2.1 Data capture

The standard MWA signal path, as described in detail in Tingay et al. (Reference Tingay2013a), Prabu et al. (Reference Prabu2014), and Ord et al. (Reference Ord2014), can be simplified to the following synopsis. Dual-polarisation dipole antennas arranged as 4 × 4 grids (tiles) are analogue beamformed, with the radio frequency signals sent to receiver boxes in the field. Each of the 256 signals (128 tiles × 2 polarisations across 16 receivers) is subsequently amplified, digitised, and processed through a coarse polyphase filter bank (PFB) in the receiver enclosures. The outputs from each receiver are then sent to one of four fine PFBs (1/4 of the tiles to each). The 32 fine PFB outputs are then converted from RocketIO (a Xilinx serial protocol) to transmission control protocol by 16 media converter servers (CISCO UCS C240) using peripheral component interconnect express mounted cards produced by Engineering Design Team Incorporated (EDT Cards). These 32 signals are then reorganised into coarse channel groups and sent to the 24 servers hosting the graphics processing unit (GPU) cards that perform the cross-multiply step of correlation. After correlation, the output is transferred offsite and archived.

The initial digital system design did not include any access to the voltage data along this path. We decided to implement voltage capture on the media conversion servers since they otherwise have a relatively small workload and had not been purchased or even specified at the time of this subsystem’s design. The media converter servers were subsequently specified to include 128 GB of RAM each, for potential buffering purposes, 24 × 2.5′′ HDD slots, and a dual-channel drive controller. The digital signal path, receivers through GPU servers, are schematically represented in Figure 1 to highlight where the voltages are recorded.

Figure 1. Simplified illustration of the MWA digital signal path. For the VCS mode, the baseband data between the fine channel polyphase filter bank (PFB) and the X-engine of the correlator are written to local RAIDs on the 16 media converter servers (highlighted with a dashed box). This gives us 100 μs resolution and frequency channels 10 kHz wide.

Each server is equipped with two redundant arrays of independent disks (RAIDs) for recording voltage data. These are each comprised of six 2.5′′ 300 GB (279 GB actual) 10K SAS drives and are combined as a hardware controlled RAID 0 (block level striping but without parity or mirroring), giving each RAID 1.44 TB of usable storage with no redundancy. The RAIDs are on independent 6 GB s^{− 1} channels in the controller, to maximise data throughput.

The 4-bit + 4-bit complex voltages from the fine PFB stream through the media conversion servers at a rate of ~ 4.2 GB s^{− 1}. After being converted by the EDT cards, these data can be streamed directly to the RAIDs. Each server has two PFB lanes passing through it, each of which is directed to a separate RAID in the unaltered post-PFB format as files on 1 s boundaries. Therefore, every second of VCS observation generates 32 × 242 MB files across 16 machines (where each file contains 1/8 of the fine channels for 1/4 of the tiles) for a combined rate of 7.744 GB s^{− 1}.

It is important to highlight that the ‘original’ digitised tile voltages are not recorded, instead it is the channelised output of the four PFBs (i.e. the data have been aggregated, filtered, amplified, and processed through two PFBs before being recorded).

2.2 Capabilities

VCS Observations consist of 30.72 MHz of bandwidth, in 1.28 MHz wide subbands, spread as desired between 80 and 300 MHz. These data have 10 kHz channels and 100 μs time resolution (corresponding to critically sampled 10 kHz channels at the Nyquist Shannon rate). These data are the same as the standard correlator input. At most, the system can record a single uninterrupted time span of ~ 100 min. Dedicated hardware, in the form of a server with 16 × 3.5′′ SAS-drive bays, is installed in Perth where the recorded data can be transferred onto a software RAID utilising a 10 GB s^{− 1} link to the MRO. Significant time, however, is required between subsequent VCS observations to transport the data from the site since archiving correlated observations takes precedence.

