Instrumentele Groep Fysica

PuMa, the first Dutch Pulsar Machine

Index: [introduction] puma-in-.htm [central problem] [the challenge: PuMa] [PuMa hardware] [PuMa software] [project management] [specifications]

Introduction

Pulsars are neutron stars, which are rotating rapidly and emitting a bundle of broadband electro-magnetic radiation. This effect resembles a stellar lighthouse. The received pulsar signals are weak. The signal-to-noise ratio (SNR) is well below 1 for most pulsars. One should think of white noise with an amplitude modulation on the order of a few percent or less.
Typical pulsar rotation frequencies are in the range from 0.2 Hz to 642 Hz. The emitted radiation may covers the entire radio-spectrum; only a few pulsars are known to emit also x-ray and gamma radiation.

Methods

There are three distinct disciplines in pulsar observations:

Searching for new, unknown pulsars. Up to know, about 700 pulsars are found. However, some estimates indicate that about 5000 pulsars should be observable.
Precise determination of the arrival times of pulses is the next important experimental discipline. The required accuracy of this kind of measurements is better than 10 ns with respect to UTC.
Detailed pulse profile studies can take place in a number of varieties.

The Problem of Pulsar Observations

Pulsar observations suffer from two hurdles. First, there is the poor signal-to-noise ratio. The solution of this problem is to use long observations. Two ways are available to increase the number of measurements: by increasing the measuring period and by increasing the analyzed sky bandwidth. The first method works quite well to a certain extend, but becomes impractical thereafter. The second method immediately reveals the second hurdle: the effects of dispersion.
A work around of these hurdles is by splitting of the incoming antenna signal in multiple bands, each with a small bandwidth, and processing all these bands in parallel. In traditional pulsar machines, the incoming antenna signal is split with common heterodyne radio techniques using many analog bandpass filters in parallel. After each filter, there is a square-law power detector and a data acquisition stage. For practical reasons, the number of filters and detectors is of the order of 20. This implies, that a compromise has to be found between the total analyzed sky bandwidth and the effects of dispersion.

The Challenge

In 1996 the project started to develop a new pulsar machine, which we call PuMa. PuMa will become the replacement of the existing FFB and will be installed at WSRT.
The design goals of this project are the following:

to develop a machine that takes the full advantage of the incoming 80 MHz signalband, even under severe dispersion;
is optimized for searching and timing and profile studies;
allows for a great amount of experimental freedom;
and all within a limited budget and on a limited time scale.

The Setup

The intermediate frequency stages (IF) at WSRT outputs the antenna signal as 8 pairs of analog signals. Each pair represents the X and Y polarizations of a 10 MHz signalband. The analog signals of both polarizations of a pair are discretized immediately after the IF stage using ADCs sampling at 20 MHz. The digital data is transferred to a cluster of digital signal processors (DSPs). Each DSP receives a time series and starts executing an FFT and data reduction software. Once a DSP is finished processing its data, it transfers the results to a mass storage system and is available for the reception of a new time series. There are sufficient DSPs in the cluster to guarantee the seamless processing of all data.

Digital Signal Processor Boards

The digital signal processors are the core building blocks of PuMa. From the vast variety of available DSPs, the Analog Devices ADSP2106X SHARC has been selected

With a 20 MHz sampling rate, it takes about 100 ms to fill a DSP with 2048 samples. Processing these takes 1 ms for the FFT and another 1 ms for data reduction and storage routines. Thus in total at least 21 DSPs are required to handle the data from a single 10 MHz channel. Because of the built-in multiprocessor features of the SHARC it was decided to use 24 DSPs per cluster; the nearest multiple of 6 SHARCs.
The PuMa SHARC board comprises the following:

A data acquisition interface stores the data in the SHARCs. This interface is also responsible for the synchronization between the boards within a cluster.
6 SHARCs to provide the required processing power.
A VME interface for the connection to the mass storage system and for the supervisory control system. The VME interface is designed using the VIC64 chipset and allows for A32/D64 master and slave transfers.

Mass Storage and Supervisory System

There are two identical mass storage and supervisory systems, one per crate. Each system consists of a HP 743 VME workstation running HP-RT and four harddisks.

