Karabo: The European XFEL software framework Design Concepts

Burkhard Heisen for CAS

December 2014

Karabo: The European XFEL software framework

Design Concepts

The star marks concepts, which are not yet implemented in the current release


Functional requirements 2

DAQdata readout

online processing

quality monitoring (vetoing)

SCprocessing pipelines

distributed and GPU computing

specific algorithms (e.g. reconstruction)

Controldrive hardware and complex experiments

monitor variables & trigger alarms

DMstorage of experiment & control data

data access, authentication authorization etc.

setup computation & show scientific results

allow some control & show hardware status

show online data whilst running

A typical use case:

Accelerator Undulator Beam Transport

Sample Injector

DM SC

ControlDAQ

Tight integration of applications


Functionality: What are we dealing with?

1. Data containers (transport and storage through serialization)

2. Data transport (communication patterns)

3. Devices (distributed end points)

4. States and state machines (when can what be called/assigned on the devices)

5. Log Messages (active, passive, central, local)

6. (Slow) Control-Data Logging

7. (Fast) Data acquisition

8. Time synchronization/tagging (time stamps, cycle ids, etc.)

9. Real-time needs (where necessary)

10. Notifications and Alarms

11. Security (who’s allowed to do what from where?)

12. Statistics (control system itself, operation, …)

13. Processing workflows (parallelism, pipeline execution, provenance)

14. Clients / User interfaces (API, languages, macro writing, CLI, GUI)

15. Experiment, Run and Configuration management

16. Software management (coding, building, packaging, deployment, versioning, …)

3


Data containers 4

Some special data containers are provided by Karabo and are exposed in the API Hash

String-key, any-value associative container Keeps insertion order (iteration possible), hash performance for random lookup Provide (string-key, any-value) attributes per hash-key Fully recursive structure (i.e. Hashes of Hashes) Serialization: XML, Binary, HDF5 Usage: configuration, device-state cache, DB-interface, message protocol, etc.

Schema Describes possible/allowed structures for the Hash. In analogy: Schema would

be for Hash, what an XSD document is for an XML file Internally uses Hash

RawImageData Specialized class for transporting image-like data Easily convertible to numpy in Python and to CpuImage<T> in C++ Optimized serialization into HDF5 Internally uses Hash


STATUS: Data containers 5

Recent changes None

Future work Improve Hash serialization implementation with respect to the XML format to allow

for slashes “/” in hash-keys

Find a proper data object (eventually plus some description) to exchange data with

the DAQ layer

Open issues

Understand the conceptual difference between using standardized objects vs.

generic container + description throughout the system (see also DAQ section)


Data transport – Message broker based

Basic communication between objects is established via a central

message broker using a publish-subscribe pattern (topic based) Each communicating object is an instance of the SignalSlotable class which

connects to a configurable broker (host/port/topic)

The SignalSlotable API allows to register regular functions of any signature (currently

up to 4 arguments) to be remotely callable (such a function is called: Slot)

Slots can be uniquely addressed by a pair of strings, the instanceId (string name of

the SignalSlotable object) and the functionName (string name of the function)

Slot registration can be done during construction or later at runtime without extra tools

Slot calls can be done cross-network, cross-operating-system and cross-

language (currently C++ and Python)

The language’s native data types are directly supported as arguments

Additionally supported arguments are Karabo’s data objects (e.g. Hash and Schema)

Data packets are on the fly compressed/decompressed if reaching some size

threshold

6

New


DETAIL: Data transportBroker based communication API – Four Patterns

① Signals & Slots SLOT ( function, [argTypes] ) SIGNAL ( funcName, [argTypes] ) connect ( signalInstanceId, signalFunc, slotInstanceName, slotFunc ) emit ( signalFunc, [args] )

7

SLOT(onFoo, int, std::string);

void onFoo(const int i, std::string& s) { }SIGNAL(“foo”, int, std::string);

connect(“Device1”, “foo”, “Device2”, “onFoo”);

connect(“”, “foo”, “Device3”, “onGoo”);

connect(“”, “foo”, “Device4”, “onHoo”);

emit(“foo”, 42, “bar”);

Device1

Device2

Device3

Device4

Emit

Notify

Notify

Notify

SLOT(onGoo, int, std::string);

void onGoo(const int i) { }

SLOT(onHoo, int, std::string);

void onHoo(const int i, std::string& s) { }


DETAIL: Data transportBroker based communication API – Four Patterns

② Direct Call call ( instanceId, funcName, [args] )

8

Device2

Call Notify

Device1

SLOT(onFoo, std::string);

void onFoo(const std::string& s) { }call(“Device2”, “onFoo”, “bar”);

③ Request / Reply request ( instanceId, funcName, [reqArgs] ).timeout( msec ).receive( [repArgs] )

Device2Device1

SLOT(onFoo, int);

void onFoo(const int i) { reply( i + i ); }

int number;

request(“Device2”, “onFoo”, 21).timeout(100).receive(number);

Request

Notify

Notify

Reply


DETAIL: Data transportBroker based communication API – Four Patterns 9

④ Asynchronous Request / Reply requestNoWait ( req_instanceId, req_funcName, rec_instanceId, rec_funcName,