Recording the data in this mode gives the user maximum flexibility in processing the data, in particular how the 256 data streams (polarisations × tiles) are combined. The simplest merger of these data is the incoherent sum, where the voltages from each tile are multiplied by their complex conjugate, to form the power, and subsequently summed. This preserves the field of view (FoV) while increasing the sensitivity over a single tile by a factor of $\sqrt{N}$, where N is the number of tiles summed. All the data presented in this paper were combined in this fashion unless noted otherwise.

Alternately, more than an order of magnitude increase in sensitivity can be gained if a phase rotation is applied to each voltage stream before summing the voltages to form a coherent beam at the cost of a drastic reduction in the FoV for an individual phase centred beam. To attain the same sky coverage as the incoherent sum, thousands of coherent beams need to be processed. Since these data are recorded from the ‘standard’ data path, it is also possible to cross-correlate these data offline in a manner similar to normal but with control over the temporal and spectral integration (within the constraints of the raw data and the available compute resources).

3.1 Pulsar science

Arguably the primary science application of the voltage capture mode is for pulsar observations. Following their serendipitous discovery at 81.5 MHz by Hewish et al. (Reference Hewish, Bell, Pilkington, Scott and Collins1968), much of the early research on pulsars was at low frequencies (Taylor & Manchester Reference Taylor and Manchester1977), however the eventual quest to find more pulsars in the galactic plane, which is highly sky-noise dominated at low frequencies, and also to achieve higher precision in timing pushed the observations to higher frequencies (≳ 1 GHz). With the advent of multiple new low-frequency arrays (LOFAR) including MWA, the LOFAR (van Haarlem et al. Reference van Haarlem2013), and the long wavelength array (Taylor et al. Reference Cordes and Lazio2012), a renaissance in low-frequency pulsar astrophysics is on the horizon. In fact, a number of low-frequency pulsar results have already started being published from these and other instruments (e.g. Bhat et al. Reference Bhat2014, Archibald et al. Reference Archibald, Kondratiev, Hessels and Stinebring2014, Dowell et al. Reference Dowell2013, Stovall et al. Reference Stovall2014).

While state-of-the-art pulsar backends are capable of providing far superior time resolution via phase-coherent dedispersion over large bandwidths (e.g. van Straten & Bailes Reference van Straten and Bailes2011), the VCS functionality of the MWA is well matched to a wide range of pulsar science goals at low frequencies, particularly for long-period pulsars (spin period, P ≳ 100 ms) with dispersion measures (DMs) ≲ 200 pc cm^{− 3}, where scattering is not large enough to significantly smear the emission across the pulse period. This parameter space samples the vast majority of the local (≲ 1 kpc) pulsar population. Our 10 kHz channelisation means the effective achievable time resolution will largely be limited by the dispersive smearing within the channel, which is ~ 1 ms at DM = 100 pc cm^{− 3} at a frequency of 200 MHz. In targeted pulsar observations, this limitation is greatly reduced by coherent dedispersion. The 100 μs resolution presents an obvious challenge in observing millisecond pulsars (MSPs). It is, however, enough resolution to construct high-quality pulse profiles of pulsars with periods ⩾ that of PSRJ0437-4715 (see Bhat et al. Reference Bhat2014 and Figure 2). It is, in principle, possible to reconstruct the time resolution of the full bandwidth (~32.6 ns) by inverting the second PFB stage, although this is yet to be attempted.

Figure 2. Pulse profiles of four different pulsars observed with the VCS mode covering a range of DMs and periods (see Table 1 for values) to demonstrate the flexibility of the instrument. There are 64 bins, spanning the pulse period for each pulsar.

Our successful detection of PSR J0437 − 4715, a binary MSP with P = 5.75 ms and DM = 2.65 pc cm^{− 3}, also provides an excellent demonstration of the timing stability of our recording system. Multiple observations of this pulsar and other objects such as the Crab pulsar have been made over time durations of 1 h without encountering any recording glitches. A close inspection of commissioning data (Table 1) confirms that, barring some issues such as power level modulations caused by the Orbcomm satellites within the FoV (see Section 4) and abrupt changes in power levels that result from changes in the beam-former settings, the data are in general highly stable. While we are unable to accurately determine the exact instrumental offset in our timing owing to our limited data, we have verified our time stamping accuracy, to first order, by successfully combining multiple observations of PSRs J0630 − 2834 and J0534+2200 that span time intervals in the range of ~ 30 min to 20 days (Table 1). The pulse profiles from different observations align at a level suggesting that the offset is limited to ≲ 0.5 ms. We aim to further characterise this more accurately as more observations accrue over the course of time.