The applied disks are fast-wide-differential Cheetah devices (Seagate) with capacities of 9 GB per disk. The disks are operated in parallel to achieve the highest possible throughput. The sustained throughput to a single disk is measured to be 5 MB/s and by using four disks in parallel we aim to achieve an aggregate throughput in excess of 15 MB/s.

PuMa Software

Data taking in PuMa is accomplished by autonomous hardware. The software starts an each DSP as soon as the data taking of a time series is completed. For PuMa, different modes of operation are defined, corresponding to the scientific aims, and for each mode a special program exists.

Mode 0: raw compression

Mode 0 reduces the data flow to the prescribed maximum using simple compression techniques. The initial time series, sampled with 20 MHz can be reduced by means of resampling in software at lower frequencies. Low pass filters, i.e. Finite Impulse Response filters, are applied to avoid the effects of aliasing. Mode 0 is intended to be used for two purposes:

since it contains only very simple algorithms and stores fairly unprocessed data onto disk, it may be of help validating the upstream electronics (AD-modules and WSRT IF stages).

the high temporal resolution of the output stream allows sophisticated off-line analysis routines, such as coherent dedispersion, to be performed using the full WSRT bandwidth.

Mode 1: searching

Mode 1 resembles much the analog filterbank technique. The incoming data series are Fourier transformed and in this way the input signal is decomposed into many frequency bands.
Because both polarizations are present at this stage, it is possible to compute all four Stokes parameters instead of only the power signal. Using the Stokes parameters, it is possible to study the linear and circular polarization of pulsar signals.

Mode 2: incoherent dedispersion

Dedispersion is the process of adding frequency dependent time-shifts in such a way, that it cancels the effects of dispersion. This process clearly requires adequate estimates of the dispersion and can therefore be accomplished with known pulsars. Once the frequency dependent time shifts have been added to the data, the Stokes parameters for all channels may be combined to a single set, because the contained frequency information is of no further use. In this way, enormous data reduction is achieved without the loss of scientific relevant information. This process is called incoherent dedispersion, because it uses Stokes parameters that contain no phase information.

The Project

PuMa is developed in a joined project of five partners:

The Netherlands Foundation for Radio Astronomy (NFRA-ASTRON), who supplies financial means and technical consultants. NFRA s Westerbork Synthesis Radio Telescope (WSRT) is the prime application of PuMa.
The Department of Astronomy ``Anton Pannekoek'' of the University of Amsterdam has pushed the initiative, acts as the project scientist and further raised manpower and additional funding.
The Department of Astronomy of the Utrecht University, who supports the scientific supervision of the project.
The Department of Computational Physics of the Utrecht University, who provides technical support for the DSP developments.
The Instrumentele Groep Fysica (IGF) of the Utrecht University, who act as main contractor for the realization of this project.

Technical Specifications


Analog inputs:
Nr. of input channels	2 x 8
Input bandwidth	10 MHz	per input channel
Sampling frequency	20 MHz
Input voltage range	-1.1 V .. 1.1 V
Dynamic range	12 bits
Passband ripple	~ 1 dB
Stopband attenuation	60 dB
Total output stream	320 MSample/s
DSP boards:
Nr. of boards	32
DSPs per board	6
Estimated performance	120 MFLOPS	per DSP
	20.000 MFLOPS	in total
Nominal FFT rate	80.000 /s	2048 points complex FFTs
DSP internal memory	256 kB
Clock frequency	40 MHz
VME throughput	35 MB/s	peak performance, D64 mode
Mass Storage System:
Processor	2 x HP 743
Disks	8 x Cheetah
Total disk capacity	72 GByte
Maximum throughput	30 MB/s	(guestimated)
Minimum time to fill disks	40 min.

For further information please contact :
P. van Haren
Instrumentele Groep Fysica
Postbus 80.004, Sorbonnelaan 4
3508 TA Utrecht
The Netherlands

Telephone: +(31) 030 - 253 2293
Fax: +(31) 030 - 253 3267
email: D.Killian@fys.ruu.nl