[reqArgs] )

Device2Device1

SLOT(onFoo, int);

void onFoo(const int i) { reply( i + i ); }requestNoWait(“Device2”, “onFoo”, “”, “onBar”, 21);

Request

Notify

Notify

Reply

SLOT(onBar, int);

onBar(const int i) { … }New


STATUS: Broker communication 10

Recent changes Fundamental change of how messages are consumed

Before: Each Slot presented an own consumer on the broker, forcing the broker

to route (using “Selectors”) messages by selecting on instanceId and

functionName. Larger installation caused huge number of consumer clients on

the broker and needed a thread per Slot on the device Now: Each object is a consumer on the broker, routing is done only utilizing the

instanceId only. Slot selection is done on the client side. Using an own queuing

system on the SignalSlotable the number of threads used per instance is

decoupled from the number of slots (can be single threaded context)

Heartbeats get first priority (using own topic and by placing on front of queue)

Future work Performance and scalability tests

Check whether trouble with heartbeats is finally solved

Open issues


Data transport – P2P 11

Another fundamental communication pattern between objects is realized by

connecting so-called input and output channels to form a direct point-to-point

connection (shortcutting the broker)

Unlike slots (which are functions) input and output channels are named objects

with a read/write/update API

The SignalSlotable API allows to create one or more such channels per

SignalSlotable instance

Technically output channels are (multi-client capable) TCP servers, input channels

are clients

The connection between them is established by referring to instanceId and

channelName instead of host and port.

Host and port are transparently communicated during connection time using the

broker based communication

Channels are highly configurable and are intended to serve the need of flexible

streaming data pipeline setups


DETAIL: Data transportP2P communication

One channel is specific for one data object (e.g. Hash, Image, Byte-Array)

For input and output channels within the same application data exchange will

happen by handing over pointers in memory instead of transmitting via TCP

Users can register two function call-backs indicating availability of data (e.g.

onData) and (optionally) the end of the data stream (e.g. onEndOfStream) on the

input channel

Input channels configure whether they share the sent data with all input channels

connected to the same output or whether they receive a copy of each data token

Output channels may specify a special hostname (in case of multiple adapters) to

which the clients are routed to

12

[…]P2P

Data

Message Broker

P2P control

New


STATUS: P2P communication 13

Recent changes Added possibility to select the interface on which to communicate

Future work More performance and scalability tests

Whilst reading is already asynchronous to the users-code execution (IO during

processing), for writing this is currently not true (was implemented but removed for

instability issues)

Asynchronous must again be implemented for performance improvement

Open issues Think carefully whether this communication could also be used (performance issue)

for transporting fast DAQ data from Device to DAQ-Layer


Devices (distributed end points) 14

The distributed end points follow the “Device Server Model” Similar to: TANGO or DOOCS

End points are controllable objects managed by a device server

Instance of such an object is a Device, with a hierarchical name

Device classes can be loaded at runtime (plugins)

Devices inherit SignalSlotable and wrap the communication API into a simpler

subset

Actions pertaining to a device given by its properties, commands, and channels i.e. get, set, monitor some property or execute some command write/read some data to/from a channel and update when done

Properties, commands and channels are statically described (expectedParameters

function) and further described via attributes in the device class. This description is

saved in form of a Schema. Dynamic (runtime) extension (Schema injection) of

expectedParameters is possible.

Devices can be written in either C++ or Python


DETAIL: DevicesConfiguration - API 15

Class: MotorDevicestatic expectedParameters( Schema& s ) { FLOAT_ELEMENT(s).key(“velocity”) .description(“Velocity of the motor”) .assignmentOptional().defaultValue(0.3) .maxInc(10) .minInc(0.01) .reconfigurable() .allowedStates(“Idle”) .commit();

INT32_ELEMENT(s).key(“currentPosition”) .description = “Current position of the motor” .readOnly() .warnLow(10) […]

SLOT_ELEMENT(s).key(“move”) .description = “Will move motor to target position” .allowedStates(“Idle”) […]}

// Constructor with initial configuration MotorDevice( const Hash& config ) { […] }

// Called at each (re-)configuration requestonReconfigure( const Hash& config ) { […] }

Any Device uses a standardized API to describe itself. This information is shipped as Schema object and used by interested clients (GUI, CLI other devices)

No need for device developers to validate any parameters. This is internally done taking the expectedParameters as white-list

We distinguish between properties and commands and associated attributes, all of them can be expressed within the expected parameters function

Properties and commands can be nested, such that hierarchical groupings are possible

Attribute

Property

Command


DETAIL: DevicesCreating a new device 16

plugins

1. Write a class (say: MyDevice) that derives

from Device

2. Compile it into a shared library (say

libMyDevice.so)

3. Select a running Device-Server or start a

new one

4. Copy the libMyDevice.so to the plugins

folder of the Device-Server

5. The Device-Server will emit a signal to the

broker that a new Device class is

available, it ships the expected parameters

as read from static context of the

MyDevice classGUI

libMy

Device.so

signalNewDeviceClassAvailable (.xsd)

