Online reconstruction overview¶
- Table of contents
- Online reconstruction overview
Introduction¶
The MAUS framework executes four types of component. Each component creates or manipulates spills. In this context, a spill is a JSON document representing information about particles.
- Inputters - these input or generate spills e.g. read live DAQ data and convert into spills, read archived DAQ data and convert into spills, or generate simulated spill data.
- Transforms - these analyse the data within the spill in various ways, derive new data and add this to the spill.
- Mergers - these summarise data within a series of successive spills and output new spills with the summarised data.
- Outputters - these take in merged spills and output spills, for example saving them as JSON files or, where applicable, extracting image data from them and saving this as image files.
Basically, the framework repeats the following loop until there are no more spills:
- Reads a spill from an inputter.
- Transforms the spill.
- Passes the transformed spill to a merger.
- Passes a merged spill to an outputter.
Parallel transformation of spills¶
As each spill is independent, spills can be transformed in parallel. Celery, a distributed asynchronous task queue for Python is used to implement parallel transformation of spills. Celery uses RabbitMQ as a message broker to handle communications between clients and Celery worker nodes.
A Celery worker executes transforms. When a worker starts up it registers with RabbitMQ - the RabbitMQ task broker can be local or remote. By default Celery spawns N sub-processes where N is the number of CPUs on a node, though N can be explicitly set by a user. Each sub-process can execute one transform at a time. Multiple Celery workers can be deployed, each of which with one or more sub-processes. Celery allows MAUS to execute highly-parallelised transforming of spills.
For MAUS, each Celery worker needs a complete MAUS deployment running on the same node as the worker. The MAUS distributed execution framework, configures Celery as follows:
- The framework gets the names of the transforms the user wants to apply e.g.
MapPyGroup(MapPyBeamMaker, MapCppSimulation, MapCppTrackerDigitization)
. This is termed a transform specification. - A Celery broadcast is invoked, passing the transform specification, the MAUS configuration and a configuration ID (e.g. the client's process ID).
- Celery broadcasts are received by all Celery workers registered with the RabbitMQ message broker.
- On receipt of the broadcast, each Celery worker:
- Checks that the framework's MAUS version is the same as the worker's. If not then an error is returned to the client.
- Forces the transform specification down to each sub-process.
- Waits for the sub-processes to confirm receipt.
- If all sub-processes update correctly then a success message is returned to the framework.
- If any sub-process fails to update then a failure message, with details, is returned to the framework.
- Each Celery sub-process:
- Invokes
death
on the existing transforms, to allow for clean-up to be done. - Updates their configuration.
- Creates new transforms as specified in the transform configuration.
- Invokes
birth
on these with the new configuration. - Confirms with the Celery worker that the update has been done.
- Invokes
- Celery workers and sub-processes catch any exceptions they can to avoid the sub-processes or, more seriously, the Celery worker itself from crashing in an unexpected way.
MAUS uses Celery to transform spills as follows:
- The framework gets the next spill from its input.
- A Celery client-side proxy is used to submit the spill to Celery. It gets an object which it can use to poll the status of the "job".
- The client-side proxy forwards the spill to RabbitMQ.
- RabbitMQ forwards this to an available Celery worker. If none are available then the job is queued.
- The Celery worker picks an available sub-process.
- The sub-process executes the current transform on the spill.
- The result spill is returned to the Celery worker and there back to RabbitMQ.
- The framework regularly polls the status of the transform job until it's status is successful, in which case the result spill is available, or failed, in which case the error is recorded but execution continues.
Document-oriented database¶
After spills have been transformed, a document-oriented database, MongoDB, is used by the framework to store the transformed spills. This database represents the interface between the input-transform and merge-output phases of a spill processing workflow.
The framework is given the name of a collection of spills and reads these in order of the dates and times they were added to the database. It passes each spill to a merger and then takes the output of the merger and passes it to an outputter.
Use of a database allows the input-transform part of a workflow to be separate from the merge-output part. It also allows them to operate in concurrently - one process can input and transform spills, another can merge transformed spills and output the merged results. This also allows many merge-output workflows to use the same transformed data, for example to generate multiple types of histogram from the same data.
Histogram mergers¶
Histogram mergers take in spills and, from the data within the spills, update histograms. They regularly output one or more histograms (either on a spill-by-spill basis or every N spills, where N is configurable). The histogram is output in the form of a JSON document which includes:
- A list of keywords.
- A description of the histogram.
- A tag which can be used to name a file when the histogram is saved. The tags can also be auto-numbered if the user wants.
- An image type e.g. EPS, PNG, JPG, or PDF. The image type is selected by the user.
- The image data itself in a base64-encoded format.
Histogram mergers do not display or save the histograms. That is the responsibility of other components.
