Evolution of the ReadOut System of the ATLAS experiment

The ReadOut System (ROS) is a central and essential part of the ATLAS DAQ system. It receives and buffers data of events accepted by the first-level trigger from all subdetectors and first-level trigger subsystems. Event data are subsequently forwarded to the High-Level Trigger system and Event Builder via a 1 GbE-based network. The ATLAS ROS is completely renewed in view of the demanding conditions expected during LHC Run 2 and Run 3, to replace obsolete technologies and space constraints require it to be compact. The new ROS will consist of roughly 100 Linux-based 2U high rack mounted server PCs, each equipped with 2 PCIe I/O cards and two four 10 GbE interfaces. The FPGA-based PCIe I/O cards, developed by the ALICE collaboration, will be configured with ATLAS-specific firmware, the so-called RobinNP firmware. They will provide the connectivity to about 2000 optical point-to-point links conveying the ATLAS event data. This dense configuration provides an excellent test bench for studying I/O efficiency and challenges in current COTS PC architectures with non-uniform memory and I/O access paths. In this paper we will report on the requirements for Run 2 and on design choices for a system complying with and possibly exceeding them, as well as discuss the results of performance measurements for different computer architectures, highlighting the effects of non-uniform resource distributions. Finally we will present the status of the project and outlook for operation in 2015.