GUI-Srv


DETAIL: DevicesCreating a new device 17

plugins

GUI

GUI-Srv

MyDevice1

factory: create(“MyDevice”, xml)6. Given the mask of possible parameters the

user may fill a valid configuration and emit

an instantiate signal to the broker

7. The configuration will be validated by the

Device factory and if valid, an instance of

MyDevice will be created

8. The constructor of the device class will be

called and provided with the configuration

9. The run method will be called which starts

the state-machine and finally blocks by

activating the event-loop

10. The device will asynchronously listen to

allowed events (slots)

signalInstantiate(“MyDevice”, xml)


Device “flavors”

Equipment

Control

DAQ

Equipment

with Data

Composite

Device

DAQ

Equipment

without Data

Workflow

Node

PCLayer

Node

Service

Device

e.g. motor, pump, valve, sensor

e.g. commercial camera e.g. digitizer, beam position monitor, 2D-detectors

e.g. calibrationManager, projectManager, brokerMonitor


DETAIL: DevicesDevices taking part in distributed system 19

HV Pump

Simulate

Store

Cali-

brate1

Cali-

brate2

Load

Digitizer

Logger

Disk

Storage

GUI

Server

GUI(s)

Terminal(s)

Camera

Device-Server

Application

Message Broker

(Event Loop)

Device Instance


STATUS: Devices 20

Recent changes

Future work Best practices and all concepts for hierarchical device structures must be defined

The composed-in DeviceClient API needs more functionality to make composition

easier

Open issues


States and state machines 21

OK

Initialization

Stopped

Started

none

start

errorFoundreset

stop

Start Stop State Machine

Error

Any property setting or command execution on a Device

can be restricted to a set of allowed states (using the

allowedStates attribute)

The state of a device can be changed by simply setting the

state property (string) to the desired value

The GUI is state and allowed states aware and

enables/disables buttons and properties pro-actively

Devices may optionally implement a finite state machine

(FSM) following the UML standard

In this case an incoming slot call is not directly

implemented but triggers and event into the state

machine

User defined hooks are executed as consequence of a

start-to-finish event processing algorithm.

Possible hooks are: guard, src-state-on-exit, transition-

action, tgt-state-on-entry, on-state-action

New

New


DETAIL: States and state machinesFinite state machines – There is a UML standard 22

State Machine: the life cycle of a thing. It is made of states, transitions and processes incoming events.

State: a stage in the life cycle of a state machine. A state (like a submachine) can have an entry and exit behaviors

Event: an incident provoking (or not) a reaction of the state machine

Transition: a specification of how a state machine reacts to an event. It specifies a source state, the event triggering the transition, the target state (which will become the newly active state if the transition is triggered), guard and actions

Action: an operation executed during the triggering of the transition

Guard: a boolean operation being able to prevent the triggering of a transition which would otherwise fire

Transition Table: representation of a state machine. A state machine diagram is a graphical, but incomplete representation of the same model. A transition table, on the other hand, is a complete representation


DETAIL: StatesFSM implementation example in C++ (header only) 23

// AllOkState MachineFSM_TABLE_BEGIN(AllOkStateTransitionTable)// SrcState Event TgtState Action GuardRow< StartedState, StopEvent, StoppedState, StopAction, none >,Row< StoppedState, StartEvent, StartedState, StartAction, none >FSM_TABLE_ENDFSM_STATE_MACHINE(AllOkState, AllOkStateTransitionTable, StoppedState, Self)

// EventsFSM_EVENT2(ErrorFoundEvent, onException, string, string)FSM_EVENT0(EndErrorEvent, endErrorEvent)FSM_EVENT0(StartEvent, slotMoveStartEvent)FSM_EVENT0(StopEvent, slotStopEvent)

// StatesFSM_STATE_EE(ErrorState, errorStateOnEntry, errorStateOnExit)FSM_STATE_E(InitializationState, initializationStateOnEntry)FSM_STATE_EE(StartedState, startedStateOnEntry, startedStateOnExit)FSM_STATE_EE(StoppedState, stoppedStateOnEntry, stoppedStateOnExit)// Transition ActionsFSM_ACTION0(StartAction, startAction)FSM_ACTION0(StopAction, stopAction)

// StartStop MachineFSM_TABLE_BEGIN(StartStopTransitionTable)Row< InitializationState, none, AllOkState, none, none >,Row< AllOkState, ErrorFoundEvent, ErrorState, ErrorFoundAction, none >,Row< ErrorState, EndErrorEvent, AllOkState, EndErrorAction, none >FSM_TABLE_ENDKARABO_FSM_STATE_MACHINE(StartStopMachine, StartStopMachineTransitionTable, InitializationState, Self)FSM_CREATE_MACHINE(StartStopMachine, m_fsm);FSM_SET_CONTEXT_TOP(this, m_fsm)FSM_SET_CONTEXT_SUB(this, m_fsm, AllOkState)FSM_START_MACHINE(m_fsm)

Transition table element

Regular callable function (triggers event)

Transition table element

Regular function hook (will be call-backed)