Example histogram mergers, and generic super-classes to build these, currently exist for histograms drawn using PyROOT (ReducePyTOFPlot and @ReducePyROOTHistogram
) and matplotlib (ReducePyHistogramTDCADCCounts
and ReducePyMatplotlibHistogram
).
Saving images¶
An outputter (OutputPyImage
) allows the JSON documents output by histogram mergers to be saved. The user can specify the directory where the images are saved and a file name prefix for the files. The tag in the JSON document is also used to create a file name.
The outputter extracts the base-64 encoded image data, unencodes it and saves it in a file. It also saves the JSON document (minus the image data) in an associated meta-data file.
Web front-end¶
The web front-end allows histogram images to be viewed. The web front-end is implemented in Django, a Python web framework. Django ships with its own lightweight web server or can be run under Apache web server.
The web front-end serves up histogram images from a directory and supports keyword-based searches for images whose file names contain those key words.
The web pages dynamically refresh so updated images deposited into the image directory can be automatically presented to users.
The interface between the online reconstruction framework and the web front-end is just a set of image files and their accomanying JSON meta-data documents (though the web front-end can also render images without any accopanying JSON meta-data).
Design details¶
Run numbers¶
Each spill will be part of a run and have an associated run number. Run numbers are assumed to be as follows:
- -N : Monte Carlo simulation of run N
- 0 : pure Monte Carlo simulation
- +N : run N
Transforming spills from an input stream (Input-Transform)¶
This is the algorithm used to transform spills from an input:
CLEAR document store run_number = None WHILE an input spill is available GET next spill IF spill does not have a run number # Assume pure MC spill_run_number = 0 IF (spill_run_number != run_number) # We've changed run. IF spill is NOT a start_of_run spill WARN user of missing start_of_run spill WAIT for current Celery tasks to complete WRITE result spills to document store run_number = spill_run_number CONFIGURE Celery by DEATHing current transforms and BIRTHing new transforms TRANSFORM spill using Celery WRITE result spill to document store DEATH Celery worker transforms
If there is no initial
start_of_run
spill (or no spill_num
in the spills) in the input stream (as can occur when using simple_histogram_example.py
or simulate_mice.py
) then spill_run_number
will be 0
, run_number
will be None
and a Celery configuration will be done before the first spill needs to be transformed.
Spills are inserted into the document store in the order of their return from Celery workers. This may not be in synch with the order in which they were originally read from the input stream.
Merging spills and passing results to an output stream (Merge-Output)¶
This is the algorithm used to merge spills and pass the results to an output stream:
run_number = None end_of_run = None is_birthed = FALSE last_time = 01/01/1970 WHILE TRUE READ spills added since last time from document store IF spill IS "end_of_run" end_of_run = spill IF spill_run_number != run_number IF is_birthed IF end_of_run == None end_of_run = {"daq_event_type":"end_of_run", "run_num":run_number} Send end_of_run to merger DEATH merger and outputter BIRTH merger and outputter run_number = spill_run_number end_of_run = None is_birthed = TRUE MERGE and OUTPUT spill Send END_OF_RUN block to merger DEATH merger and outputter
The Input-Transform policy of waiting for the processing of spills from a run to complete before starting processing spills from a new run means that all spills from run N-1 are guaranteed to have a time stamp < spills from run N.
is_birthed
is used to ensure that there is no BIRTH-DEATH-BIRTH redundancy on receipt of the first spill from the document store.
Document store¶
Spills are stored in documents in a collection in the document store.
Documents are of form {"_id":ID, "date":DATE, "doc":SPILL}
where:
- ID: index of this document in the chain of those successfuly transformed. It has no significance beyond being unique in an execution of the Input-Transform loop which deposits the spill. It is not equal to the spill_num (Python
string
) - DATE: date and time to the milli-second noting when the document was added. A Python
timestamp
. - DOC: spill document. A Python
string
holding a valid JSON document.
Collection names¶
For Input-Transform,
- If configuration parameter
doc_collection_name
isNone
,""
, orauto
thenHOSTNAME_PID
, whereHOSTNAME
is the machine name andPID
the process ID, is used. - Otherwise the value of
doc_collection_name
is used. doc_collection_name
has default valuespills
.
For Merge-Output,
- If configuration parameter
doc_collection_name
isNone
,""
, or undefined then an error is raised. - Otherwise the value of
doc_collection_name
is used.
Miscellaneous comments¶
- Currently Celery timeouts are not used, transforming a spill takes as long as it takes.
- Celery task retries on failure option is not used. If the transformation of a spill fails first time it can't be expected to succeed on a retry.
- If memory leaks arise, e.g. from C++ code, look at Celery rate limitss, which allow the time or number of tasks before sub-process is killed and respawned, to be defined. Soft rate limits would allow
death
to be run on the transforms first.
Updated by Rogers, Chris over 10 years ago ยท 14 revisions