• No results found

Software can always be improved, and SERPent is no exception. One of the weaknesses of SERPent, highlighted in Section 2.4.2, is the computational performance.

Parallelisation has made massive improvements in the performance. However, the adopted method of parallelisation is not the most efficient. The Python fork process parallelisation, pre-allocates jobs to CPUs before reducing them. This results in a plateau in performance at higher NCPUs, because of the ‘heavy’ jobs issue (Figure 2.11). An alternative exists in Python 2.7, which contains a multiprocessing module. A recent study on Python parallelisation approaches was conducted by Singh et al. (2013), who showed that the Python Process/ Queue method from the multiprocessing module performed better than other methods from the same module and the Parallel Python module. The Process/ Queue method creates a queue of jobs from which CPUs can access and take as and when they’ve finished processing a previous job which is added to a second queue containing finished jobs. Once all jobs are finished the second queue can be read for post-

processing. This will create an even performance increase by removing the performance bias from CPUs with a heavy jobs.

Another weak link in SERPent is the behaviour of theAIPStask REFLG, which has not only been too aggressive at times, but has recently (November 2013) been seen to create more flag entries than the input table (reported by a number of people; Anita Richards; Rob Beswick; Megan Argo; Danielle Fenech etc.). This is in complete contradiction to one of its primary functions which is to condense the FG row entries. Investigations are ongoing to understand the reason(s) for this behaviour. Despite this,AIPShas an internal limit to how many FG rows can be read in (10 million), which is not always sufficient. A passage which does this job outside of AIPS(i.e. within Python/ Parseltongue) has been considered as a possible solution.

Following further discussion at the e-MERLIN early science meeting, the L-band RFI may be time dependent. Low declination observations (such as COBRaS ∼ 40◦) will suffer from heightened levels of RFI from horizon sources and another big contributor to RFI has been identified as Manchester airport, just North of Jodrell Bank. Investigations will start in the future to determine and indentify different sources of RFI at L-band, which is the first step in order to mitigate the level of RFI during any particular observing run. Empirical studies are currently underway at UCL and the University of Manchester, UK, to find the optimal flagging parameters for SERPent, for each antenna or baseline. This is an important investigation which will affect both the flagging and computational performance of SERPent, as some antennas are more stable than others within the array and differ in the amount of RFI in the local environment.

An e-MERLIN L-band RFI mask has also been created (Megan Argo, private com- munication), in the form of an FG table, which is applied to the dataset before being processed by SERPent. Whilst this can be added in the calibration pipeline, efforts may be made to have the option to introduce an FG mask into SERPent. The obvious benefits to this is that any known RFI will be removed, improving SERPent’s ability to remove the rest. This RFI mask has been implemented on recent L-band datasets presented in Chapter 3.

Radio Interferometric Calibration

Pipeline for COBRaS

The only true wisdom is in knowing that you know nothing.

Socrates This chapter addresses the interferometric calibration procedures for the COBRaS pipeline, which follows from the SERPent module (Chapter 2) for the initial reduction and flag- ging. This chapter also includes discussions on wide-field, wide-bandwidth imaging with e-MERLIN and mosaicing techniques with a heterogeneous array.

3.1

Calibration Coefficients

This chapter contains the methods which are used to calculate the relations between the measured interferometric visibilities and the true sky brightness described in Chapter 1. The interferometric equation (the two-dimensional simplified interferometric equation) is reproduced here as Equation 3.1 for reference,

V (u, v) = Z l Z m B (l, m) e−2π i (lu + mv)dl dm, (3.1) 87

where V are the visibilities measured by the interferometer, B is the sky brightness, u, v are the coordinate system for the interferometer and l, m are the corresponding sky coordinates.

The correlator outputs visibilities which contain errors representing effects within the interferometric system. These can range from atmospheric effects distorting the wavefront (which affect the phases at different antennas), differences in system electronics over the observing time, frequency dependent antenna gains etc., all affecting the polarised electro- magnetic wavefront. To a good approximation, interferometers behave as linear systems (Fomalont and Perley 1999). The visibilities are sampled at discrete times and therefore the formula relating the measured visibilities V to the sky brightness B for antennas i and j (baseline i j) can be written linearly as

Vi j(t) = Z l Z m Gi j(t) Bν(l, m) e−2πi (ui j(t) l + vi j(t) m)dl dm. (3.2)

Equation 3.2 can be abbreviated to

Vi j(t) = Gi j(t) Bi j(t) + i j(t) + ηi j(t) , (3.3)

where t is the time of the observation, Gi j(t) is the baseline based complex gain, i j(t) is

the baseline based complex offset and ηi j(t) is the stochastic complex noise. The complex

offset and complex noise derive from the use of two complex correlators (Sine and Cosine correlators).

Initial calibration is conducted on an antenna basis where the baseline based complex gain Gi j(t) is the product of two associated antenna based complex gains gi(t) and gj(t):

Gi j(t) = gi(t) gj∗(t)

= ai(t) aj(t) ei(φi(t) − φj(t)), (3.4)

where ai(t) is the antenna based amplitude correction and φi(t) is the antenna based

phase correction for antenna i.

For e-MERLIN, there are two effects which need calibration, the complex electronic gain coefficients for phase and amplitude calibration (Gi j = gigj∗) and the bandpass

gain coefficients for bandpass calibration (Pi j = pipj∗). These coefficients account for

3.8 and 3.9.