The pilot project "Tools and methods of
visual analytics as part of the workload management system (WMS) ProdSys2/PanDA
of the ATLAS experiment at the Large Hadron Collider (LHC)" is aimed at
developing state-of-the-art visualization and analytical approaches, and tools
for monitoring distributed WMS using the example of data processing and
analysis system of the ATLAS experiment [1] at LHC [2] (CERN, Switzerland).
Initial motivation of the project is associated with the experiments at LHC,
but both quantitative and qualitative requirements are common for many
experiments in high energy and nuclear physics (HENP). Therefore this project
is of interest to a wide range of scientific groups, as the discussion under of
the 26th Symposium on Nuclear Electronics and Computing (NEC2017) has shown
[3].
Internal information
(such as, description of format and every stage of processing of data from
physical experiments) of distributed data processing systems should be
processed, analyzed and presented to users of the system in a suitable and
compact form. For this purpose, specialized monitoring systems are being
developed. The paper [4], also published in this issue of the journal,
describes the monitoring system currently existing in the ATLAS experiment -
BigPanDA. The system includes following possibilities of data visualization:
interactive interfaces, parametric tables and various charts (histograms, bar
and pie charts, simple two-dimensional graphics). Until now, the requirements
for the monitoring system have been limited to the use of basic visual analysis
methods for a limited class of tasks, and data dimensionality limited to three
dimensions. However, the constant increase in the amount of processed data and
the complexity of the computing infrastructure of WMS of the ATLAS experiment,
as will be shown below, produces new challenges related to the visual analysis
of large volumes of multidimensional data.
Under this project it is proposed to apply a
fundamentally new approach to monitoring the operation of complex distributed
systems, and to change the "classical" monitoring of WMS
(particularly, in the field of HENP) using visual analysis methods [5,6] to
improve the quality of metrics for data state definition. The application of
these methods will allow to remove the restriction on the number of dimensions
of the analyzed data, providing the generation of multidimensional geometric
interpretations for visual analysis. As a result, the functionality of the
monitoring system will be expanded by the ability to detect explicit
correlations between different data objects and the use of multi-level
interactive interfaces. In addition, the joint application of visual analytics
and "machine learning" methods will increase the level of automation
in the functioning of data processing and analysis systems primarily in the
field of particle physics.
The scientific research cycle on modern
installations can lasts for decades (for example, scientific communities on the
LHC were established more than 25 years ago, the data collection was started 9
years ago and will last at least another 10 years). During the lifetime of the
experiments, the installations and detectors are modernizing, information
technologies are qualitatively changing. At the same time, the software and
hardware infrastructure as well as the data processing model (computing model)
are evolving: the number of data centers is constantly increasing (new
architectural solutions are emerging), the scenarios for launching and
performing analysis, data processing and modeling tasks are changing, software
versions are being updated, the technologies of storage and access to data and
metadata becomes different. Under a constantly evolving and complex computing
infrastructure as well as a simultaneous growth of information flow, it becomes
difficult to control the operating of data management systems, organize data
processing and simulation of the experiment, predict and timely detect possible
anomalies in the functioning of individual hardware and software components of
the computing infrastructure and data processing systems. To solve these
problems specialized analytical tools are being developed in modern
experiments. A separate crucial task is to search for patterns or analysis of
anomalies in the operating of complex distributed systems, correlations between
the actions of the operator's service and the behavior of the system,
prediction of system performance in case of software changes.
The second generation of the ProdSys2
(Production System) [7] of the ATLAS experiment, in conjunction with the
workload management system PanDA [8] (Production and Distributed Analysis
system), that is a complex set of hardware and software tools for organizing,
planning, starting and executing computing tasks and jobs (Figure 1).
ProdSys2/PanDA is responsible for all stages of data processing, analysis and
modeling, including simulating of physical processes and functioning of the
detector using the Monte Carlo method, (re)processing of physical data,
performing highly specialized tasks (e.g., setting a high-level trigger or a
software quality assurance task). Using the ProdSys2/PanDA software, the ATLAS
scientific community, individual physical groups and scientists have access to
hundreds of WLCG [9] (Worldwide LHC Computing Grid) computing centers,
supercomputers, cloud computing resources and university clusters.
