CABReport0002 - david-macmahon/wiki_convert_test GitHub Wiki

Location and Attendance

  • Date: 29 September 2009
  • Location: The Boardroom, Atlantic Imbizo Conference Center, Cape Town
  • Present: Matthew Bailes, John Ford, Yashwant Gupta, Justin Jonas, Alan Langman, Jason Manley, Aaron Parsons, Jonathan Weintroub, Dan Werthimer
  • Absent: Glenn Jones

New acronym, reflecting decentralized structure

CASPER will now be the "Collaboration for Astronomy Signal Processing and Electronics Research"

Overall Direction

  • Not exclusively an FPGA shop, with a desire both to be seen and to actually be “platform neutral”.
    • CASPER is really about creating Open Source tools for solving astronomy signal processing problems quickly.
  • CASPER is, presently, FPGA-focused, however might become:
    • a central clearing house, using web pages/wikis, for more general developments in areas such as GPUs, CPUs, Back End Software, initiatives such as AIPY, coherent data storage & software signal processing
    • or for any technology which enables fast development times of digital radio-astronomy instrumentation
  • General theme is “packetized” and using "COTS" solutions rather than re-inventing protocols, data transport systems etc (for larger systems) (smaller?)
  • Aaron: People are starting to look at CASPER for what lies beyond the correlator. Thus the output formats should mesh seemlessly with existing solutions for post-processing.
    • Real time imaging
  • Matthew: CASPER boards allow the possibility to create ...
  • Invite individuals with alternate approaches to the 2010 workshop, e.g. have NVIDIA GPU people give a presentation

Focus for coming year: Infrastructure

  • ROACH II based on Virtex 6 is targeted for release in 2010
    • Keep ROACH_2 as simple as possible
    • Driven by SA SKA
    • Concern about feature creep: a small group of people should vet suggested features
    • (Suggested committee by Dan: Alan, Francois, Mo)
    • Expect design of ROACH II to go much faster than I.
    • Yashwant: 10 Gig connector will affect what switches people buy
  • Stabilization of Simulink based tools and libraries
    • This is a high priority before considering open source took kits
  • In 2010 research the feasibility of an Open Source toolset
    • (Add some thoughts on practicality and possible frameworks)
  • Standardization of the instrumental interface
  • Documentation (as usual)
    • fobbing off to user community may be poor model
    • require that new contributions to the library are checked in with complete documentation package.
  • Still need in person training for new recruits, but decentralize this.
    • Consider maintaining a register of geographically convenient training sites whom new recruits can contact

Publish the benefits of CASPER to the broader astronomy community

  • Focus on “Design Wins”
  • Link to published science results using CASPER technology
  • Briefs on actual CASPER applications deployed and used for science
  • Also, memos describing suggested CASPER architectures for different real applications which are not necessarily CASPER wins,, e.g. SKA, MWA, EVLA, etc
  • Within next year actually build and deploy a large N correlator. See note below by Jason Manley:

Prospects for large-N correlator deployment in 2010

Instrumentation research appears to need a science driver to motivate fast progress. Fortunately, there are many such drivers for the packetised correlator and 2010 will see CASPER's packetised correlator growing both in complexity and in scale.

PAPER is pushing ever larger correlator systems with 64 antennas deployed in 2010 and 128 antennas by year-end. It is expected that ROACH's resources will be sufficient for a 64 antenna correlator. However, a correlator of 128 dual polarisation antennas is better suited to the next generation hardware (ROACH-II), which is only anticipated to go into production at year end. Fortunately, it is believed that porting of the correlator from ROACH-I to ROACH-II will be trivial and so it seems likely that larger-N correlators will become a reality in early 2011.

Additionally, 2010 will see smaller systems deployed but with focus on new features and system reliability, design stability and standardisation rather than gross size. KAT-7's 8-input correlator is driving the addition of delay compensation (coarse and fine) and fringe rotation, developed in collaboration with GMRT. The first phase of this system should be ready for testing early in the second quarter of 2010, with deployment scheduled for mid year. GMRT will make use of a larger versions of the same system, compiled for 32 inputs.

In addition, various aspects of the design have been standardised, including the data exchange format and control and monitoring protocols. The system will eventually have a fully-automated built-in-selftest, significant for system reliability as the design becomes ever larger and more complex. Effort in designing flexible software frameworks to enable interation with computer systems and extending the processing pipeline towards real-time imaging is ongoing.

Integrated beamforming functionality is expected towards year-end as part of KAT-7's second phase.

Specific development ideas

  • Need an RFI solution
    • Basic RFI block
    • Re-quantization
    • Kurtosis processing
    • Source cancellation
  • Categories of Correlator
    • Large N large bandwidth
    • Large N small bandwidth (Quite a challenging case in terms of efficient FPGA utilization, and an important one, since it's typical of EOR applications.
    • Small N wide bandwidth

Concerns

  • Feature creep on ROACH
    • e.g. (Dan) ACTEL CPLD
  • Took a year to port libraries from 7.1 to 10.1. Solution needed.
    • Move cores to a different language which is easier to port
    • Would still look like Simulink at top level
    • Jason: Possible intermediate solution is to add another layer on top of Xilinx Sysgen... to wrap Xilinx primitives (logical blocks, counters, memories etc) into our own primitives. These 'CASPER' primitives would then be used by all CASPER DSP blocks, rather than the Xilinx natives directly. When porting to new Xilinx tool, only need to port CASPER primitives. DSP blocks would remain largely unaffected.

CASPER workshop 2010

  • Much thanks to SA SKA for hosting the 2009 workshop in such style
  • (Add venue and dates once decided)
    • being voted on presently between NRAO Green Bank, NRAO Socorro, and CfA