The design of a distributed key-value store for petascale hot storage in data acquisition systems

Abstract

Data acquisition systems for high energy physics experiments readout terabytes of data per second from a large number of electronic components. They are thus inherently distributed systems and require fast online data selection, otherwise requirements for permanent storage would be enormous. Still, incoming data need to be buffered while waiting for this selection to happen. Each minute of an experiment can produce hundreds of terabytes that cannot be lost before a selection decision is made. In this context, we present the design of DAQDB (Data Acquisition Database) -a distributed key-value store for high-bandwidth, generic data storage in event-driven systems. DAQDB offers not only high-capacity and low-latency buffer for fast data selection, but also opens a new approach in high-bandwidth data acquisition by decoupling the lifetime of the data analysis processes from the changing event rate due to the duty cycle of the data source. This is achievable by the option to extend its capacity even up to hundreds of petabytes to store hours of an experiment's data. Our initial performance evaluation shows that DAQDB is a promising alternative to generic database solutions for the high luminosity upgrades of the The ATLAS experiment at the LHC implements physical event building in a two-step process: fragments are transferred from the detector to the readout system (ROS) via pointto-point optical links connected to FPGA-based PCIe cards. The data are buffered in the ROS and made available for the High-Level Trigger (HLT). The HLT processing unit collects and processes fragments of a given event. A data collection manager (DCM) dynamically distributes the available resources and orchestrates all the data flow process [9].Another LHC experiment, CMS, implements physical event building in a similar strategy, the event builder assembles the fragments in the RAM, and a central entity supervises the allocation of available building resources to the events. The acquisition of all the data, from the detector to the selection of the interesting events are coupled. The DAQ system can buffer data up to 90 seconds before the HLT selection, and it can store only selected events for few days [10].

Let’s get our hands dirty: a comprehensive evaluation of DAQDB, key-value store for petascale hot storage

“…Eventually, an independent DAQDB server instance is to be run on each CPU. DAQDB uses 64-bit adaptive radix tree (ART) [3] with 8b per ART node, with DAQ-specific optimizations [1]. Both event data and data structure are kept on PMem.…”

“…In the second evaluation we are analyzing how DAQDB performance depends on event fragment size in more detail. We also study if there is any performance impact when both the readout and filtering threads are running in parallel with the latter using the GetAny feature [1] to request the next available event ID from the store. The performance characteristics are presented in Figure 4.…”

“…Decoupling real-time data acquisition from event selection is a promising approach for future data acquistion (DAQ) systems of large-scale experiments like the Large Hadron Collider (LHC) at CERN. In [1], facing the new requirements of the High Luminosity upgrade of the LHC, we analyzed this approach in detail and proposed a possible design using a large distributed key-value store (KVS) as a generic high-bandwidth data storage system. In this paper we present the first performance evaluation of our open-source implementation of this design -Data Acquisition Database (DAQDB) [2].…”