Characteristics of the system can be reflected by the following indicators:
implementation of more than a million computing tasks per day at 200+ computer
centers by thousands of users utilizing more than 300,000 nodes.
Fig.1.
Data processing workflow at the ATLAS experiment
In a distributed system, there is always a
competition between different threads of computing tasks. For example, in the
period preceding the main physical conferences, the number of data analysis
tasks distinctly increases (from Figure 2 it follows that in the ATLAS
experiment there was a sharp increase in the number of computing jobs performed
in certain months). Therefore, when starting computing jobs, there can be
significant delays due to the lack of a free computing resource. As a rule, the
end user is interested not so much in the process of computing jobs execution
as in the ability to predict the data processing (or analysis) completion time
and get a scientific result in a predetermined time period (it can be hours for
analysis tasks, or weeks for data processing). The tasks flow itself has
several phases of execution (e.g., modeling, digitization, reconstruction,
creation of objects for physical analysis, physical analysis itself) and each
stage (phase) can be performed on geographically distributed computing centers,
which includes the transfer of initial input data between computing centers and
can affect the total execution time of the entire computing tasks flow.
Possible hardware malfunctions may require redistribution of tasks between
computing centers, that also provides additional ambiguity in the prediction of
the data processing metrics. At the moment there is no central portal for
monitoring, the operator is forced to view data transfer schedules, task
performing schedules, tables with information about the computing centers and
individual components functioning. The creation of a single portal and the
ability to visualize the operation of systems will allow to optimize the
utilization of a computing resource, significantly automate and simplify the
workflow of data processing, and thereby accelerate the delivering of
scientific results.
Fig.2.
Number of completed computing jobs grouped by computing “regions” during 2017
(average per week)
Information that was accumulated over the
entire period of operation of WMS of the ATLAS experiment (this is more than 14
years) contains records of the progress of execution of more than 10 millions
of computing tasks and about 3,000 millions of computing jobs. Based on such
statistics it allows to use "machine learning" methods to make
analytical calculations and to forecast the software functioning. The work
related to ProdSys2/PanDA, in which the development and usage of "machine
learning" methods for the analysis of processing data [10,11] were
started, showed that the calculated metrics based on the predictive analytics
allow to increase the efficiency of the processing of real and simulated data,
providing more thorough planning of the analysis process (determined by
individual users or a group of users), predicting possible failure or abnormal
behavior of the system (produced by agents of control services).
In the modern world, problems related to
processing and analysis of multidimensional data are among the most urgent
tasks. Many different methods and hardware-software tools are developed to
solve these tasks, that also includes both automatic and interactive solutions.
It is worth noting that data visualization can be used within these methods.
Currently, the approach of visual analytics is widely used. This approach was
preceded by solutions for solving tasks of multidimensional data analysis by a
variety of visual methods.
A review of the literature describing specific
applications using visual methods makes it possible to assert that, in reality,
interactive systems with multidimensional data are often given less importance
than systems for displaying the results of applying data analysis methods.
Examples include systems such as the situational awareness system AdAware [12],
the visual analysis system in aircraft engineering [13], the SAS Visual
Analytics software package [14], designed to process and analyze large amounts
of business data. Experience has shown that classical methods of parallel
coordinates, Andrews curves, Chernoff faces and other similar mnemonic
graphical representations are widely used for visual representation of
multidimensional data. All these visual methods are based on the fact that the
analyzed tuples of numerical data are interpreted as the values
of the parameters of such mnemonic graphical maps. Examples of
such mappings are shown in Figures 3,4.
Fig.3.
Fisher's Iris data set presented in the form of the Andrews curves [15]
Fig.4.
Chernoff faces for medical data [16]
It should be noted that all systems using the
above visual methods are essentially configured to internal processing of
multidimensional data and presentation of it to analysts in a convenient form.