Table 1. Pulsars detected after processing MWA voltage data. The voltage streams were combined into incoherent beams and then processed using a PRESTO pipeline. These observations were performed at various times within the commissioning period of the VCS, hence the variety of bandwidths. PSRs J0534+2200 (the Crab) and J0528+2200 were detected within the same beam during an observation, similarly PSRs J0630-2834 and J0742-2822 within a single 31 min observation.

^aThese observations were done with only 32 tiles.

^bThe bandwidth spanned 15.36 MHz, but one of the central course channels was not recorded so only 14.08 MHz of data were summed in the dedispersed time series.

^cAll values are taken from the ATNF Pulsar catalogue http://www.atnf.csiro.au/people/pulsar/psrcat/ (Manchester et al. Reference Manchester, Hobbs, Teoh and Hobbs2005)

Pulse broadening resulting from multipath scattering in the interstellar medium (ISM) is an important consideration at the MWA’s frequencies. Based on the galactic electron density models (NE2001; Cordes & Lazio Reference Cordes and Lazio2002) and an observationally-established scaling relation between scattering, DM, and the observing frequency (Bhat et al. Reference Bhat, Cordes, Camilo, Nice and Lorimer2004), a scatter broadening $\tau _{\text{scatt}}$ ~ 1 ms is expected near the low end of the MWA band (~100 MHz) toward pulsars at distances ~ 1 kpc located within the galactic plane (b = 0°). At larger distances within the plane, scattering can be substantial, owing to the non-linear scaling of scattering with DM, with $\tau _{\text{scatt}}$ ~ 100 ms to be expected toward the galactic centre at a distance of ~ 3 kpc at ~ 100 MHz. Off the galactic plane, scattering will likely be far less severe, and is expected to be ≲ 0.3 ms at frequencies ≳ 200 MHz, thereby retaining sensitivity to the detection of most nearby pulsars, including MSPs.

As seen from Table 1 and Figure 2, our commissioning analysis so far investigated DMs up to 123 pc cm^{− 3}, and little scattering is evident at DMs ≲ 50 pc cm^{− 3} at ~ 200 MHz. Besides the Crab, known for its atypical scattering (Ellingson et al. Reference Ellingson2013; Bhat et al. Reference Bhat2007), PSR J0742-2822 is the only pulsar that shows scattering in our data (Figure 2). While generally considered to be a hindrance for most pulsar studies including searches, scattering measurements provide useful means of characterising the ISM turbulence, for which the MWA frequency band is well suited.

Even with the 100 μs, 10 kHz limitations of the VCS, useful scintillation and profile studies are possible even for MSPs, as vividly demonstrated in Bhat et al. (Reference Bhat2014). The 100-μs resolution, while not a limitation for scintillation studies where the primary goal is to investigate the time and frequency modulations of the integrated pulse emission, will likely pose a limitation in the profile studies of MSPs with P ~ a few ms. Similarly, the 10 kHz resolution may limit the scope of scintillation studies to DMs ≲ 40 pc cm^{− 3} in the MWA band. However, cyclic spectroscopy could be implemented to characterise scattering at higher DMs (Demorest Reference Demorest2011). Low-frequency observations can also potentially yield accurate DM measurements, provided the frequency-dependent evolution of the pulse profile is modelled in the analysis; this is important for timing-array applications such as the search for gravitational waves (e.g. Manchester et al. Reference Manchester2013).

