Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.naic.edu/~tghosh/vfuture/soc07/Brisken_swcorr-usvlbi.pdf Дата изменения: Tue Jun 5 20:45:39 2007 Дата индексирования: Mon Apr 11 16:40:24 2016 Кодировка: Поисковые слова: п п п п п п п п п п п п п п п

Software Correlators for VLBI

Walter Brisken May 14, 2007 U.S. VLBI meeting


Why software correlators?




Rapid Development Core software available for free DiFX written by grad student Adam Deller in < 1 year Flexibility Limitations are not built in Arbitrary number of inputs, spectral channels, integration time Special processing modes can be implemented e.g., pulsar binning Can support multiple file formats Scalability Can add more processing power to increase performance CPU power is fully usable for any number of inputs


DiFX architecture

Diagram by Adam Deller


DiFX for the VLBA
Immediate goal: integrate DiFX into VLBA operations Use existing path for generation of correlator jobs Use same delay model (Calc 9.1) Add to DiFX direct Mark5 access Add operator tools that interact via multicasts Convert output to FITS-IDI



A prototype cluster
Use Mark5s as a cluster for a software correlator Upgrade motherboards to dual-dual core XEON CPUs Cost ~= $15,000 18 Mark5s should correlate 10 stations at 128 Mbps in real-time Software correlation does not interfere with Mark5 playback to the hardware correlator


Note 1 : for 10 stations about 2 Gflop compute power per Msample/sec is needed. Note 2 : Computing costs about ј storage cost for the VLBA's typical usage pattern.



First Fringes!


First with old version Found some VLBA specific bugs New version now working

Phase





Amplitude


10 Antennas on existing Mark5 units sustained 16 Mbps equivalent speed


DiFX Multicast System (draft plan)


Emitted by several software components:
­ ­ ­

DiFX Fringe Tool Queue Manager Progress Warnings Data quality



Messages contain information on:
­ ­ ­



Simplicity through decoupling Driven by EVLA's multicast success


Fringe Tool


Used to provide real-time feedback of data quality to operators Data Products:
­ ­



Baseline amplitudes/phases/delays Auto-correlation spectra



Results to be Multicast to operator display Clock search on-the-fly possible


Optimistic Time-line


Project Start : Feb 23, 2007 First fringes goal : Apr 1, 2007
­

Actual : Apr 23, 2007 Actual : May 1, 2007



First correlation off Mark5A : Jul 1, 2007
­



Modules-to-FITS in 1 day : Jun 1, 2007 Usable by operations : Sep 1, 2007 Routine correlation : Nov 1, 2007


Open issues


Playback of incomplete modules
­

Is this even possible using XLRRead()? Will it be affordable?



Can this be scaled to 4 Gbit/sec on 10 stations?
­