They do not provide an opportunity to work with the analyzed data directly and
their multivariate geometric interpretations using corresponding visual
representations that are natural for human. The approach of visual analytics
involves solving data analysis tasks, and, in particular, tasks of
multidimensional data analysis, using a conducive interactive visual interface.
One of the most common forms of visual
analytics is the solution for tasks of multidimensional data analysis by the
visualization method [17]. The solution of the task of the initial data
analysis by the visualization method consists of the sequential solution of the
following two subtasks (Figure 5).
Fig.5.
Data analysis using the visualization method
The first subtask is to get a representation
of the analyzed data in the form of a certain graphic image (the visualization
of the original data), which is solved using a computer. The resulting graphic
images serve as a natural and convenient means to represent the spatial
interpretation of the initial data to the person (analyst). Spatial
interpretation is one or more spatial objects (i.e., the spatial scene), which
are set in compliance to the analyzed data. The second subtask, which is no
less important, is the visual analysis of the graphical representation of the
analyzed data obtained as a result of solving the first subtask, while the
analysis results are interpreted with respect to the original data, and which
is solved directly by the analyst. The spatial scene is visually analyzed using
the enormous potential possibilities of the analyst's spatial and conception
thinking during the analysis. As a result of solving this problem, the analyst
makes some judgments about the spatial scene. Thus, judgments about the object
under consideration are formulated. The process of visual analysis of the
graphic image is not strictly formalized. The efficiency of visual analysis is
determined by the experience of the person who carries out this image analysis
and his propensity for spatial and conception thinking. Looking at the
resulting image, a person is able to solve 3 main tasks: analysis of the shape
of spatial objects, analysis of their relative positions and analysis of
graphic attributes of spatial objects.
Under this project, it is proposed to use the
multidimensional geometric modeling of the initial data, which are considered
as multidimensional tabular data about computing jobs (Table 1).
Table 1. Representation of computing jobs (multidimensional tabular
data)
A geometric interpretation is carried out to
solve the defined task. Rows of the table correspond to multidimensional points
in the space En, 
, and the values
of the computing job parameters are the coordinates of
multidimensional points. It is suggested to interpret the difference in
parameter strings as the Euclidean distance between the points of this
multidimensional space (the longer the distance is, then the lines are more
different). With this interpretation, the analysis of the distance between
points of the N-dimensional space is assigned to the analysis task of
similarities and differences in records of computing jobs.
It is proposed to use a visual presentation of
points in the N-dimensional space to analyze the distance between these points.
At the beginning, the original set of points is projected onto one of the
three-dimensional spaces. Wherein:
•
The multidimensional point pi is projected into the
sphere Si.
•
If the distance between the points of the
N-dimensional space p1 and p2 is less than the threshold distance d given by
the analyst in an interactive mode, then a cylinder is constructed to connect
the spheres S1 and S2.
•
The color of the cylinder simulates the distance
between the points p1 and p2 from red (small distance) to blue (long distance).
Then, the graphic projection of spheres and
cylinders onto the picture plane is performed, and followed by its
corresponding visual analysis. The resulting set of spheres and cylinders forms
a spatial scene with a given geometry and optical (color) characteristics.
Thus, visual analysis of the spatial scene
will allow to judge the distance between the original multidimensional points.
In the process of solving the analysis task, it is proposed to set the initial
large value of d, and then reduce it and select subsets of multidimensional
points, depending on the resulting image in the picture plane. It should be
noted that under this approach, the analyst in the process of data analysis
does not passively contemplate the spatial scene, but has the possibility of
interactive engagement with it.
Interpretation of the parameters of data
objects and corresponding metrics in the N-dimensional space (the dimension is
determined by the number of parameters to be investigated) to evaluate the
mutual influence and determine their correlations.