3.2 Single pulse astrophysics

The last ten years has seen renewed interest in high time resolution single pulse detections, with particular enthusiasm for both rotating radio transients (RRATs; McLaughlin et al. Reference McLaughlin2006, Burke-Spolaor et al. Reference Burke-Spolaor2011, Keane et al. Reference Keane, Kramer, Lyne, Stappers and McLaughlin2011) and FRBs (Lorimer et al. Reference Lorimer, Bailes, McLaughlin, Narkevic and Crawford2007, Thornton et al. Reference Thornton2013, Spitler et al. Reference Spitler2014). Both the large FoV (hundreds or even thousands of square degrees) as well as the flexibility in observing frequency make low-frequency arrays, such as the MWA, powerful tools in the search for single pulse emission. As mentioned in Section 2.2, an incoherent sum of the elements preserves this large FoV and generates a data set that is practical to search through using modest compute resources. Once detected, the power of the interferometric array can be used to localise and study these pulses as long as the voltages have been preserved.

We have verified the ability of the MWA to observe short-duration single pulse emission using two separate observations of the Crab pulsar (PSR J0534+2200, Table 1) to observe giant pulses (Hankins et al. Reference Hankins, Kern, Weatherall and Eilek2003) which, based on higher frequency observations (~5 GHz), are thought to be intrinsically ~nanosecond duration bright pulses. In the MWA band, these intrinsically short pulses are scattered to widths of tens of milliseconds (see Figure 3). We detected 51 and 47 pulses above 6 σFootnote ¹ from the observations. The brightest pulse we detected was 39 σ¹ (Figure 3), and shows a well-defined sharp peak followed by an exponential scattering tail.

Figure 3. Dedispersed (56.76 pc cm^{− 3}) total power from one of the detected Crab giant pulses observed with the VCS at 192.64 MHz. The long scattering tail (~40 ms) extends further than a single pulse period (33 ms). The median power from each fine channel was removed before dedispersion and time steps are averaged to 400 μs. Reference times on the abscissa denote seconds from the beginning of the observation.

For further information on FRBs see Trott, Tingay, & Wayth (Reference Trott, Tingay and Wayth2013) for the prospects of detecting FRBs with the MWA over a variety of potential spectral indices and Tremblay et al. (Reference Tremblay2015, in preparation) which describes the FRB pipeline being used by the MWA.

3.3 Solar science

The standard imaging modes of the MWA are being used to study the Sun (Oberoi et al. Reference Oberoi2014) with a variety of aims, including the investigation of the quiet Sun, the detailed study of Type II and III bursts, and the characterisation of coronal mass ejections (described in Bowman et al. Reference Bowman2013, and references therein). However, the temporal resolution of the standard MWA imaging modes can significantly under-sample solar radio emission variability. Intense solar bursts evolve rapidly in time and frequency, but previous MWA observations show that even in a quiet state, low level radio emission from the Sun evolves strongly in both time and frequency (Oberoi et al. Reference Oberoi2011). As such, there is a need for significantly higher time resolution observation modes with the MWA. Observations of the Sun for which voltages are captured can be used for high time resolution beamforming, either incoherent or coherent as mentioned in Section 2.2, to form high time resolution dynamic spectra (for example, as seen in Figure 4). Furthermore, the voltages can be correlated at high time resolution post-observation, in order to undertake high time resolution imaging of the Sun. Both high cadence beamforming and imaging modes are identical to those to be used for pulsars and searches for FRBs (Sections 3.1 and 3.2).

Figure 4. Dynamic spectrum spanning 5 s where the MWA’s 128 tiles have been incoherently summed from observations on MJD 56 544. Here, 12 coarse channels (half of the MWA’s typical bandwidth) were averaged to 10 ms increments and the median value of each 10 kHz fine channel was subsequently subtracted to highlight the variable emission. The vertical bands are the result of coarse channel edges where sensitivity is reduced. The narrowband, sub-second solar features highlighted here would be smoothed out with the typical integration times used in standard imaging modes. Times on the ordinate reference Unix Time.

In addition to Solar studies, the VCS is anticipated to conduct observations of interplanetary and ionospheric scintillations. High time resolution observations of satellite beacons will also be used to study the variability of ionospheric Faraday rotation.