The paper considers the application of a software complex for data processing simulation in computing systems. The BM@N experiment data storage and processing system of the NICA complex is used as a simulated infrastructure. Simulation is performed in order to obtain recommendations on the organization of experimental data processing during a session with available allocated resources. The paper presents the results of simulating the processing of data that will be received during the BM@N experiment session, with several scenarios for distributing job flows across existing processing centers. In addition, some recommendations for organizing experimental data processing are proposed. The status of the work and future plans for the development of the software complex are formulated.
A software complex for simulating the processing of data coming from the experimental
facility of the NICA complex is being developed at the Meshcheryakov Laboratory of Information
Technologies of the Joint Institute for Nuclear Research [
Currently, work on the data processing simulation of the BM@N experiment is in progress
[
Requirements for the improvement of the software complex are formed in the process of its using. The main aim of the improvement is its further application for data processing simulation in any data storage and processing centers.
Before starting to simulate data processing, it is necessary to determine the structure of the data storage and processing center, the configurations of various equipment, as well as the characteristics of data flows and job flows. All these parameters are input data for simulation.
The simulated distributed data storage and processing system with the equipment parameters is shown in Figure 1.
The system consists of two centers and includes several levels of data storage. The data of particle collision events with a fixed target (raw data), which are selected by the trigger (Trigger BM@N), is first written to the Data reception buffer, then it arrives at the Intermediate data storage level, after which the full volume of raw data is sent to the local buffers (EOS LHEP and EOS LIT) of the respective data processing centers. All storage devices have a limited volume, and data
Class RawToDigit DigitToDst GenToSim SimToDst
DstToAna 3. Simulation transmission channels have a limited bandwidth. Data processing jobs are executed on the compute nodes of the above processing centers (NCX LHEP, T2 LIT and Supercomputer). Data processing jobs arrive at the compute nodes and are performed if there are free slots, i.e. processor cores. If all slots are occupied, the job goes to the queue and waits until some slot is free. Each job needs a certain amount of data to execute. Data is transferred from the storage to the compute node before the job starts.
A 30-day session of the BM@N experiment is considered in the article. It is assumed that the session will last continuously, except for minute breaks between runs. Each run lasts 2 minutes. Experimental data is written to files during each run. The recording is carried out continuously up to the maximum size of each file (35 GB). The speed of data selection by the BM@N trigger is 10 000 events per second. The average size of a single event is about 0.2 MB. The maximum planned volume of experimental data per session, taking into account the described characteristics, will be approximately 350 TB.
Raw binary data, which is received from the installation, is subject to processing, namely, digitization (the final format is digit) and conversion to reconstruction data (DST format), after which physical analysis of experimental data is performed. In addition to processing experimental data, it is planned to process simulation data (sim), which will also be converted into reconstruction data (DST) for subsequent physical analysis.
There are several classes of jobs in accordance with the described process of data processing. The characteristics of classes are described in Table 1. Job flows for the simulation of data processing were formed on the basis of the data from the table.
Avg. event processing time on one processor (ms)
In the example considered below, simulation is performed in order to provide recommendations on the organization of experimental data processing during the session with available allocated resources [fig. 1]. First of all, there were simulated three scenarios, where the data processing jobs described in Table 1 are distributed in different ways across the compute nodes [tab.2].
The results of simulating three scenarios showed that one of the computing resources could be fully allocated for processing simulated data up to reconstruction data (GenToSim and SimToDst jobs). As for the processing of experimental data, at best only 20% of raw data can be processed to digit data, and less than 10% of jobs on processing data to reconstruction data will have time to complete before the end of the session, which lasts 720 hours. Figure 2 shows graphs reflecting the number of completed experimental data processing jobs (RawToDigit and DigitToDst jobs) on computing resources in accordance with Scenario 2. Figure 2 does not show graphs of completing DstToAna jobs since the number of completed jobs is too small and is approximately 1%. This is due to the full load of all computing resources.
Surely, the result is unsatisfactory. Such a small amount of experimental data can be processed by the end of the session. After the end of the session, one will have to wait a few more months before the end of processing all the data acquired from the experiment. It may be noticed that a large amount of time is required to fully process a single data file for RawToDigit and DigitToDst jobs [tab. 1]. In this regard, it was decided to simulate a scenario similar to Scenario 2, in which express file processing will be added. Express processing will be run simultaneously with full processing. Express processing consists in processing 1% of the file in RawToDigit jobs to obtain preliminary results.
A software complex to simulate the processing of data, which comes from the experimental facility of the NICA complex, is being developed at MLIT JINR. The presented simulation results of the data processing of the BM@N experiment showed that with the described parameters of equipment, data flows and jobs flows, no more than 20% of raw data could be processed by the end of the session, lasting 30 days. It is also proposed to perform express data processing simultaneously with full processing, which will allow obtaining preliminary results every hour throughout the experiment. It is possible to conclude that the measurements obtained are correct.
Requirements for the improvement of the software complex are formed in the process of its using. The main aim of the improvement is its further application for data processing simulation in any data storage and processing centers. Thus, at the next stage, it is planned to develop a module for launching jobs similar to the pilot, as well as to conduct computational experiments taking into account the fact that the equipment does not have absolute reliability, i.e. to calculate the probability of equipment failure and recovery time.
This work supported by JINR grant for young scientists № 21-602-02.
The authors express their gratitude to Doctor of Physical and Mathematical Sciences, Professor, main researcher of the Meshcheryakov Laboratory of Information Technologies, Gennady Ososkov for his assistance in the work, valuable advice and fruitful discussions.