**2. WLCG**

As mentioned in section 1, the LHC experiments are designed to search for rare events with the signal/noise ratio as low as 10−13. This Physics requires a study of enormous number of pp and Pb-Pb collisions resulting in the production of data volumes of more than 10 PetaBytes per one data taking year. The original estimates elaborated when the LCG TDR [13] was put together were about ∼ 15 PetaBytes (PB) of new data each year which translates into ∼ 200 thousands of CPUs/processor cores and 45 PB of disk storage to keep the raw, processed and simulated data.

Fig. 2. Distribution of WLCG computing centers

grids, EGI [16] and OSG [17], or by several regional or national grids.

**2.2 Hierarchical (Tier) structure, the roles of different Tier-sites**

Currently, the WLCG integrates over 140 computing sites, more than 250 thousands CPU cores and over 150 PB of disk storage. It is now the world's largest computing grid: the WLCG operates resources provided by other collaborating grid projects: either the two main global

Grid Computing in High Energy Physics Experiments 185

The WLCG has a hierarchical structure based on the recommendations of the MONARC project [18], see Figure 3. The individual participating sites are classified according to their resources and level of provided services into several categories called Tiers. There is one Tier-0 site which is CERN, then 11 Tier-1 centers, which are large computing centers with thousands of CPUs, PBs of disk storage, tape storage systems and 24/7 Grid support service (Canada: TRIUMF, France: IN2P3, Germany: KIT/FZK, Italy: INFN, Netherlands: NIKHEF/SARA, Nordic countries: Nordic Datagrid Facility (NDGF), Spain: Port d'Informació Científica (PIC), Taipei: ASGC, United Kingdom: GridPP, USA: Fermilab-CMS and BNL ATLAS). Then there are currently about 140 Tier-2 sites covering most of the globe. The system also recognizes Tier-3 centers, which are small local computing clusters at universities or research institutes. The raw data recorded by the LHC experiments (raw data) is shipped at first to the CERN Computing Center (CC) through dedicated links. CERN Tier-0 accepts data at average of 2.6 GBytes(GB)/s with peaks up to 11 GB/s. At CERN, the data is archived in the CERN tape system CASTOR [19] and goes through the first level of processing - the first pass of reconstruction. The raw data is also replicated to the Tier-1 centers, so there are always 2 copies of the raw data files. CERN serves data at average of 7 GB/s with peaks up to 25 GB/s [20]. The Tier-0 writes on average 2 PB of data per month to tape in pp running, and double that in the 1 month of Pb-Pb collisions, (cf. Figures 4,5). At Tier-1 centers, the raw data replicas are permanently stored as mentioned before and several passes of the data re-processing are performed. This multiple-stage data re-processing is performed using methods to detect

Nowadays, 200 thousands cores does not sound like much and one can find them in large computer centers. 50 PB of a disk storage is however not that common. In any case, at the time the LHC Computing Grid was launched there was no single site within the LHC community able to provide such computing power. So, the task of processing the LHC data has been a distributed computing problem right from the start.
