Knocking on Heaven's Door - Lisa Randall [119]
The triggers need to identify the potentially interesting collisions and discard the ones that won’t contain anything new. The events themselves—what leaves the interaction point and gets recorded in the detectors—must be sufficiently distinguishable from usual Standard Model processes. Knowing when the events look special tells us which events to keep. This makes the rate for readily recognizable new events even smaller still. The triggers have a formidable task. They are responsible for winnowing down the billion events per second to the few hundred that have a chance of being interesting.
A combination of hardware and software “gates” accomplishes this mission. Each successive trigger level rejects most of the events it receives as uninteresting, leaving a far more manageable amount of data. These data in turn get analyzed by the computer systems at 160 academic institutions around the globe.
The first-level trigger is hardware based—built into the detectors—and does a gross pass at identifying distinctive features, such as selecting events containing energetic muons or large transverse energy depositions in the calorimeters. While waiting a few microseconds for the result of the level-one trigger, the data from each bunch crossing are held in buffer. The higher-level triggers are software based. The selection algorithms run on a large computer cluster near the detector. The first-level trigger reduces the billion per second event rate to about 100,000 events per second, which the software triggers further reduced by a factor of about a thousand to a few hundred events.
Each event that passes the trigger carries a huge amount of information—the readouts of the detector elements we just discussed—of more than a megabyte. With a few hundred events per second, the experiments keep well over 100 megabytes of disk space per second, which amounts to over a petabyte, which is 1015 bytes, or one quadrillion bytes (how often do you get to use that word?), the equivalent of hundreds of thousands of DVDs worth of information, each year.
Tim Berners-Lee first developed the World Wide Web to deal with CERN data and let experimenters around the world share information on a computer in real time. The LHC Computing Grid is CERN’s next major computational advance. The Grid was launched late in 2008—after extensive software development—to help handle the enormous amounts of data that the experimenters intend to process. The CERN Grid uses both private fiber-optic cables and high-speed portions of the public Internet. It is so named because data aren’t associated with any single location but are instead distributed in computers around the world—much as the electricity in an urban area isn’t associated with one particular power plant.
Once the trigger-happy events that made it through are stored, they are distributed via the Grid all over the globe. With the Grid, computer networks all over the globe have ready access to the redundantly stored data. Whereas the web shares information, the Grid shares computational power and data storage among the many participating computers.
With the Grid, tiered computing centers process the data. Tier 0 is CERN’s central facility where the data get recorded and reprocessed from their raw form to one more suitable for physics analyses. High-bandwidth connections send the data to the dozen large national computing centers constituting Tier 1. Analysis groups can access these data if they choose to do so. Fiber-optic cables connect Tier 1 to the roughly 50 Tier 2 analysis centers located at universities, which have enough computing power to simulate physics processes and do some specific analyses. Finally, any university group can do Tier 3 analyses,