DETAIL: StatesFSM implementation example in Python 24

# AllOkState MachineallOkStt = [# SrcState Event TgtState Action Guard (‘StartedState’, ‘StartEvent’, ‘StoppedState’, ‘StartAction’, ‘none’), (‘StoppedState’, ‘StopEvent’, ‘StartedState’, ‘StopAction’, ‘none’)]FSM_STATE_MACHINE(‘AllOkState’, allOkStt, ‘InitializationState’)

# EventsFSM_EVENT2(self, ‘ErrorFoundEvent’, ‘onException’)FSM_EVENT0(self, ‘EndErrorEvent’, ‘slotEndError’)FSM_EVENT0(self, ‘StartEvent’, ‘slotStart’)FSM_EVENT0(self, ‘StopEvent’, ‘slotStop’)

# StatesFSM_STATE_EE(‘ErrorState’, self.errorStateOnEntry, self.errorStateOnExit )FSM_STATE_E( ‘InitializationState’, self.initializationStateOnEntry )FSM_STATE_EE(‘StartedState’, self.startedStateOnEntry, self.startedStateOnExit)FSM_STATE_EE(‘StoppedState’, self.stoppedStateOnEntry, self.stoppedStateOnExit)

# Transition ActionsFSM_ACTION0(‘StartAction’, self.startAction)FSM_ACTION0(‘StopAction’, self.stopAction)

# Top MachinetopStt = [ (‘InitializationState’, ‘none’, ‘AllOkState’, ‘none’, ‘none’), (‘AllOkState’, ‘ErrorFoundEvent’, ‘ErrorState’, ‘none’, ‘none’), (‘ErrorState’, ‘EndErrorEvent’, ‘AllOkState’, ‘none’, ‘none’)]FSM_STATE_MACHINE(‘StartStopDeviceMachine’, topStt, ‘AllOkState’)self.fsm = FSM_CREATE_MACHINE(‘StartStopMachine’)self.startStateMachine()


STATUS: States 25

Recent changes A hook for performing some (periodic) action whilst being in a state was added to

the FSM

A clean way of implementing devices without FSM is available (and is now

recommended)

Future work In case of no FSM: Device-side validation of command executions and property

settings against allowed states attribute

Open issues


Data logger

All property changes of all devices are archived centrally and in an event-

driven way The archive can be used to debug the system at a later point

The data logger allows fast retrieval of two kinds of information: Values of a property in a selected time range (feeding e.g. trend line plots in

GUI) The full configuration of a device at a given time point

By default all devices and all their properties are logged. However, entire

devices or individual properties of those may be flagged to be excluded from

logging

Logging is done in a per-device fashion and for any device currently 3 append

able text files are generated: *_configuration.txt: Stores all changes of the device properties *_schema.txt: Stores all changes of the device schema *_index.txt: Index file for speeding up queries

Changed


Data logger

Any regular device has a

DataLogger_<deviceName> companion

A DataLoggerManager composite

device couples the life-time of the two

companions

DataLogger

DeviceA

DataLogger

Manager

DataLogger

DeviceB

DeviceA DeviceB


STATUS: Data Logger 28

Recent changes A hook for performing some (periodic) action whilst being in a state was added to

the FSM

A clean way of implementing devices without FSM is available (and is now

recommended)

Future work Further scaling will be done by running DataLoggers on several hosts (connected to

a shared file system) as configured via the DataLoggerManager

A second device will be implemented that reads the generated files and

asynchronously populates a RDBMS

Open issues Is the data we log complete? Should not command executions also be part of the

logged data?


Data acquisition 29

via broker

direct TCP channels

Data aggregation, integration & dissemination

Multiple aggregator instances to handle all slow & fast data

Borrowed from Djelloul Boukhelef


Concept thoughts: DAQ integration 30

via broker

direct TCP channels

Data aggregation, integration & dissemination

Multiple aggregator instances to handle all slow & fast data

Borrowed from Djelloul Boukhelef

DAQ

Equipment

with Data

DAQ

Equipment

without Data

Equipment

Control

PCLayer

Node

Aggregator

Workflow

Node


STATUS: Data Acquisition integration 31

Burkhard Heisen (WP76)

Future work Think about the best way how to transport the data (which is send by Karabo

devices) to the DAQ layer

Open questions Requirements for sending data from Karabo devices to DAQ layer instead of

sending data between devices for workflow purposes are different

No “smartness” needed (like load balancing, multi-cast, etc.)

Writing to file can be done more generic, than further processing (what format

is the best)

Can we and should we try to use the same API and implementation for scientific

workflows and DAQ sinking?


