s&c system overview · – archive centric architecture – common software layer – multiple...
TRANSCRIPT
S&C System Overview
Tuesday, 14 February 12
Outline
• Context• System architecture
Tuesday, 14 February 12
SKA high level schematic
SKA Conceptual
Block Diagram
RegionalScience
Centre(s)
RegionalEngineering
Centre(s)
ScienceComputing
Facility
OutlyingStation
OutlyingStation
OutlyingStation
OutlyingStation
OutlyingStation
Dense Aperture
Array
Operationsand
MaintenanceCentre
SignalProcessing
Facility
HighLow
Sparse Aperture Array
On Site
SKA HQ(Off Site)
Global
High Performance ComputingData Storage
Digital Signal ProcessingBeamformingCorrelation
Outlying stationson spiral arm
(only one arm is shown)
Wide band single pixel
feeds (WBSPF)
Phased array feeds(PAF)
Dish Array
Tuesday, 14 February 12
Physical layout of SKA1 receptors
Core (50%)
Inner (20%)
Mid (30%)
1km
5km
100kmTuesday, 14 February 12
Science information flow
Science proposalDesign studiesSurvey strategySurvey targetsSurvey validationSurvey publication
Survey Science Teams
Science CaseScientific requirementsSST interactionsScience commissioning
SKA Science Office
ObservingScience data processing and reprocessingData validation
SKA Telescope
Archive of observationsAccess to data products
Science Archive
Tuesday, 14 February 12
Science data flow
Telescope Operating SystemSKA operations realm
Cloud based science analysis and visualisation
Science proposal
Observation definition
TOS data store
SKA science realm
Colle
ctor
s
Data
ex
cisio
n
Corre
lato
r
Data
ex
cisio
n
Data
ro
utin
g
Visib
ility
proc
essin
g
Imag
e pr
oces
sing
Science product archive
Sky models, calibration, pulsar
data, visibility data, etc.
SKA telescope realm
Local science analysis and visualisation
Tim
e se
ries
proc
essin
g
Pulsa
r sea
rch,
tim
ing,
tra
nsie
nt
dete
ctio
n
Tuesday, 14 February 12
Data level definitions
Enhanced data products e.g. Source identification and association
Validated science data products (released by Science Survey Teams)
Calibrated data, images and catalogues
Measurement sets
Correlator output
Beam-former output
ADC outputs
7
6
5
4
3
2
1
SST
SST
SKA
SKA
SKA
SKA
SKA
DefinitionLevel Responsibility
Enhanced data products e.g. Source identification and association
Validated science data products (released by Science Survey Teams)
Catalogs, timing results, transient detections
Pulsar search, pulsar timing, transient detections
Time series
Beam-former output
ADC outputs
7
6
5
4
3
2
1
SST
SST
SKA
SKA
SKA
SKA
SKA
DefinitionLevel Responsibility
Tuesday, 14 February 12
Outline
• Context• System architecture
Tuesday, 14 February 12
Overview of architecture
• An example of an architecture that could grow into SKA architecture
• Developed by ASKAP leads (Humphreys and Guzman) and so looks like ASKAP architecture
• Other possible architectures will be evaluated during Stage 1 of PEP
Tuesday, 14 February 12
Key driving assumptions
• Automated, real-time data processing• Service observing (mostly)• All required information flows from M&C
seamlessly• Data rich environment• HPC essential• Multiple types of data products• 2+ Tier data archive hierarchy
Tuesday, 14 February 12
Comments on other architectures
• ASKAP and LOFAR– Similar in design to each other– Closest in scale and requirements to SKA
Phase 1• MeerKAT (KAT-7)
– Architected for full system simulation
Tuesday, 14 February 12
Comments on other architectures
• ALMA– Archive centric architecture– Common software layer– Multiple (regional) data access centres
• LSST– Processing and archive is a unified system,
the Data Management System (DMS)– Impressive data/process model
Tuesday, 14 February 12
Reference S&C architecture
• Straightforward SOA-like approach– Telescope Operating System – Common Processor– Science Data Archive
• Enterprise Service Bus– Except for visibility and time series data
Tuesday, 14 February 12
Data model
• Each subsystem has own, distinct data persistence needs– Alternate (e.g. ALMA): one persistent store
• Multiple data formats– MeasurementSets, FITS, Catalogs– Translation to VO required– Alternate (e.g. LSST): data models translate
easily or trivially to VO
Tuesday, 14 February 12
Architectural goals
• System decomposition• Loose coupling• Reliability and Robustness• Support technology evolution• Promote software consistency
Tuesday, 14 February 12
Loose coupling
• Hardware platform independence• Language independence• OS independence• Implementation independence• Location and server transparency• Asynchronous communication
• Loose coupling not appropriate for large data
Tuesday, 14 February 12
Reliability and robustness
• SKA will be remotely operated– Must be reliable and robust
• Have the ability to operate in degraded mode where appropriate – Identification of single points of failure. Relevant to
location services, alarm management, logging and monitoring components.
• Support the deployment of redundant / high-availability services where possible. – Services must be stateless or idempotent
Tuesday, 14 February 12
Support technology evolution
• Technologies very likely to change during 10 year development
• Must support evolution of hardware and software components
Tuesday, 14 February 12
Promote software consistency
• Development of SKA software is likely to be distributed over multiple teams around the world
• Unless carefully managed, this may lead to fragmented software development processes, design and implementation.
• The architecture must aim to control this fragmentation and promote consistency.
