14 Distributed Computing
625
servers are implemented on inexpensive commodity computers with large disks.
This is in turn stimulating multi-layer implementations, for example the D2D2T
(disk-to-disk-to-tape) scheme, where Disk servers store and retrieve data on diskbased pseudo-tape servers, which in turn store and retrieve data on “real” magnetic
cartridge Tape servers. Both disk-based tape servers and D2D2T schemes are
important because the large growth in data volumes in particle physics is the result
of a very modest growth in file size multiplied by a large growth in the number
of files, resulting in a situation which can be very inefficient for traditional tape
systems (the so-called small file problem in tape cartridge storage).
A “Hierarchical Storage Manager” (HSM) is implemented between the Data
servers and Tape servers using specific Control and Information servers. HSM
systems are not a common feature of conventional computing environments. HSM
systems that fulfill the needs of particle physics data processing are even less
common. Furthermore, the interaction between Data servers and Tape servers can be
rather complex and no clean, practical interfaces are available. Therefore, each data
center implements its own customized solution using a combination of communitywritten and commercial software. Examples of Data server/HSM implementations
are dCache interfaced to Tivoli or Enstore; GPFS interfaced to IBM HPSS; and
CERN’s Castor-2 which has its own data server and HSM implementations.
Depending on the solution, the HSM system may be more or less intertwined with
the Data server software, although the tendency is to separate as much as possible
the HSM and Data server implementations.
Experiment specific Information servers are deployed in order to manage and
make accessible a view of all their files containing custom details about their
characteristics, for example trigger conditions, detector configuration or beam type.
This is often called a “File Catalog” and may in itself be a quite complex database.
14.4 The Main Workflows of Data Processing
in an Experiment
Most experiments implement two “organized” data processing workflows: one for
data coming from the actual detector (awkwardly referred to as “real” data) and
another for simulations resulting from the techniques described in the contribution “Detector Simulation” (awkwardly referred to as “Monte Carlo” data). An
additional very important “organized” workflow derives calibration constants from
detector data taken either in normal mode or in special conditions for calibration
(cosmic rays, single beam, sending pulses to the front-end electronics, etc.).
Altogether, these three workflows are referred to as “production” and constitute the
process by which “raw” data is transformed into “analysis datasets”. Workflows
related to the use of these analysis datasets by individual physicists are described in
a later section.
625
servers are implemented on inexpensive commodity computers with large disks.
This is in turn stimulating multi-layer implementations, for example the D2D2T
(disk-to-disk-to-tape) scheme, where Disk servers store and retrieve data on diskbased pseudo-tape servers, which in turn store and retrieve data on “real” magnetic
cartridge Tape servers. Both disk-based tape servers and D2D2T schemes are
important because the large growth in data volumes in particle physics is the result
of a very modest growth in file size multiplied by a large growth in the number
of files, resulting in a situation which can be very inefficient for traditional tape
systems (the so-called small file problem in tape cartridge storage).
A “Hierarchical Storage Manager” (HSM) is implemented between the Data
servers and Tape servers using specific Control and Information servers. HSM
systems are not a common feature of conventional computing environments. HSM
systems that fulfill the needs of particle physics data processing are even less
common. Furthermore, the interaction between Data servers and Tape servers can be
rather complex and no clean, practical interfaces are available. Therefore, each data
center implements its own customized solution using a combination of communitywritten and commercial software. Examples of Data server/HSM implementations
are dCache interfaced to Tivoli or Enstore; GPFS interfaced to IBM HPSS; and
CERN’s Castor-2 which has its own data server and HSM implementations.
Depending on the solution, the HSM system may be more or less intertwined with
the Data server software, although the tendency is to separate as much as possible
the HSM and Data server implementations.
Experiment specific Information servers are deployed in order to manage and
make accessible a view of all their files containing custom details about their
characteristics, for example trigger conditions, detector configuration or beam type.
This is often called a “File Catalog” and may in itself be a quite complex database.
14.4 The Main Workflows of Data Processing
in an Experiment
Most experiments implement two “organized” data processing workflows: one for
data coming from the actual detector (awkwardly referred to as “real” data) and
another for simulations resulting from the techniques described in the contribution “Detector Simulation” (awkwardly referred to as “Monte Carlo” data). An
additional very important “organized” workflow derives calibration constants from
detector data taken either in normal mode or in special conditions for calibration
(cosmic rays, single beam, sending pulses to the front-end electronics, etc.).
Altogether, these three workflows are referred to as “production” and constitute the
process by which “raw” data is transformed into “analysis datasets”. Workflows
related to the use of these analysis datasets by individual physicists are described in
a later section.