Real time needs (where necessary) 32


Karabo itself does not provide real time

processes/communications

Real time processes (if needed) must be

defined and executed in layers below

Karabo. Karabo devices will only

start/stop/monitor real time processes

An example for a real-time system are the

Ethercat based solutions from the company

Beckhoff which we can interface to

Interlock/Supervisory code can be

implemented at either PLC (realtime) and

Karabo

BeckCom

Motor1 Motor2 Pump1

PLC-CPU

Motor1 Motor2 Pump1

TCP (own protocoll)

Gather/Scatter

Ethercat


Time synchronization (time stamps, cycle ids, etc.) 33


Concept: Any changed property will carry timing information as attribute(s) Time information is assigned per property

Karabo’s timestamp consists of the following information: Seconds since unix epoch, uint64 Fractional seconds (up to atto-second resolution), uint64 Train ID, uint64

Time information is assigned as early as possible (best: already on hardware) but

latest in the software device

On event-driven update, the device ships the property key, the property value and

associated time information as property attribute(s)

Real-time synchronization is not subject to Karabo

Correlation between control system (monitor) data and instrument data will be

done using the archived central DB information (or information previously

exported into HDF5 files)


DETAIL: Time synchronizationDistributed Train ID clock 34


Concept: A dedicated machine with a time receiver board

(h/w) distributes clocks on the Karabo level

Scenario 1: No time information from h/w Example: commercial cameras Timestamp is associated to the event-driven data in

the Karabo device If clock signal is too late, the next trainId is calculated

(extrapolated) given the previous one and the interval

between trainId's

The interval is configurable on the Clock device and

must be stable within a run. Error is flagged if clock

tick is lost.

Scenario 2: Time information is already provided by h/w The timestamp can be taken from the h/w or the

device (configurable). The rest is the same as in

scenario 1.

Clock

Device

Time receiver board

signals:

1. trainId

2. epochTime

3. interval

creates timestamp and associates to trainId


Central services - Name resolution/access 35


The only central service technically needed is the broker, others are optional Start-up issues

Any object connecting to the same broker (host/port/topic) must have a unique ID (string)

All communication objects will finally derive the SignalSlotable class which can be instantiated with a given ID (configured) or generates one if no ID is provided

If no instance ID is provided the ID is auto-generated locally Servers: hostname_Server_pid Devices: hostname-pid_classId_counter

Any instance ID is validated (by request-response trial) prior startup to be unique in the distributed system


DETAIL: Access levels We will initially have five access levels (enums) with intrinsic ordering

ADMIN = 4 EXPERT = 3 OPERATOR = 2 USER = 1 OBSERVER = 0

Any Device can restrict access globally or on a per-parameter basis Global restriction is enforced through the “visibility” property (base class)

Only if the requestor is of same or higher access level he can see/use the device The “visibility” property is part of the topology info (seen immediately by clients)

Parameter restriction is enforced through the “requiredAccessLevel” schema-attribute Parameter restriction typically is set programmatically but may be re-configured

at initialization time (or even runtime?) The “visibility” property might be re-configured if the requestors access level is higher

than the associated “requiredAccessLevel” (should typically be ADMIN) The default access level for settable properties and commands is USER The default access level for read-only properties is OBSERVER The default value for the visibility is OBSERVER

36



DETAIL: Access levels A role is defined in the DB and consists of a default access level and a device-

instance specific access list (overwriting the default level) which can be empty. SPB_Operator

defaultAccessLevel => USER accessList

SPB_* => OPERATOR Undulator_GapMover_0 => OPERATOR

Global_Observer defaultAccessLevel => OBSERVER

Global_Expert defaultAccessLevel = EXPERT

After authentication the DB computes the user specific access levels considering current time, current location and associated role. It then ships a default access and an access level list back to the user. If the authentication service (or DB) is not available, Karabo falls back to a

compiled default access level (in-house: OBSERVER, shipped-versions: ADMIN) For a ADMIN user it might be possible to temporarily (per session) change the

access list of another user.

37



DETAIL: Security 38


Header […]__uid=42__accessLevel=“admin”

Body […]

Broker-Message

Device

Locking:if is locked: if is __uid == owner then ok

Access control:if __accessLevel >= visibility: if __accessLevel >= param.accessLevel then ok

GUI-Srv

Central DB

1. Authorizes

2. Computes context based access levels

username

password

provider

ownIP*

brokerHost*

brokerPort*

brokerTopic*

userId

sessionToken

defaultAccessLevel

accessList

GUI or CLI


Statistics (control system itself, operation, …) 39


Concept: Statistics will be collected by regular devices OpenMQ implementation provides a wealth of statistics (e.g. messages in

system, average flow, number of consumers/producers, broker memory used…)

Have a (broker-)statistic device that does system calls to retrieve information

Similar idea for other statistical data


Logging (active, passive, central, local) 40


Concept: Categorized into the following classes Active Logging Additional code (inserted by the developer) accompanying the

production/business code, which is intended to increase the verbosity of what is currently

happening.

