Modeling and Validating Time, Buffering, and Utilization of a Large-Scale, Real-Time Data Acquisition System

Abstract

Abstract-Data acquisition systems for large-scale high-energy physics experiments have to handle hundreds of gigabytes per second of data, and are typically implemented as specialized data centers that connect a very large number of front-end electronics devices to an event detection and storage system. The design of such systems is often based on many assumptions, small-scale experiments and a substantial amount of over-provisioning.In this paper, we introduce a discrete event-based simulation tool that models the dataflow of the current ATLAS data acquisition system, with the main goal to be accurate with regard to the main operational characteristics. We measure buffer occupancy counting the number of elements in buffers; resource utilization measuring output bandwidth and counting the number of active processing units, and their time evolution by comparing data over many consecutive and small periods of time. We perform studies on the error in simulation when comparing the results to a large amount of real-world operational data. We show which efforts are required to minimize the error for such a configuration, and explain possible reasons for the most important outliers we are observing. Furthermore, we use this tool to derive an operational envelope of the system, which describes the minimal amount of resources required to fulfill certain real-time guarantees.

Buffer Provisioning for Large-Scale Data-Acquisition Systems

“…The third histogram used as input parameter for the single buffer model can be greatly simplified by using the following formula (see also [15]). According to this formula, the average processing time over N p records for a processing time p i and the number of processed events q i is:…”

“…Thus, we configure for first buffering stage of the split buffer we set the number of Data Handlers to one, mimicking single buffer configuration. We follow the methodology described in [15] for the single buffer model, but extend the validation for the split buffer model as described in Section 3. The simulation is run for 60 simulated seconds.…”

“…Previous related work includes [15], which describes a simulation model to study the single-stage buffering system of the existing ATLAS "Phase-0" experiment, a model which in this paper is called single buffer simulation model. By contrast, this work is about a split buffer simulation model for a large buffer like the one it will be required for ATLAS "Phase-2", and which is distributed across multiple storage entities.…”

“…Model validation is done from two sides. First, the split buffer model is run and compared against the existing system by using single buffer model results [15] as reference, described in Section 4. Second, a small-scale emulated DAQ system was created and the simulation model compared against the output metrics of the smallscale emulated DAQ system, as described in Section 5.…”

“…This section mainly describes the foreseen ATLAS "Phase-2" data acquisition system, but also overviews the DAQ of other systems.. Details for the ATLAS "Phase-0" system, the corresponding simulation model and results can be found in [15]. Figure 2 shows the dataflow organization of the ATLAS "Phase-2" DAQ.…”