Communications - June 2010

The schema of SDSS data includes more than 70 tables, though most user queries focus on only a few of them, referring, as needed, to spectra and images. The queries aim to spot objects with specific characteristics, similarities, and correlations. Patterns of query expression are also limited, featuring conjunctions of range and user-defined functions in both the predicate and the join clause.

Simulation scientific data. Earth science employs simulation models to help predict the motion of the ground during earthquakes. Ground motion is modeled with an octree-based hexahedral mesh19 produced by a mesh generator, using soil density as input (see Figure 2). A “solver” tool simulates the propagation of seismic waves through the Earth by approximating the solution to the wave equation at each mesh node. During each time step, the solver computes an estimate of each node velocity in the spatial directions, writing the results to the disk. The result is a 4D spatio-temporal earthquake data set describing the ground’s velocity response. Various types of analysis can be performed on the data set, employing both time-varying and space-varying queries. For example, a user might describe a feature in the ground-mesh, and the DBMS finds the approximate location of the feature in the simula-

figure 3. Workflow of the atLas experiment.

tion data set through multidimensional indexes.

Combined simulation and observational data. The ATLAS experiment ( http://atlas.ch/), a particle-physics experiment in the Large Hadron Collider ( http://lhc.web.cern.ch/lhc/) beneath the Swiss-French border near Geneva, is an example of scientific data processing that combines both simulated and observed data. ATLAS intends to search for new discoveries in the head-on collision of two highly energized proton beams. The entire workflow of the experiment involves petabytes of data and thousands of users from organizations the world over (see Figure 3).

We first describe some of major ATLAS data types: The raw data is the direct observational data of the particle collisions. The detector’s output rate is about 200Hz, and raw data, or electrical signals, is generated at about 320MB/sec, then reconstructed using various algorithms to produce event summary data (ESD). ESD has an object-oriented representation of the reconstructed events (collisions), with content intended to make access to raw data unnecessary for most physics applications. ESD is further processed to create analysis object data (AOD), a reduced event representation suitable for user analysis. Data volume decreases gradually from raw to ESD

to AOD. Another important data type is tag data, or event-level metadata, stored in relational databases, designed to support efficient identification and selection of events of interest to a given analysis.

Due to the complexity of the experiment and the project’s worldwide scope, participating sites are divided into multiple layers. The Tier-0 layer is a single site—CERN itself—where the detector is located and the raw data is collected. The first reconstruction of the observed electrical signals into physics events is also done at CERN, producing ESD, AOD, and tag data. Tier- 1 sites are typically large national computing centers that receive replicated data from the Tier-0 site. Tier- 1 sites are also responsible for reprocessing older data, as well as for storing the final results from Monte Carlo simulations at Tier- 2 sites. Tier- 2 sites are mostly institutes and universities providing computing resources for Monte Carlo simulations and end-user analysis. All sites have pledged computing resources, though the vast majority is not dedicated to ATLAS or to high-energy physics experiments.

The Tier-0 site is both computation-and storage-intensive, since it stores the raw data and performs the initial event reconstruction. It also serves data to the Tier- 1 sites, with aggregate sustained transfer rates for raw, ESD,