Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.atnf.csiro.au/research/pulsar/orange10/pdf/AidanHotanOrange2010.pdf Дата изменения: Mon Sep 27 10:28:53 2010 Дата индексирования: Sun Feb 13 23:48:10 2011 Кодировка:

The Legacy of CPSR2

8 years of safe, reliable, high-precision pulsar timing.
Aidan Hotan
Curtin University
Orange 2010


Summary


Design philosophy, data pipeline (Bailes et al.). Operational deployment. The user experience. Key science outcomes. Lessons learned. Convergence of instrument technology. The future.
Orange 2010


Design Philosophy


Baseband recording and coherent dedispersion. Maximise use of "standard" components.




Digitisation remains the most problematic stage. Reduce non-recoverable engineering (NRE) costs. Easier to debug and develop incrementally. Leads to logical hardware upgrade paths. Allows for very flexible experimental configurations. Scope for multiple re-processing attempts.
Orange 2010



Push computation into the software domain.



4x IF, clock, 1PPS

Software Pipeline


2-bit Digitiser
DMA

The "firehose" strategy. No single node can process in real time. Data divided into sequential time slices. Nodes written to in a wrapping sequence.
Monitoring and control

64 MB/s each Primary Node 2 (Memory Buffer)

Serial control



Primary Node 1 (Memory Buffer)



Archives



Switch (1Gb) Node 1 Node 2 Node 3

Switch (1Gb) Node 1 Node 2 Node 3
Raw data files written to disk for processing

Node 14 Switch (100Mb)

Node 14
Orange 2010


CPSR2: Commissioning Stage
(Around back) Parallel cables from FFD to DMA cards in head nodes. Caltech FFD, 2x 64 MHz bands. Ethernet switch (x2). Giganet switch (soon discarded due to driver problems). 2 head nodes with extra RAM; EDT DMA cards overclocked to 64 MB/s. 28 dual-processor servers with SCSI HDDs, 1Gb ethernet. Remarkably reliable, 8 years with few failures. Later installed in RFI enclosure, heat dissipation problems.

Orange 2010

IF & clock cables.


CPSR2: The User Experience


Operating close to the edge of performance:


Only a few seconds of primary buffer space. Linux is not a real-time operating system.
­

Waiting on a system call could crash the observation.



Required 2/3 of theoretical 1Gb network capacity. Disk writing on nodes could become a bottleneck.
­

Particularly when drives started to fill.



Detecting and dealing with single node failures. Complicated chains of commands (and 30 xterms) required to start the system.
Orange 2010


Key Science Outcomes


Highest precision timing (Verbiest, Hotan, etc.)


Stability (8 years with no hardware changes). Simplicity (few hardware-induced systematics). Full Stokes available for small computational cost. Time resolution limited only by bandwidth and the number of frequency channels kept. Can fold multiple pulsars simultaneously (clusters).
Orange 2010



Polarisation (Ord, Karastergiou, etc.)




Giant pulses with high time resolution (Knight)





Convergence of Technology


Hardware correlation and filterbanks:


Wide band, but limited number of channels. Leads to dispersion smearing in pulse profile. A niche technique for precise measurement. Bright sources over small bandwidths only.



Coherent dedispersion:




Today, hardware is fast enough that we can coherently dedisperse wide bands, or create large numbers of filterbank channels.
Orange 2010


In a digital filterbank, faster FPGAs allow for more frequency channels.

Incoherent channels, wide bandwidth.

Coherent dedispersion, small band.

In a baseband recorder, faster computers and interconnects allow for wider bandwidths. Baseband recorders still have the advantage of flexibility.


Lessons Learned


Complete re-install of all nodes after security breach on ATNF network.


Much improved methods for Operating System and server hardware management now available.



Two different control systems developed, neither of which were ideal.


Plan the full software pipeline in advance! The finishing touches are important, especially to promote wider usage amongst external users. So the observers don't wake the developers at 3 AM.
Orange 2010



Documentation is important!



Useful Spin-offs


PSRCHIVE (http://psrchive.sourceforge.net )


Well maintained, general software suite for pulsar data processing, with multiple archive formats. Evolved from PSRDISP, now supports multiple instruments and multi-threaded operation. Complete memory-based data pipeline, with support for dspsr and monitoring hooks.
Orange 2010



DSPSR (http://dspsr.sourceforge.net)




PSRDADA (http://psrdada.sourceforge.net )



The Future


APSR (ATNF Parkes Swinburne Recorder)


8-bit sampling, up to 1 GHz of bandwidth. Nyquist samples tapped from PDFB. Commercial 1G-sample PCI cards now available. Complete remote control using web interface and standard server tools.



Coherent searching? Transient detection? Beam-formed data?
Orange 2010

APSR web interface