Tuesday, 14 February 12
System description
• Telescope Operating System– Observation preparation, scheduling and
execution (data acquisition); – Operator displays, startup, shutdown and
maintenance software. – Data acquisition coordination
• antenna motors, beamformers, correlator, timing synchronisation, etc. levels 1 - 5 data products
Tuesday, 14 February 12
System description
• Common Processor– on-line or near real-time data processing,
producing level 5 data products • Science Data Archive
– archiving data products coming from the common processor (level 5 data products), pushing data products to regional centres and some user access for data analysis.
Tuesday, 14 February 12
SKA Common Software
• Provides software infrastructure for all three components
Operating System
CommunicationMiddleware
DatabaseSupport
3rd Party Toolsand Libraries
Access Control
Monitoring Archiver
Live Data Access
Logging System
Alarm Service
ConfigurationManagement
SCA Application Framework(Support for C++/Java/Python languages)UIF Toolkit
DevelopmentTools
SKA Applications
1. Base Tools
2. Core Services
3. High-level APIs andTools
Scheduling Block
Service
Tuesday, 14 February 12
Benefits of Common Software
• Minimises the problem of integration hell by providing a common way to implement interface between components and common services
• Minimises the support burden required by the maintenance staff that would otherwise have to absorb development differences from the very distributed and heterogeneous SKA development teams.
• Minimises reinventing the wheel by letting application (business domain) developers focusing on domain-specific application code rather than technical infrastructure code, also referred sometimes as plumbing code.
• Allows an organised way of sharing and reusing design concepts and patterns.
• Promotes a cohesive set of development practices, standards and tools across different development teams
Tuesday, 14 February 12
Capabilities of SCS
• Observation or experiment data model definition and management
• Logging and monitoring data collection and archiving• Alarm Management• Component communication infrastructure• Component and common configuration. For example, station
composition, station/element location, etc• User Interface (UI) framework• Authentication and access control • Software development tools: build system, unit test frameworks,
deployment system, etc.
Tuesday, 14 February 12
TOS and CP subcomponents
Central ProcessorTelescope Operating System
ServiceContract
ServiceContract
ServiceContract
ServiceContract
ServiceContract
ServiceContract
ServiceContract
Monitoring Archiver Log Archiver Executive
Processing PipelinesData Services
Telescope Observation
Manager
Alarm Management Service
Operator Displays Observation Management Portal
FacilityConfiguration
Manager
ServiceContract
ServiceContract
Sky Model Service Light Curve Service
Scheduler User Interface
ServiceContract
ServiceContract
RFI Source Service
ServiceContract
Central Processor Manager
ServiceContract
Calibration Data Service
Enterprise Service Bus
Tuesday, 14 February 12
Common Processor
• Responsible for transforming data from correlator or direct from beamformers into science products
• Specific functionality for imaging– Conditioning of the data to remove known errors such as RFI, – Calibration of observed visibilities; – Processing of calibrated visibilities into images/cubes; – Source finding/extraction within the produced images; – Detection of transient sources and production of light curves; – Quality evaluation of all data products; – Provenance tracking of all data products as they are produced and
processed.• Also time series processing, software correlator
Tuesday, 14 February 12
Common Processor
• Both a hardware and software subsystem,• High-performance computer and a suite of calibration,
imaging and analysis software. • Highly optimised and parallel software, designed
specifically to support processing high-rate data streams in quasi real-time.
• The high data rate, and likely inability to buffer (store) the visibilities for more than perhaps 24 hours, requires processing occur at the same rate as data acquisition.
Tuesday, 14 February 12
Visibility data buffer
• What to do with that huge data flow?– 3.3 TB/s input
• Two options– Buffer– Stream
Tuesday, 14 February 12
Buffered
• Store data to non-volatile storage medium for part or the whole of an observation
• Benefits– Enables multi-pass algorithms, which iterate over the
dataset multiple times; – Simplifies failure handling, allowing handling of most
common processor hardware failures without data loss; – Enables optimisation of the ordering of processing– Allows potentially smaller memory (RAM) footprint given
there is no need for all spectral channels to be processed simultaneously.
Tuesday, 14 February 12
Feasibility of buffering
• For single write and single read, required a storage system capable of sustaining simultaneous 3.3TB/s write and 3.3TB/s read would be required.
• Current record is from Jaguar supercomputer at Oak Ridge National Laboratory– Demonstrated bandwidth of 240GB/s (3.5% SKA1 requirement).
• SSD’s? Phase change memory
Tuesday, 14 February 12
Streamed
• Correlator integrations stream through pipelines• Eliminates costly buffer• Drawbacks
– No support for algorithms which require iteration over the dataset;
– Node failures will usually result in the loss of data; – Few data ordering optimisations possible; – Jitter (variation) in processing time may result in data
loss; – Potentially larger memory (RAM) footprint.
Tuesday, 14 February 12
Pipelines
• Streaming: acyclic (one-pass) processing• Buffering: cyclic (multi- pass) processing • Each pipeline consists of
– one or more data sources – one or more processing stages – one or more data sinks.
• System pipelines– data conditioner– calibration pipelines.
• One or more science oriented processing pipelines– imaging – source finding and fitting
Tuesday, 14 February 12
Data conditioner pipeline
Tuesday, 14 February 12
Science Data Archive
Resource Layer
Fast Online Storage
Offline/Nearline Storage
Database Management
System
Data Access/Management Layer
Data Ingest Services
Data Access and
Management Services
Data Replication
Services
Presentation Layer
VO Interfaces
Web Based Data Product
Access
Data Curation and Management
Interface
S
Astronomers
Central Processor
Regional Science Centres
Tuesday, 14 February 12
Summary
• Reference architecture exists• Based on ASKAP architecture• Conservative SOA-like approach
– Data do not flow across ESB• Many questions remain
– e.g. scaling to large number of antennas• Will explore other options as part of PEP
Stage 1
Tuesday, 14 February 12