data to the beamformer units is realised by the combined routing capabilities of a cross-point switch on each Redback-2, the ATCA backplane full-mesh, interconnections between the Redback-2 FPGAs and within the FPGA itself.

The beamformer operation reduces data throughput and the processing load of the correlator is also significantly decreased because only corresponding beams must be correlated, instead of all PAF elements. The capacity of the beamformer is limited by the number of clock cycles between new data values and the number of beamformer multipliers. Within these constraints the beamformer needs to flexibly allocate the number of complex-multiply-accumulate (CMAC) operations to the beams. This number can vary from beam to beam.

The beam-forming operation is followed by a fine filterbank that further channelises the signal to the final required frequency resolution. The fine filterbank is a critically-sampled polyphase filter which channelises each 1 MHz band into 54 fine frequency channels that are 18.52 kHz wide. Oversampling in the coarse filterbank results in data outside the 1 MHz band, which are discarded. The fine filterbank is operated in block mode to minimise FPGA memory usage. Each 1 MHz beamformer unit produces nine sets of beam pairs at a time for one sample of the 1 MHz coarse filtered data from all PAF ports. The fine filterbank, however, requires blocks of consecutive samples for one single dual-polarisation beam. A data ‘corner-turn’ is therefore necessary between the two modules so that contiguous blocks of samples for a given dual-polarisation beam are presented to the filterbank inputs. Since the corner-turn requires significant storage resources, external DDR3 SDRAM (synchronous dynamic random access memory) is used. Associated with the fine filterbank is fine delay tracking (integral-sample delays are removed using a buffer in the DragonFly-2), and fringe stopping.

7.2 PAF ACM correlator

The beamformer also generates data needed to calibrate the phased array feed. This consists of correlations between calibration signals and PAF port signals and the ACM for the port signals. The compute load of the full ACM is much higher than the beamforming itself and approximately equal to that of the full correlator that follows the beamformers. For this reason, the ACM module operates with a 25% duty cycle on a subset of the 1 MHz channels. All calibration data are sent to the TOS via user datagram protocol (UDP) packets. The TOS then runs the algorithms to determine the weighting coefficients needed for each beam, and uploads these to the hardware as necessary. A summary of the key signal processing steps performed in the beamformer FPGAs is shown in Figure 10.

Figure 10. Key signal processing tasks performed by the BETA beamformer FPGAs.

It is also possible to access critically-sampled data from several beams at the output of the beamformers via UDP over 1 Gb ethernet, but only at about 10% of the full duty cycle and 1/4 the bandwidth. This mode of operation allows for software-based correlation as an alternative to the main data path, which has been useful for initial imaging, system debugging, delay searching and low-level verification of FPGA signal processing firmware.

7.3 Central correlator

The links between the beamformer and correlator are also 10 Gb s^{− 1} optical connections. As the payload is different to that from the digitiser, this stage uses a different but related communications protocol to the pedestal-central building link. The beamformer data that arrives at the Redback-2 correlator boards is redistributed amongst the processing FPGAs so that each FPGA processes a number of fine frequency channels with data from a single dual polarisation beam at one time.

At any one time the fine filterbank in the beamformer processes up to 64 coarse 1 MHz channels for a single dual polarisation beam and each coarse channel generates 54 fine channels. The correlator processes all these data as they are generated. As the input to the correlator switches between beams and coarse frequency channels, intermediate correlation values are accumulated in external SDRAM. Data in the SDRAM are triple buffered with the buffers switching every 5 s. A summary of the signal path is shown in Figure 11.

Figure 11. Diagram of operations performed in a single correlator cell.

After accumulation, data are dumped to the external correlator control computer via the 1 Gb ethernet connection available on the Redback-2 boards. A network switch is used to aggregate all data onto 10 Gb ethernet before being ingested by the control computer. This computer formats the data into a form suitable for transmission to the remote supercomputing facilities in Perth.