Code Tracing Macro based, no overhead if disabled, for low-level purposes

Code Logging Conceptual analog to Log4j, network appender, remote and at runtime

priority (re-)configuration

Passive Logging Recording of activities in the distributed event-driven system. No extra

coding is required from developers, passive logging transparently records system relevant

events. Broker-message logging Low-level debugging purpose, start/stop, not active during

production Transactional logging Archival of the full distributed state (see DataLogger)


Project

The project is an organizational structure for logically related devices The project does not describe:

Which device-server should run on what host Which plugin is loaded to what device-server

The project acts on top of existing (running) device-servers and loaded plugins

It describes initial configurations, runtime configurations, macros, scenes, monitors

and resources for a set of logically connected devices

Example projects could be: Detector_FXE Laser_FXE DAQ_FXE

Macros have an API to work with the project

Projects are associated to a user (can be a functional user)

The project itself is a set of files, it does not maintain a state (like “started” or “stopped”)

41


Project 42

Centralized project storing via a Karabo service device “ProjectManager” Implement in Python (code already exists in GUI code)

Analog to DataLoggerManager or CalibrationManager within Karabo Framework

Implement an output (loading project) and an input (saving project) channel

Allow multi-user (read) and single-user (write) access


Detail: Project file organization The project is saved as a zipped folder named <projectname>.krb

The folder contains a project.xml file with the following structure: <project>

<devices>[…]</devices> <macros>[…]</macros> <scenes>[…]</scenes> <monitors>[…]</monitors> <resources>[…]</resources>

</project>

And sub-folders containing files which are referenced by the above mentioned project.xml Devices

Containing <device>.xml files Macros

Containing <macro>.py files Scenes

Containing <scene>.svg files Resources

Containing any files (images, specific configurations, notes, etc.) etc.

43



STATUS: Project 44

Recent changes Logical grouping of devices of same class is possible

Groups allow multi-edit functionality (very useful for work-flow configurations)

Future work Central project handling must be implemented

The Monitors section must be implemented Monitors are a user defined collection of properties that will be associated to a experimental

run (or a control scan)

Open questions Current idea is to introduce another top hierarchy level -> a project group

Groups should be logical associations and only point/link to the physical projects

A project group could reflect all settings an experiment needs by aggregating all specialists

projects (like laser, detector, daq, experiment) with the user project

Experiment configurations could be started by copying a (template) group and then modifying it

by the individual experts until the specified setup is reached

Still not completely clear whether this approach will cover all needs for experiment control


Processing workflows (parallelism, pipeline execution, provenance) 45


Concept: Devices as modules of a scientific workflow system Configurable generic input/output channels on devices One channel is specific for one data structure (e.g. Hash, Image, File, etc.) New data structures can be “registered” and are immediately usable Input channel configuration: copy of connected output’s data or share the data with

other input channels, minimum number of data needed ComputeFsm as base class, developers just need to code the compute method IO system is decoupled from processing system (process whilst transferring data) Automatic (API transparent) data transfer optimization (pointer if local, TCP if remote) Broker-based communication for workflow coordination and meta-data sharing GUI integration to setup workflows graphically (drag-and-drop featured) Workflows can be stored and shared (following the general rules of data privacy and

security) executed, paused and stepped

Parallel execution


DETAIL: Processing workflowsParallelism and load-balancing by design 46


TCP

Memory

Devices within the same device-server: Data will be transferred by handing over pointers

to corresponding memory locations Multiple instances connected to one output

channel will run in parallel using CPU threads

Devices in different device-servers: Data will be transferred via TCP Multiple instances connected to one output

channel will perform distributed computing

CPU-threads

Distributed processing Output channel technically is TCP server, inputs are clients Data transfer model follows an event-driven poll architecture, leads to load-balancing

and maximum per module performance even on heterogeneous h/w Configurable output channel behavior in case no input currently available: throw, queue,

wait, drop


DETAIL: Processing workflowsGPU enabled processing 47


Concept: GPU parallelization will happen within a compute execution The data structures (e.g. image) are prepared for GPU parallelization Karabo will detect whether a given hardware is capable for GPU computing at runtime,

if not falls back to corresponding CPU algorithm Differences in runtime are balanced by the workflow system

IO whilst computing

Pixel parallel processing

(one GPU thread per pixel)Notification about new data possible to obtain

GPU

CPU


Clients / User interfaces (API, languages, macro writing, CLI, GUI) 48


Concept: Two UIs – graphical (GUI) and scriptable command line (CLI) GUI

Have one multi-purpose GUI system satisfying all needs See following slides for details

Non-GUI We distinguish APIs for programmatically set up of control sequences (others call

those Macros) versus and API which allows interactive, commandline-based control (IPython based)

The programmatic API exists for C++ and Python and features: Querying of distributed system topology (hosts, device-servers, devices, their

properties/commands, etc.): getServers, getDevices, getClasses instantiate, kill, set, execute (in “wait” or “noWait” fashion), get, monitorProperty,

monitorDevice Both APIs are state and access-role aware, caching mechanisms provide proper

Schema and synchronous (poll-feel API) although always event-driven in the back-end

The interactive API integrates auto-completion and improved interactive functionality suited to iPython


GUI: What do we have to deal with?