Interpretation of data in the N-dimensional
space (dimensionality is determined by the number of parameters to be
investigated) and projection into 3-dimensional space will allow clustering of
objects (by specified parameters and metrics) in order to detect objects with
non typical set of values of specified parameters (i.e.,
anomalous objects). In the case of distributed workflow management system, such
data objects are computing tasks and jobs (computing task consists of a set of
computing jobs) that describe the processing of data in the corresponding
specialized systems (ProdSys2 and PanDA). The collection of the necessary
information about data processing (object parameters and metrics) is performed
during the definition of computing tasks and jobs, and their subsequent
execution (it is possible to take into account parameters and metrics which
describe the current state of the computing environment and the computing resources
used).
Determination of the sequence of stages in
which there was a failure, a delay or an error during the data processing, and
the identification of a set of values of the specified parameters in the
initial stage that caused the failure (based on the methods developed in
section 4.1).
The popularity of data is determined by the
number of accesses of analysis tasks to datasets (i.e., data objects) and the
number of requests for additional dataset replicas/copies (at different
computer centers). When the number of accesses to datasets is increased, then
the additional data replicas should be created automatically, and that replicas
should be “tied” geographically to the computer center with the highest demand.
The studies carried out based on data from the
monitoring system showed that the "popularity" of data is
dramatically reduced after about 45 days, which allows to make a decision to
delete replicas of not used datasets from disks and to transfer them to tape.
Visual analytics methods will allow to cluster
the data by geographical location of storage (computing center) and their
demand for analysis depending on time. This will optimize the requirements for
creation/deletion of additional/redundant data replicas.
Development of visualization tools and
methodology of using them for cluster analysis of tuples of parameters of
computing jobs of the PanDA system. Visual analysis will allow to identify
similar computing jobs, as well as to reveal anomalous computing jobs, and at
the same time determining parameters that caused this anomaly.
Development of visualization tools and
methodology of using them to analyze execution durations of computing jobs of
the PanDA system. Visual analysis will allow to determine the influence of a
certain set of parameters (i.e., basic set of parameters, which can be extended
in the future) on the execution time of computing jobs, and the identification
of a set/range of values of individual parameters indicating an
increase in the execution time of computing jobs (i.e., the deviation from the
average time, it affects tasks with execution time exceeding 
, it is assumed that this is
about 7.5% of the number of all tasks).
Development of visualization tools and
methodology of using them to analyze the popularity of data in relation to
time. Visual analysis will reveal the increase/decrease in the number of
accesses to data as a function of time, thus providing a convenient visual
method for decision making process of the dynamic management of the data
replicas distribution.
Creation of projection graphic images of the
N-dimensional geometric interpretations of parameters of computing jobs, 2D
histogram representation of the number of computing jobs (Y axis) grouped by
their execution durations (X axis), and identification of a group of computing
jobs with "increased" execution duration (more than average).
To solve the defined research task, key
parameters of computing jobs are defined:
•
Duration of the computing job execution
(<duration> = endtime - starttime)
•
Name of the job flow-group (gShare)
•
Computing center for the job processing (nucleus)
•
Number of events to be processed (nEvents)
•
Additional set of parameters (extended set of
parameters for the basic set):
o
The name of the analysis/processing stage
(processingType); the amount of input data for the job (inputFileBytes); the
type of input data (inputFileType); the amount of output data of the job
(outputFileBytes); the initial priority of the job (assignedPriority); CPU time
to process one event (cpuTimePerEvent); the hardware architecture on which
calculations are performed for the job (cmtConfig); the number of cores
(actualCoreCount); software release (atlasRelease); CPU efficiency per core
(CPU eff per core); the average size of memory pages allocated to the process
by the operating system and currently located in RAM (avgRSS); the average
fraction of shared memory used by the CPU (avgPSS); the average size of the
allocated virtual memory (avgVMEM); the maximum size of memory pages allocated
to CPU by the operating system (maxRSS); the maximum share of the total memory
used by CPU (maxPSS); the maximum size of the allocated virtual memory
(maxVMEM)
•
Indication parameters:
o
Step of restarting the job (attemptNr); error codes
(brokerageErrorCode, ddmErrorCode, exeErrorCode, jobDispatcherErrorCode,
pilotErrorCode, supErrorCode, taskBufferErrorCode)
The format of the raw data representation:
•
Combination of job parameters into groups
•
The representation of input data in the form of
matrices corresponding to a group of parameters, where rows correspond to job
records, and columns correspond to the parameters of a certain group:
o
- matrix with jobs’
durations, where n - the number of jobs;
o
- matrix with the basic set
of job parameters (gShare, nucleus, nEvents);
o
- matrix with the
additional/extra set of parameters;
o
- matrix with the set of
indication job parameters (attemptNr, errorCodes).
