MOTIVATION

1. Prospects for the use of remote real time computing over long distances in the ATLAS Trigger/DAQ systemR. W. Dobinson (CERN), J. Hansen (NBI), K. Korcyl (INP), J.Pinfold (UofA), M. Turala (INP) ATLAS TRIGGER/DAQ SYSTEM • MOTIVATION • ATLAS Trigger/DAQ system uses a three-level architecture to reduce an initial rate of 109 particle interactions per second down to 100 Hz of events of size 2 MB – the throughput acceptable for mass storage. At the final, the third level (Event Filter) a massive processing power of O (2000) CPUs is required to analyze data and achieve the required reduction factor. A large processing farm will be built at CERN, where the experiment will take place. However, there are number of advantages in situating some of this capacity in the member states: • it is easier to fund computing equipment at home, as opposed to CERN • it is easier to involve existing staff, in particular IT professionals • computing equipment can be shared with other users • To use remote, home based, computing capacity efficiently will require very high performance networking at an affordable price. For example, if half the ATLAS event farms were situated off the CERN site, then several 10s of Gigabits per second of long haul traffic would be needed to flow to remote farms. Remote Event Processing Farms ATLAS Detectors ROB ROB ROB ROB PF Front End Network PF Dedicated light paths SFI SFI SFI L2PU L2PU L2PU Back End Network L2PU L2PU Packet switched WAN (GEANT) 10 GE PF PF PF PF PF PF Local Event Processing Farms Remote Event Processing Farms COPENHAGEN EDMONTON (CANADA) GEANT + Danish National CRACOW Dedicated ligth-path GEANT + Polish National CERN EQUIPMENT We have developed and used custom measuring equipment to determine the quality of service for connections between CERN and locations in Europe and Canada. The equipment is based on an Alteon Gigabit Ethernet (GE) network interface card (NIC) reprogrammed to act as an IP traffic generator, a custom clock card and commercial Global Positioning System equipment. We can measure one-way latency, loss and reordering on a packet-by-packet basis, as a function of load, up to Gigabit Ethernet speed. RESULTS – summary: We have measured packet latency, loss and reordering as a function of throughput up to GE line speed between CERN and NBI and between CERN and the University of Alberta. The connections are error-free up to line speed with a small percentage of out-of-order packets for the highest loads We have achieved 800 Mbps streaming rate with the TCP/IP by tuning protocol parameters. The hardware setup (number of PC CPUs, their speed and type of NIC card) was not optimal. This, together with an extra processing needed for handling the out-of-oder packets, limited the throughput from approaching more closely to line speed. The connection to Cracow was limited to 100 Mbps with an allocated channel. The line was not error-free (we measured a few percent of packet loss for all loads). The achieved throughput for streaming with TCP/IP was only 25% of the allocated bandwidth showing the sensivity of the TCP/IP to packet loss. CONCLUSIONS We have shown that long distance error-free Gbps connections can be established now over packet switching networks or dedicated light-paths. Data transfers rates using the TCP/IP protocol have been achieved up to 800 Mbps (further optimizations are possible including the use of alternative TCP/IP protocol stacks). We plan to use the present trans-Atlantic connection to support ATLAS test-beam activities. We plan to test a 10 GE connection later this year.