7.4 Data rate

The total data output of the BETA correlator is summarised as follows. For the cross-correlations we have six antennas producing 6 × 5/2 = 15 individual baselines. Each baseline produces nine dual-polarisation beams and each beam produces 304 × 54 = 16 416 fine frequency channels with four stokes parameters per channel. This is a total of 15 × 9 × 16 416 × 4 = 8 864 640 complex numbers.

For the auto-correlations we have six individual antennas with nine dual-polarisation beams per antenna. Each of these produces 16 416 fine frequency channels with two auto-correlations for the orthogonal polarisations and one cross-polarisation product. This is a total of 6 × 9 × 16 416 × (2 + 1) = 2 659 392 complex numbers.

In addition to the products above, a 64-bit integration count is recorded for each beam and fine frequency channel every cycle, so the data can be normalised by the number of samples integrated. This adds another 147 744 64-bit numbers every 5 s.

With an integration cycle of 5 s and a long-term accumulator word-length of 32 bits, the total data output rate from the correlator is (8 864 640 + 2 659 392 + 147 744) × (32 × 2) bits every 5 s, or 142.5 Mb s^{− 1} (using 1024² to convert to Mb).

The ingest pipeline combines these values with appropriate meta-data to define a CASA measurement set. The addition of meta-data causes a small increase to the overall data rate, resulting in a final figure of 153.9 Mb s^{− 1} reaching the disk. This means that a 1 h observation with 5 s integrations and 18.5 kHz frequency resolution requires 68 GB of storage space. Consequently, a 12 h observation produces 816 GB or 0.8 TB of data.

9.1 Monitoring and control

The architecture of the TOS is discussed in detail elsewhere (Guzman & Humphreys Reference Guzman and Humphreys2010), so only a brief outline will be given here. The software is built on a set of tools and libraries collectively known as EPICSFootnote ¹⁰ . It is designed specifically to support remote operations and large area surveys through a model of automated service observing.

A high-level system known as the executive is responsible for the execution of scheduling blocks, effectively orchestrating the various hardware and software systems during data acquisition. The implementation of scheduling blocks provides a mechanism for declaring complex observations via the combination of scheduling block parameters and a script, known as the observation procedure. The observation procedure is implemented using Python and exposes telescope control and monitoring interfaces via an application programming interface (API) known as the observation procedure library (OPL).

Expert users, or those performing novel observations, can exploit the flexibility provided via the procedure scripting capability. However, common observing modes are provided via scheduling block templates, providing a pre-tested procedure (script) and requiring the user only specify basic parameters.

9.2 Image processing and calibration

The computational challenges associated with interferometry over a wide field of view are known to be significant (Cornwell, Voronkov, & Humphreys Reference Cornwell, Voronkov and Humphreys2012). Since BETA utilises only one sixth of the total number of ASKAP antennas (and therefore has many fewer and shorter baselines), and one quarter of the number of beams, it will provide a tractable benchmark on the path to a full ASKAP imaging pipeline, at the expense of sensitivity and a more complicated PSF (see Figure 4).

Calibration, imaging, and other science processing is carried out on the ASKAP central processor. The hardware platform is a Cray XC30 supercomputer consisting of 472 compute nodes. Each node contains 64 gigabytes of main memory and two Intel Xeon E5-2690 v2 CPUs, each with ten processing cores. All compute nodes have access to a global Lustre filesystem with a capacity of 1.4 PB and capable of sustained I/O rates of 25 Gb s^{− 1}. The Lustre file system provides a buffer for datasets not yet processed as well as for images and other data products that are yet to be archived.

As existing data reduction packages were neither capable of dealing with the high data-rates nor the wide-field imaging challenges of a PAF-equipped interferometer, a custom data reduction package has been developed. Visibilities are gridded with AW-projection and weighted using a single-pass Wiener filter. Continuum images are deconvolved using the multi-scale, multi-frequency-synthesis algorithm described in Rau & Cornwell (Reference Rau and Cornwell2011). For spectral line cubes, multi-scale CLEAN (Cornwell Reference Cornwell2008) is used instead.

