Hardware Reference
In-Depth Information
5. Minimize file size. Stochastic simulations tend to produce multi-gigabyte
files, and as their use is expanding, the file size is not expected to sta-
bilize. For FEM applications, the file size depends on the size of the
FE meshes that are produced by simulation processes and the number
of results per mesh. For highly nonlinear problems like the ones dis-
cussed in this section, the number of timesteps and the level of mesh
refinement can be extremely high. This is multiplied by the number of
simulations stored in a single file, which in turn depends on the number
of parameters that the user is varying. But in a time of cheap hardware,
the requirement of minimizing file size might appear irrelevant. For in-
dustrial users however, storage economics is a complicated matter. It
is not uncommon for the TCO/terabyte/year to be equal to $15,000 or
more if all cost items are factored in, such as labor costs, backup, redun-
dancy, servers, floor space, software licenses, etc. Some IT departments
are also using Total Cost of Data Ownership (TCDO), which considers
the actual amount of data stored per terabyte, factoring in the cost of
waste (usable and unallocated and allocated but not used space), RAID,
and backup overhead costs. The TCDO of 1 TB of data/year may be
multiple times higher than the TCO of 1 TB of storage/year.
6. Use industry standards. Avoid niche and untested platforms and sys-
tems. Industrial customers avoid things they cannot manage and that,
upon failure, will disrupt their processes. Custom-made solutions are
rare and, when present, their lifetime is surprisingly long. Customers
prefer not to use exotic solutions such as file systems or parallel access
patterns that are not fully supported by existing software vendors or well
known by their own IT. Otherwise, the cost of failure can be extremely
high.
Apart from the customer requirements, the Voice of the Customer|there are
additional requirements emanating from the software vendors|the Voice of
the Business:
1. Use proprietary formats. As much as data is valuable for customers,
software vendors do not like competitors to build software on top of
that data. Most software vendors use proprietary file formats that, to
some extent, conceal the structure of the data in the file.
2. Allow for different persistence concepts. Modern software architectures
have departed from the typical Plain Old Data Structure (POD) type
of data models. The advent of Object-Oriented Programming (OOP)
has introduced extensive and complicated relationships between data.
OOP is combining the data and the methods to manipulate the data in
one object entity. Modern software systems often require large networks
of objects that represent specific and limited concepts. Such networks
develop to graphs with cyclic dependencies. Collecting the data for per-
sistence from such graphs is not trivial.
 
Search WWH ::




Custom Search