Data source:
•
Infrastructure ElasticSearch at the University of
Chicago [18]
•
Indices "jobs_archive_*"
o
Search conditions
§
Acceptable statuses of jobs: jobStatus IN
("finished", "failed");
§
Source of jobs: prodSourceLabel = 'managed'
§
Type of processing data and the stage of
processing: REGEXP_LIKE (jobName, “^mc(.*\.){3}simul\..*”)
This research task implies an extension of the
task/problem in section 5.1 applied to computing tasks of ProdSys2 and the
solution proceeds from the results of the research task given in 5.1, implies a
similar approach, but taking into account the specifics of the data objects
that are under consideration - computing tasks, on the basis of which sets of
computing jobs are formed.
1. Approbation of the approach of visual analytics using the example of
applying the cluster analysis of tuples of the computing jobs parameters of
PanDA system and estimating the distribution of computing jobs’ durations.
2. Expansion of the developed approach with respect to computing tasks of
ProdSys2.
3. Integration of the developed prototypes of visualization and analytical
tools into the monitoring infrastructure of the ProdSys2/PanDA system.
4. Evaluation of the modification of the existing ProdSys2/PanDA
monitoring using the visual analytics approach.
As a result of the project, a visual analytics
system will be developed to monitor workflow management systems. The developed
system will be an extended analytical service of the existing ATLAS monitoring
system. By means of the developed system, the monitoring functionality will be
significantly expanded, allowing to simulate, predict the further course and
state of the experiment. Visual analytics will form the basis of a decision
support system and strategic planning.
Cooperation with the ATLAS experiment at LHC,
availability of access to experimental data and demonstration of the created
solution and prototype on an existing data processing system, will provide a
unique testing ground for developing analytical research technologies and
application of visual analytics methods and will allow this project to be among
the world's most important developments of the given area.
The results of the project will be in demand
for the creation of the software for the NICA collider (JINR, Dubna), for the
high-luminosity LHC (HL-LHC), and for visualization of scientific information
on mega-facilities such as XFEL and FAIR.
The ATLAS collaboration in joint with Russian
research centers and universities participate in this pilot project.
List of research
centers and universities:
•
National Research Nuclear University
"MEPhI"
o
Laboratory of Scientific Visualization
o
Department of Analysis of Competitive systems
o
MEPhI group in the ATLAS experiment
•
National Research Center "Kurchatov
Institute"
o
Laboratory of Big Data Technologies
•
National Research Tomsk Polytechnic University
•
Joint Institute for Nuclear Research
o
Laboratory of Information Technologies
•
Brookhaven National Laboratory
•
University of Iowa
•
University of Chicago
•
University of Texas at Arlington
•
European Organization for Nuclear Research (CERN)
One of the objectives of the project is to
work with students, including the preparation of bachelors, masters and
graduate students who manage to use advanced tools for scientific visualization
and work with data from a physical experiment. The courses "Visual
Analytics" and "Scientific Visualization" are taught at the NRNU
MEPhI, on the basis of the results of the project, special courses for graduate
students with major in particle and nuclear physics, and system engineering
will be created. In addition, the University of Dubna and the Institute of
Cybernetics of TPU expressed interest in creating joint courses on the subject
of the project (the University of Dubna has created a course for
training/educating specialists for work at the NICA collider. TPU actively
participates in the scientific program in the field of particle physics: the
COMPASS experiment on a super-proton synchrotron (SPS, CERN), and ATLAS and CMS
experiments at LHC).