Client-Server (network protocol, optimizations)

User management (login/logout, load/save settings, access role support)

Layout (panels, full screen, docking/undocking)

Navigation (devices, configurations, data, …)

Configuration (initialization vs. runtime, loading/saving, …)

Customization (widget galleries, custom GUI builder, composition, …)

Notification (about alarms, finished pipelines, …)

Log Inspection (filtering, configuration of log-levels, …)

Embedded scripting (iPython, macro recording/playing)

Online documentation (embedded wiki, bug-tracing, …)

49

Kerstin Weger (WP76)


Client-Server (network protocol, optimizations) 50

Master

Central DB

GUI-Srv

Message

Broker

GUI-Client

I only see device “A”

onChange information only

related to “A”

Concept: One server, many clients, TCP Server knows what each client user sees (on a

device level) and optimizes traffic accordingly

Client-Server protocol is TCP, messages are

header/body style using Hash serialization (default

binary protocol)

Client side socket will be threaded to decouple from

main-event loop

On client start server provides current distributed

state utilizing the DB, later clients are updated

through the broker

Image data is pre-processed on server-side and

brought into QImage format before sending



User management (login/logout, load/save settings, access role support) 51

Concept: User centralized, login mandatory Login necessary to connect to system

Access role will be computed (context based)

User specific settings will be loaded from DB

View and control is adapted to access role

User or role specific configuration and wizards are

available

Central DB

1. Authorizes

2. Computes context based access role

username

password

userId

accessRole

session



Layout (panels, full screen, docking/undocking) 52

Six dock-able and slide-able (optionally tabbed) main panels Panels are organized by functionality

Navigation Custom composition area (sub-GUI building) Configuration (non-tabbed, changes view based on selection elsewhere) Documentation (linked and updated with current configuration view) Logging Notifications Project

Panels and their tabs can be undocked (windows then belongs to OS’s window

manager) and made full-screen (distribution across several monitors possible)

GUI behaves natively under MacOSX, Linux and Windows



Graphical interface - Overview 53

Burkhard Heisen (CAS Group)

drag & dropLive

Navigation Custom SceneConfiguration

ProjectLog Messages

Interactive Command LineDocumentation

Bug Reporting

User centric and access-controlled setup (login at startup)

Dock-able and resizable multi-panel, all-in-one user interface

Live navigation showing all device-servers, plugins, and device instances

Automatically generated configuration panel, allowing to read/write/execute

Project panel for persisting configurations, macros, scenes, resources, etc.

PowerPoint like, drag & droppable, tabbed custom scene

Centralized logging information, notification handling, documentation, etc.


Navigation (devices, configurations, data, …) 54

Concept: Navigate device-servers, devices,

configurations, data(-files), etc. Different views (tabs) on data

Hierarchical distributed system view Device ownership centric (view compositions) Hierarchical file view (e.g. HDF5)

Automatic (by access level) filtering of items

Auto select navigation item if context is selected

somewhere else in GUI



Configuration (initialization vs. runtime, loading/saving, …) 55

Concept: Auto-generated default widgets for

configuring classes and instances Widgets are generated from device information (.xsd

format)

2-column layout for class configuration (label,

initialization-value)

3-column layout (label, value-on-device, edit-value)

for instance configuration

Allows reading/writing properties (all data-types)

Allows executing commands (as buttons)

Is aware about device’s FSM, enables/disables

widgets accordingly

Is aware about access level, enables/disables

widgets accordingly

Single, selection and all apply capability



Customization (widget galleries, custom GUI builder, composition, …) 56

Concept: Combination of PowerPoint-like editor and online

properties/commands with changeable widget types Tabbed, static panel (does not change on navigation)

Two modes: Pre-configuration (classes) and runtime configuration (instances)

Visual composition of properties/commands of any devices

Visual composition of devices (workflow layouting)

Data-type aware widget factory for properties/commands (edit/display)

PowerPoint-like tools for drawing, arranging, grouping, selecting, zooming of text,

shapes, pictures, etc.

Capability to save/load custom panels, open several simultaneously



DETAIL: CustomizationProperty/Command composition 57

drag & drop

Display widget (Trend-Line)

Display widget

Editable widget



DETAIL: CustomizationProperty/Command composition 58

drag & drop

Display widget

(Image View)

Display widget

(Histogram)



DETAIL: CustomizationDevice (workflow) composition 59


drag & drop

Whole devices can be dragged

(from left side) as pipeline nodes

Dragging individual parameters

from right is still possible (e.g.

control parameters)

Devices can be grouped and

edited as group (connections

and configurations)

Distributed computing will

happen if different hosts are

involved

Display of per node or node-

group utilization


Macro editing and execution 60


Macro editing and execution from

within GUI possible

Macro parameters and functions

integrate automatically into

configuration panel

Macros are running within the

GUI’s event loop (direct widget

manipulation possible)

Macro API can be interactively

executed in embedded IPython

interpreter

Asynchronous operations use

Python 3’s coroutines and the

yield from keyword (extension

written allowing this for IPython)Courtesy of M. Teichmann


Notification (about alarms, finished runs, …) 61

Concept: Single place for all system relevant notifications, will link-out to more

detailed information Can be of arbitrary type, e.g.:

Finished experiment run/scan Finished analysis job Occurrences of errors, alarms Update notifications, etc.

Intended to be conceptually similar to now-a-days smartphone notification bars

Visibility and/or acknowledgment of notifications may be user and/or access role

specific

May implement some configurable forwarding system (SMS, email, etc.)



