Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.eso.org/~qc/concept/Phoenix.pdf Дата изменения: Tue Jan 8 19:53:10 2013 Дата индексирования: Sun Feb 3 07:40:01 2013 Кодировка: Поисковые слова: phoenix

Largely automated science processing (working title: phoenix)
Reinhard Hanuschik, 2012-05-10

Pre-conditions
· · · all associations (ABs) already done, by standard QC workflow all associations point to certified and archived master calibrations (the ones to virtual calibrations can't be processed so they need to be filtered) we use only selected pipelines ("workhorse"/"flagship"/"bread and butter"): o o o o · certified (reviewed by SDP) instruments with homogeneous data properties (science data types, calibration plans, headers stable over years) instruments with sufficiently demanded modes (we don't aim for completeness but for optimal turnback of investment) pipelines with stable and robust processing

products get seamlessly ingested as IDPs o o o aim: no involvement by QCG needed acceptable: starting set of configuration done by EDP and then reviewed/certified by QCG automatic ingestion done as part of the workflow (we never had that before but it seems reasonable and doable)

Process
Model 1: · global solution: register date when executable science ABs have been created (as part of dfos workflow): (executable means: no virtual calibrations, and pipeline/instrument (mode) is registered as "phoenix certified")

database
sciproc_date sciproc_ab

sciproc_date: instr, date, abs -1-


· or

execute on central platform, have monitoring interface, have operator look into status

Model 2: · · local solution: fill a JOB_FILE during day as part of dfos workflow execute locally, over night (because then there is no load and no interference with incremental daytime QC processing)

Phoenix workflow
for each scheduled date: · download mcalibs (bulk download since many ABs might require same mcalibs) into $DFO_CAL_DIR/ read MCALIB list; ignore MASSOCs check for required gencalibs This is a single process, needs to be done once per date and INS process ABs download raw data within each job use processAB can be massively parallel; total execution time is T = t(single AB)*N(ABs)/N(cores) delete raw files immediately after processing (by post-plugin) have minimal score-like process to evaluate association quality, plus measure pipeline processing quality (could be e.g. S/N) into QC1 params ingest QC1 params and scores (?) display status, logs, scores on process monitor ingest products few (mostly primary) only, no intermediate steps (?) must be automatic(!), at the end of processing

·

·

after scheduled job: · · · feed info on process monitor delete all local mcalibs (except gencalibs) and sciproducts store process info in database (logs, scores, ingestion logs)

All of this could be done in the background, with an operator checking the process sanity. -2-


Day-to-day processing vs. back-processing
The described scenario is applicable for day-to-day processing of new science data. · load balancing: incremental daytime processing of calibrations is not affected; science jobs are auto-scheduled only during the night from experience the processing power of dfo blades was usually high enough but it might be wise to envisage a dedicated server for day-to-day processing (like the pre-img server dfo33) to monitor and schedule these tasks, the currently existing dfos tools are sufficient.

· ·

·

The same scenario could be applied for back-processing (to close the gap between now and 2011-1001), or re-processing (to process backwards the entire data history of an instrument), but: · · then we need a scheduling and monitoring tool the process then needs to be monitored and maintained

IDP ingestion
Most if not all issues with day-to-day processing are expected for the ingestion process. Issues: · which products do we select for ingestion? The trivial answer is "the final ones" but what does it mean e.g. for UVES? o standard setups between 2001 and 2006 had a master response curve, the nonstandard ones not after 2006, the master response curves have not been updated; the 2009 detector upgrade has not seen any corresponding master response curve we cannot generally provide flux calibration should we give up on it (case 1) or support it whenever possible (case 2)? Case 1: final products come sometimes flux-calibrated, sometimes not Case 2: final products never come flux-calibrated, always wave-calib only o do we provide the error file per final product? If so, how?

-3-


·

How will EDP requirements for adding keywords look like this time (for IDP/VO compliance)? o in the UVES reprocessing project 6 years ago, this turned out to be one of the workflow components that required most efforts, without paying back anything in the end (e.g. the lengthy discussion about the "proper" S/N value per spectrum costed a month of work, without this number being visible anywhere in the end) for the GIRAFFE reprocessing project finished more than a year ago, the lack of a data model has even prevented any publication this is in general true for all QC SCIENCE data products

o o ·

this must be kept at a minimum level o o otherwise we are again limited by lack of standards/concepts, or spend precious time with header compatibility issues rather than with science grade processing

·

what does minimum mean? o o as much metadata information as possible to be read from the header additional information as far as possible from configuration files or by database processes for instance: if the header requires information about "pointing accuracy", it should be possible to provide this at database level

What would all this mean for UVES?
· Typical nights with 50-100 science ABs: 1-2 hours execution time, performance was never an issue on dfo21 processed modes (since pipeline certified!): o o o ECHELLE, point source (there is also EXTENDED) ECHELLE, ABSORPTION-CELL note: the flux calibration is not certified, strictly speaking! -4-

·


o ·

no distinction for SM vs. VM, standard vs. non-standard setups

no processing: o o ECHELLE, SLICER (why not?) FLAMES/UVES alias UVES/MOS: unstable/delicate pipeline, not certified

Impact on QC workload
· · Of course: some impact on setting this up no impact on day-to-day operations: all is done automatically, no decisions to be taken; only a process monitor required provided we can automatically cdbIngest ... true for day-to-day processing; not for back-processing or re-processing (some more monitoring required, pus some development for monitoring and maybe scheduling tools)

· ·

Requirements: · · · · one attempt only: the AB either processes fine or fails that information should be stored in a database and be displayed on the ADP user interface if pipeline issues are discovered: these should be taken up by SDP QCG will not provide any processing comments, or feedback to users or pipeline developers or SDP QCG will just provide the platform for processing (both hardware and software-wise) if issues show up with ingestion or pipeline, the process could be stopped anytime and resumed after fixing the issue, without the requirement to process the backlog in general, there cannot be a new operational requirement on QCG related to phoenix (unless it would be balanced by manpower, of course) there can also be no commitment to the speed of the process (if monitoring reveals bottlenecks, these should be taken up efficiently by SOS but not tackled by QCG)

· ·

·

·

-5-


Sketch of phoenix processing workflow
Day-to-day processing: this workflow is executed once per day and per selected INS

download all mcalibs

not all available? error and exit

process ABs, in parallel (on as many cores we can find)
download raw download raw AB1: process, derive assoc scores, extract QC1, delete raw AB2: process, derive assoc scores, extract QC1, delete raw AB3: process, derive assoc scores, extract QC1, delete raw AB4: process, derive assoc scores, extract QC1, delete raw

download raw download raw

cdbIngest selected products

delete all mcalibs and raw

delete all products

cleanup system, store process information

-6-