An Analysis of Data Management Schemes for Distributed Memory MIMD Systems
M. C. Cornea‐Hasegan, Dan C. Marinsecu, Zhongyun Zhang · Purdue e-Pubs (Purdue University System) · 1992
A comparative analysis of data management schemes for Distributed Memory MIMD systems for applications which need a very large shared data space, is presented.Work allocation, data distribution, fetching and allocation policies, as well as optimization techniques for iterative computations, are discussed.Measurements for a case study, the electron density averaging for the determination of the atomic structure of viruses are analyzed.We argue that future MPPs should provide primitives to support user controlled data management, rather than solutions, e.g., virtual memory or shared memory.mization techniques for iterative computations.A case study is discussed in detail in Section 4, namely the electron density averaging for the determination of the atomic structure of VIruses. Definitions and Basic Concepts Data spaceThe data space is the set of all data necessary for the parallel computations.We are primarily concerned with computations related to physical objects.The data space consists of the physical and geometrical information associated with a set of grid points.When the physical object has a regular shape and it is covered by a uniform mesh, then geometrical information may be implicit.For example, the data space may be viewed as a linear structure consisting of Ny records, one for each grid point.The coordinates of the i-th grid point Xi = (xLxf, ... ,xn may be easily computable as xi = fj(i,X1, ... ,X q ) where Xl, X 2 , ... ,X q define the geometry of the object.If the physical shape is not regular, but the mesh is uniform, then the object may be circumscribed by a regular volume in the space with q dimensions and an indicator will show if the grid point belongs to the original object or not.In these cases, the geometrical information can be omitted, saving precious data space.Another data compression strategy can be applied to objects which exhibit some symmetry.If an object consists of several replicas of a basic unit, the data space might contain only information related to the basic unit.Anytime information about a grid point outside the basic unit is needed, the data management has to map that particular grid point into the basic unit.If g is the amount of information associated with one grid point, and Ny the total number of grid points, then the size of the data space is D = g x Ny. Data allocation unitThe data allocation unit, DAU, is the smallest subset of the data space that is manipulated, read or written by a PE.A DAU might consist of the information associated with one or more grid points.In general, it is advantageous to group together a number of neighbor grid points into a DAU to reduce the overhead of storage management.