The primary purpose of the Computing Model exercise is to describe the model and estimate the computing resources required for a full year of data taking in 2008, with the inputs as described in Section 2.2 . Two main areas of computing activity have to be considered:
E.g. reconstruction, MC simulation. Two processing passes are assumed for the dataset in a year; the model is as described in Section 2.2.5.5
The data analysis is more uncertain in its requirements, but a plausible scenario has been outlined in section Section 2.3 . It should be stressed that the model and the resource requirements below do not include the personal analysis performed at the Tier-3 resources (or their equivalent share of Tier-2s and Tier-1s). The user load includes a CPU-intensive term that is in aggregate equal to the simulation of 20% of the data rate, (this may be full simulation, fast simulation, reconstruction or other CPU-intensive work), another term that describes the analysis of the working group DPDs in a chaotic fashion, and a final contribution for limited event re-reconstruction. Given the many different analyses to be performed and the various data formats required, these are by necessity approximations.
It is important to note that the predictions that follow include allowance for inefficiencies in the usage of CPU and disk resources. For scheduled CPU usage, the efficiency is assumed to be 85%, while for chaotic usage it falls to 60%. The disk storage efficiency is taken to be 70%, while tape storage efficiencies are assumed to be 100%.
The model assumes there will be 10 Tier-1s, which will be of different sizes. We assume there are on average three Tier-2 facilities for each Tier-1; again, there will be a range of sizes of Tier-2. The CERN Analysis Facility is taken to have 100 active ATLAS users (in addition to the active 600 users associated with the Tier-2s), five times the size of a nominal Tier-2.
The simulation is (along with the analysis load) a key driver of the Tier-2 resource requirements. It also has an effect on the Tier-1 storage and CPU requirements, as they host the simulated data. Many previous experiments have been severely limited in their capacity to produce fully simulated data, and this has hampered their physics output. Table 7-1 shows, for the year 2008 data only, the components that depend on the percentage of the data that is fully simulated as separate items for the case of 20% of the full data rate simulated; this simulation rate (4x108 events simulated per year) would be barely acceptable. Table 7-2 gives the same breakdown for the full capacity required at the start of 2008, i.e. including the capacity to deal with the 2007 data as well.
It has been assumed that the Tier-1 facilities follow the processing model outlined in See Computing Model. It is further assumed that the raw data are stored on tape/slow access media, with only a small subset remaining on disk to allow software development and testing1. The actual Tier-1 and Tier-2 capacity might be larger, and indeed shared with other experiments. Here only that part that is visible to and accessible by all ATLAS members, is taken into account, and would be credited in the ATLAS accounting.
Copyright © CERN 2005