Log Inspection (filtering, configuration of log-levels, …) 62

Concept: Device’s network appenders provide active logging information which

can be inspected/filtered/exported Tabular view

Filtering by: full-text, date/time, message type, description

Export logging data to file

Logging events are decoupled from main event loop (threading)

Uses Qt’s model/view with SQLite DB as model (MVC design)



Online documentation (embedded wiki, bug-tracing, …) 63

Concept: Make the GUI a rich-client having embedded internet access. Use it

for web based device documentation, bug tracking, feature requests, etc. Any device class will have an individual (standardized) wiki page. Pages are

automatically loaded (within the documentation panel) as soon as any

property/command/device is selected elsewhere in GUI (identical to configuration panel

behavior). Depending on access role, pages are immediately readable/editable.

Device wiki pages are also readable/editable via European XFEL’s document

management system (Alfresco) using standard browsers

For each property/command the coded attributes (e.g. description, units, min/max

values, etc.) is shown.

European XFEL’s bug tracking system will be integrated



Software management (coding, building, packaging, deployment, versioning, …) 64


Concept: Spiced up NetBeans-based build system, software-bundle approach Clear splitting of Karabo-Framework (distributed system) from Karabo-Packages

(plugins, extensions)

Karabo-Framework (SVN: karabo/karaboFramework/trunk) Coding done using NetBeans (for c++ and python), Makefile based Contains: karabo-library (libkarabo.so), karabo-deviceserver, karabo-

brokermessagelogger, karabo-gui, and karabo-cli Karabo-library already contains python bindings (i.e. can be imported into python) Makefile target “package” creates self-extracting shell-script which can be installed

on a blank (supported) operating system and is immediately functional Embedded unit-testing, graphically integrated into NetBeans (c++ and python)

Karabo-Packages (SVN: karabo/karaboPackages/category/packageName/trunk) After installation of Karabo-Framework packages can be build SVN checkout of a package to any location and immediate make possible Everything needed to start a full distributed Karabo instance available in package A tool for package development is provided (templates, auto svn integration, etc.)


Software management - Tools

Continuous integration system using

Jenkins (nightly builds on different

platforms)

Jenkins automatically runs all unit-tests

for each build and tests execution of

binaries

Redmine for project management

(features, bugs, releases, versioning

integration)

Installation through software bundle

approach (all dependencies are

shipped), user does not need to compile

nor install any system packages

Deployment system for distributed

device-servers and their plugins

65

Burkhard Heisen (CAS Group)


DETAIL: Software managementThe four audiences and their requirements 66


Framework Developer SVN interaction, versioning, releases Code development using Netbeans/Visual Studio Addition of tests, easy addition of external dependencies Tools for packaging the software into either binary + header or source bundles Allow for being framework developer and package developer (see below) in one person at the

same time Package Developer

Flexible access to the Karabo framework ($HOME/.karabo encodes default location) Allow "one package - one software" project mode (each device project has its own versioning

cycle, individual Netbeans project) Standards for in-house development or XFEL developers need to be fullfilled: use parametrized

templates provided, development under Netbeans, use SVN, final code review Possibility to add further extern dependencies to the Karabo framework (see above)

System Integrator/Tester Simple installation of Karabo framework and selected Karabo packages as binaries Start broker, master, i.e. a full distributed system Flexible setup of device-servers + plugins, allow hot-fixes, sanity checks

XFEL-User/Operator Easy installation of pre-configured (binary framework + assortment of packages) karabo systems Run system (GUI, CLI)


DETAIL: Software managementUnit-testing 67


PythonC++


DETAIL: Software managementContinuous integration 68


Continuous Integration is a software development practice where members of a team integrate their work frequently, usually each person integrates at least daily - leading to multiple integrations per day. Each integration is verified by an automated build (including test) to detect integration errors as quickly as possible. [Wikipedia]

Required Features: Support for different build systems and different OS Automated builds – nightly builds Continuous builds – on demand, triggered by SVN commit Build matrix – different OS, compiler, compiler options Web interface – configuration, results Email notification Build output logging – easy access to output of build errors Reporting all changes from SVN since last successful build – easy trace of guilty developer Plugin for any virtualization product (VirtualBox, VMWare, etc.) Netbeans plugin for build triggering Easy uploading of build results (installation packages) to web repository

CI systems on the market: Hudson, CruiseControl, buildbot, TeamCity, Jenkins …


DETAIL: Software managementContinuous integration 69



Conclusions 70


XFEL.EU software will be designed to allow simple integration of existing algorithm/packages

The provided services focus on solving general problems like data-flow, configuration, project-tracking, logging, parallelization, visualization, provenance

The ultimate goal is to provide a homogenous software landscape to allow fast and simple crosstalk between all computing enabled categories (Control, DAQ, Data Management and Scientific Computing)

The distributed system is device-centric (not attribute-centric), devices inherently express functionality for communication, configuration and flow control


71

Thank you for your kind attention.


Karabo: The European XFEL software framework Design Concepts

Documents

Transcript of Karabo: The European XFEL software framework Design Concepts