Grid Computing

Grid Computing is
"The sharing that we are concerned with is not primarily file exchange but rather direct access to computers, software, data, and other resources, as is required by a range of collaborative problem-solving and resource-brokering strategies emerging in industry, science, and engineering. This sharing is, necessarily, highly controlled, with resource providers and consumers defining clearly and carefully just what is shared, who is allowed to share, and the conditions under which sharing occurs. A set of individuals and/or institutions defined by such sharing rules form what we call a virtual organization."
(The Anatomy of the Grid, Intl J. Supercomputer Applications, 2001; 15 (3), Ian Foster, Carl Kesselman, Steven Tuecke)

HEP analysis

Analysis of High Energy Physics Data

The analysis of the data of HEP (high energy physics) experiments requires a sophisticated chain of data processing and data reduction. The data recorded by the experiment has to follow a chain of reconstruction and analysis programs before they are examined by the physicist. The stages can be characterized by their data format.

DataGrid © CERN Ref Nunber: BUL-PHO-2003-010

In the first stage raw data is delivered by the data acquisition system of the experiment. The data consists out of the raw measurements of the detectors. These can be channel numbers, time measurements,charge depositions and other signals. Due to the large number of channels one can expect a single event to have a size in the order of several hundreds of kilobytes to several megabytes. Integrated over the time of data taking during a year the overall data size is expected to be in the order of a petabyte. The data has to be recorded in a safe way on permanent storage.

The data will then be processed in a complicated reconstruction step. Raw numbers are converted into physical parameters as space coordinates or calibrated energy deposits. A further pattern recognition step is translating these numbers in parameters of observed particles, characterizing them by their moments or their energy. The output of this processing step is called reconstructed data. Its size is typically in the same order as the raw data. Data in that format has to be distributed to a limited number of users for detector studies and some specific analysis.

To simplify the analysis procedure only the most relevant quantities are stored in separate streams. These data is typically referred AOD (Analysis Object Data). The size of the events is significantly reduced and this format provides the input for physics analysis. It is expected that the amount of AOD for a year of data taking is in the order of 100 TB.

In the process of the analysis of the data physicists will often perform a further data reduction. Sometimes skims for interesting events are performed. Often the physicist define their own ad hoc data format. In ATLAS these datasets are known as DPD (Derived Physics Data), in CMS as PAT Skims (Physics Analysis Toolkit).

Of special consideration are simulated event data, the Monte Carlo events. In the process of the analysis it is required to study carefully the sensitivity and coverage of the detectors. The basis of such studies are careful and detailed simulation of event simulations.

Tier structure

The Organization of the Global Infrastructure

Already at an early stage it was recognized that analysis of data on the foreseen scale would require careful planning and organization. Different computing centres would have to have different roles according to their resources and their geographical location. The MONARC working group (Models Of Networked Analysis at Regional Centres) defined a tiered structure that groups computing centres according to their characteristics.

Analysis of High Energy Physics Data

The Tier-0 is the computing centre in CERN. As the site is hosting the experiments, its computing centre has a special role. It is required that all original data are recorded to permanent storage. Then initial processing of the data has to be performed on the site to provide rapid feedback to the operation. The data is then sent to other computing centres for analysis. It´s main role is to provide analysis capacity for local users. Computing resources for simulation can be provided to the experiments on an opportunistic basis. Other resources are usually reserved for local users.

A number of large computing centres (eleven in total) take the role of Tier-1 centres for the various computing centres. They receive data directly from CERN and provide additional permanent storage. The computing centres provide also the computing resources for reprocessing of the data, required at a later stage in the analysis. As major facilities they have a special role also in providing reliable services to the community as databases and catalogues. The centres are also responsible for the collection of simulated events produced at higher Tier centres.

A large number of sites take the responsibility as Tier-2 centres. Their role is to provide the bulk of computing resources for simulation and analysis. Typically they are associated with large disk storage to provide temporary storage data that is required for analysis. In total about hundred such centres have been identified.

The model is complemented by a large number of smaller computing centres at various universities and laboratories, the Tier-3 centres. Their main role is to provide analysis capacity for local users. Computing resources for simulation can be provided to the experiments on an opportunistic basis. Other resources are usually reserved for local users.