NATIONAL RADIO ASTRONOMY OBSERVATORY Computing Architecture Jeff Kern
NATIONAL RADIO ASTRONOMY OBSERVATORY Computing Architecture Jeff Kern ng. VLA The Next Generation Very Large Array
ng. VLA Definition (from a software perspective) ng. VLA (noun): A geographically distributed large N array for PI driven science with high sensitivity, wide instantaneous bandwidth, and low operational cost. The Next Generation Very Large Array 2 Radio Futures III 2 -4 Aug. 2017, Berkeley CA
ng. VLA Computing Architecture Constraints • Geographically distributed • Need full suite of operations tools (Proposal, Observing Preparation, Data Processing) • Produce science ready products • Antennas must be self protecting • Local control for maintenance • Minimize tight central communication loops • High Sensitivity • High calibration and imaging accuracy • Large N Array (200+) • Snapshot observations likely • Avoid central bottlenecks (logging, monitoring) • Wide Instantaneous Bandwidth • Robust to elements entering and • RFI concerns leaving the array • Data Volume (again) • Large (but not unsupportable data • Low Operational Cost volumes) • Automation everywhere possible • PI Driven Science The Next Generation Very Large Array 3 Radio Futures III 2 -4 Aug. 2017, Berkeley CA
Control and Monitoring • In principle M&C of the ng. VLA is very similar to existing arrays. • Devices have gotten smarter, interface to software at a higher level. • Use time tags for time critical operations and do synchronization in HW. al. T Re • Use standard communication protocols o N • Large number of antennas and larger geographic area implies: • Subarrays will be used nearly continuously (esp. during commissioning) • Antennas need to be very independent, local monitoring and error detection • Send sky position rather than encoder commands • Operations Optimizations • Dynamic Scheduling algorithm • Need a more sophisticated weather model than typically used (variable across the array). • Hardware drivers should self diagnose when updates to parameter data needs to be made, or operation conditions are out of normal. • May need to include more self check modes in the hardware design. The Next Generation Very Large Array 4 Radio Futures III 2 -4 Aug. 2017, Berkeley CA S O e m i
Correlator Software • Correlator details TBD but assume: • • FX Architecture Assume delay tracking performed digitally in Correlator HW Subarrays will be needed early and often. Correlator Back End, cluster for processing and formatting • More conditioning of visibilities in this stage than for VLA • • (Tsys, Sub-band stitching, Telcal Application) RFI Mitigation (or even further upstream) Data Flagging and blanking Data into final archiving / processing format • Correlator functionality is easy to add in hardware. 5 • Cost is in commissioning and support (firmware and software complexity) • ng. VLA should limit correlator modes to those that have clear use cases. Com miss ione d The Next Generation Very Large Array Radio Futures III 2 -4 Aug. 2017, Berkeley CA
Data Archiving and Storage • Estimated archived data rates are few TB per hour (visibility only) • Not trivial, but possible (esp. with continued improvements in storage cost). • Baseline design is to store the “raw” visibilities. • But current paradigm doesn’t work • Moving data, and multiple copies will likely be prohibitively expensive • “Filling” data from an archival to working format, expensive and unnecessary • Beginning to work on Measurement Set V 3 with SKA • Designed for massively parallel processing • Robust to node failures, thread safe • Support backwards compatibility to MS V 2 • Compute cluster should be “close” to storage • Primary processing on PI hardware unlikely The Next Generation Very Large Array 6 Radio Futures III 2 -4 Aug. 2017, Berkeley CA
Data Processing • Algorithmic and performance requirements are still uncertain. • Needs more study now that the system design is stabilizing. • Preliminary analysis shows that most use cases are tractable • Assumes that compute cost follows historical trend • Wide-field low-frequency (<4 GHz) imaging is likely to be computationally prohibitive in early operations. • Excluded from the baseline for this reason. from SKA: s Key Difference r start te la d n a y c n e u Higher freq date • Delivered products for most projects will be science quality images • Science Ready Data Products (SRDP) project is a precursor to the ng. VLA pipeline • ALMA Calibration and Imaging Pipeline • VLA Calibration Pipeline • VLA Sky Survey Pipeline The Next Generation Very Large Array 7 Radio Futures III 2 -4 Aug. 2017, Berkeley CA
g n i n n gi to End Operations Be. End ng. VLA must be a telescope for all astronomers, not just radio “blackbelts”. • Proposal process should focus on desired products, not on hardware functionality. • Reprocessing will be expensive. Initial generation of the proper science images will be important • Archival researchers also need to be supported (next talk). • Product quality is an observatory deliverable, thus the observatory must control calibration strategy. • Challenge is to allow the flexibility that traditional radio astronomers expect. The Next Generation Very Large Array 8 Radio Futures III 2 -4 Aug. 2017, Berkeley CA
Software Architecture Summary: No blo c • The RMS community knows how to design, implement, and operate the software to run the ng. VLA. • Some changes in emphasis because of large numbers of antennas and baseline length. • Radio community is gaining experience with full lifecycle software support (Proposal to Delivery). • Lessons to be learned from current generation of NRAO telescopes. • SRDP project is pioneering science quality image production across the frequency range of the ng. VLA. The Next Generation Very Large Array 9 Radio Futures III 2 -4 Aug. 2017, Berkeley CA ker s
www. nrao. edu science. nrao. edu public. nrao. edu The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. The Next Generation Very Large Array 10
- Slides: 10