For BETA, a more traditional calibration approach is used, where individual beams are independently calibrated by observing a single reference source. While time-consuming, this method is well-tested on other telescopes and allows for bootstrapping while the sensitivity of the array is low due to the small number of available antennas.

9.3 Data archiving

Long term archival storage is provided by a pair of tape libraries (supplied by Spectra Logic) in the Pawsey centre. Two copies of all data products are stored, one copy in each library, to minimise the possibility of data loss. Initially 10 PB of tape storage (physically 20 PB due to mirroring) is allocated to the ASKAP telescope, with future expansion possible as the libraries are each able to support up to 50 PB with the addition of tape media.

This storage pool is used by two different ASKAP archives. An internal archive, known as the Commissioning Archive, provides storage for any file-based data product the engineering and commissioning teams may wish to store. A public facing archive, the CSIRO ASKAP Science Data Archive (CASDA), is currently being developed. This archive will provide access to data products via a web interface as well as via virtual observatory protocols as defined by the IVOA.

The data rate at the output of the BETA correlator is sufficiently small (compared to the full ASKAP) that it is possible to permanently archive all visibilities. This information is stored in the form of CASA measurement sets. There is sufficient storage space at the MRO to buffer approximately one week of typical data, in order to cope with any network outages.

9.4 Science processing

The focus during the operational life of BETA will be on testing beamforming and calibration techniques, the imaging pipeline and source-finding algorithms (Whiting & Humphreys Reference Whiting and Humphreys2012) with real-sky data. Due to the large dataset sizes, the full ASKAP science processing system has the difficult task of performing calibration and imaging in a semi-autonomous mode. This challenge can only be met by gaining an in depth understanding of the telescope and its behaviour under a variety of conditions. Experience gained with BETA will be critical input into this process.

Users of the BETA telescope have access to the Pawsey central processor facility, which can be used to perform data reduction in a traditional iterative and somewhat interactive fashion. As experience is built, automated processing pipelines will be deployed as a demonstration of the ASKAP quasi-realtime processing model. In addition to data reduction with the ASKAP-specific toolset, subsets of data are relatively portable and existing CASA tools are often sufficient for processing tasks. Full continuum datasets (with 1 MHz channel resolution) are transportable outside of the Pawsey centre, as are small subsets of the spectral line datasets. This allows cross-verification of processing outcomes with a variety of existing systems and software packages.

10.1 First observations

Preliminary observations using a single pixel feed to characterise the first assembled antenna have already led to interesting scientific outcomes, including the very long baseline imaging project described in Tzioumis et al. (Reference Tzioumis2010).

The first synthesis image with multiple beams obtained using data from all six BETA antennas (in 2 × 3 sub-array mode) is show in Figure 12. The field of interest was selected because it contains three sources (PKS 1547-795, PKS 1549-790 and PKS 1610-771), each with a flux density of roughly 5 Jy, two of which fall within a single primary beam and another that is approximately three beam widths away. PKS 1934-638 was observed with each beam to calibrate the phases of the array. The targets were all detected (along with several other sources in the field) and their positions and fluxes match well with data from the SUMSS catalog (Bock et al. Reference Bock, Large and Sadler1999).

Figure 12. Stokes I image formed using nine simultaneous overlapping beams and six baselines. 304 MHz of bandwidth (712–1016 MHz) was recorded using the BETA correlator with a channel width of 1 MHz.

Further observations will allow updates to key parameters (such as antenna positions), which will iteratively improve image quality over time. Early results from the full 15-baseline BETA (in preparation) show that traditional data processing methods can already yield high-quality images, with calibration errors being the main limiting factor. There are several reasons for this, including the presence of confusing sources near our chosen calibrator, and the fact that the Sun can be detected on the shortest baseline regardless of its position in the sky with respect to the target field. These issues are side effects of the wide field of view and will require refinements to existing data processing strategies to fully overcome. Experience gained from BETA observations will feed into the development of the final ASKAP processing pipeline.