Information support for the project is
provided by the journal "Scientific Visualization" [19], as well as
by the informational portals of the ATLAS experiment at CERN [20], the
Laboratory of Big Data Technologies NRC KI [21] and LIT JINR [22].
1. The ATLAS Collaboration, “The ATLAS Experiment at the CERN Large Hadron
Collider”, Journal of Instrumentation, vol. 3, S08003, 2008.
2. LHC - Large Hadron Collider, http://lhc.web.cern.ch/lhc
3. 26th International Symposium on Nuclear Electronics & Computing -
NEC’2017, http://indico.jinr.ru/conferenceDisplay.py?confId=151
4. S.Padolski, T.Korchuganova, T.Wenaus, M.Grigorieva, A.Alexeev, M.Titov,
A.Klimentov, "Data visualization and representation in ATLAS BigPanDA
monitoring", Scientific Visualization, 2018.
5. J.Thomas, K.Cook, "Illuminating the Path: The Research and
Development Agenda for Visual Analytics", IEEE Computer Society, 2005.
6. D.Popov, I.Milman, V.Pilyugin, A.Pasko, "Visual Analytics of
Multidimensional Dynamic Data with a Financial Case Study", Data Analytics
and Management in Data Intensive Domains, Springer International Publishing,
pp. 237--247, 2017.
7. M.Borodin, K.De, J.Garcia Navarro, D.Golubkov, A.Klimentov, T.Maeno,
A.Vaniachine, “Scaling up ATLAS production system for the LHC Run 2 and beyond
: project ProdSys2”, Journal of Physics: Conference Series, vol. 664, no. 6,
2015.
8. A.Klimentov et al., "Migration of ATLAS PanDA to CERN",
Journal of Physics: Conference Series, vol. 219, no. 6, 2010.
9. WLCG - Worldwide LHC Computing Grid, http://wlcg.web.cern.ch
10. M.Titov,
G.Zaruba, A.Klimentov, and K.De, “A probabilistic analysis of data popularity
in ATLAS data caching”, Journal of Physics: Conference Series, vol. 396, no. 3,
2012.
11. M.Titov,
M.Gubin, A.Klimentov, F.Barreiro, M.Borodin, D.Golubkov, "Predictive analytics
as an essential mechanism for situational awareness at the ATLAS Production
System", The 26th International Symposium on Nuclear Electronics and
Computing (NEC), CEUR Workshop Proceedings, vol. 2023, pp. 61--67, 2017.
12. Y.Livnat,
J.Agutter, S.Moon, S.Foresti, "Visual correlation for situational
awareness", IEEE Symposium on Information Visualization (INFOVIS), pp.
95--102, 2005.
13. D.Mavris,
O.Pinon, D.Fullmer, "Systems design and modeling: A visual analytics
approach", Proceedings of the 27th International Congress of the
Aeronautical Sciences (ICAS), 2010.
14. SAS the
power to know, [Online]. Available: http://www.sas.com/en_us/home.html
[accessed on 15.03.2018].
15. K.Sharopin
et al., Vizualizacija medicinskih dannyh na baze paketa NovoSpark ["Vizualization of the Medical Data on the basis of package
NovoSpark"], Izvestiya SFedU. Engineering Sciences, vol. 109, pp. 242--249,
2010. [In Russian]
16. J.Woollen,
"A Visual Approach to Improving the Experience of Health Information for
Vulnerable Individuals", PhD Thesis, Columbia University Academic Commons,
2018.
17. V.Pilyugin
et al., Nauchnaja vizualizacija kak metod analiza nauchnyh dannyh [Scientific visualization as method of scientific data analysis],
Scientific Visualization, vol. 4, pp. 56--70, 2012. [In Russian]
18. http://atlas-kibana.mwt2.org:5601/app/kibana
19. http://sv-journal.org
20. http://atlas.cern
21. http://bigdatalab.nrcki.ru
22. http://lit.jinr.ru
