Browse Prior Art Database

Data distribution in a scalable reconstructor Disclosure Number: IPCOM000016621D
Publication Date: 2003-Jul-07
Document File: 4 page(s) / 156K

Publishing Venue

The Prior Art Database



This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 34% of the total text.

Page 1 of 4


Data distribution in a scalable reconstructor

A data distribution for a scalable reconstructor consistis of a cluster of processor nodes connected via a point-to-point network. Data distribution is required for data parallel processing. In a data parallel program, each node processes a slice of the total data. Parallel processing is often only viable if data is sliced in a particular manner. For example, for 2D Cartesian Fourier transformation, data might be sliced in rows for Fourier transformation in the row direction and sliced in columns for Fourier transformation in the column direction. The sliced row data is distributed over the nodes for parallel Fourier reconstruction in the row direction. The sliced column data is distributed over the nodes for parallel Fourier reconstruction in the column direction.

    General data distribution strategies require a network on which each node can transfer data to every other node. Each node in the network conceptually has a unidirectional point-to-point connection to each other node in the network dedicated for the transfer of data for data distribution.

node 1 node 2

node 1 node 2

Figure 1: Unidirectional point-to-point network topology model of a four-node reconstructor.

Such unidirectional point-to-point connections are typically mapped on a restricted physical network. The discussed strategy does not make explicit assumptions about the actual network technology. It can be implemented by means an incomplete network of bi-directional point-to-point connections, by means of a shared network medium such as Ethernet or by means of a bus. The actual network technology determines the network bandwidth and transfer latency. The discussed strategy attempts to mitigate these effects as far as possible.

    The discussion will assume a reconstructor consisting of four nodes. The data distribution strategy can easily be extended to networks consisting of any number of nodes. The dimensionality and size of the actual data may place a limit on the degree of parallelism that can be achieved in practice.

    As an example, assume a four-channel synergy acquisition processed on a four-node reconstructor. Channel parallel processing is natural as acquisition can readily route channel data to specific nodes. Image composition however, requires that individually processed channel data in some way come together to be combined in images. This suggests two phase processing; a first channel parallel phase and a second image parallel phase.

[This page contains 1 picture or other non-text object]

Page 2 of 4

node 1

node 2

node 3

node 4

C1 [1]

C1 [1]

C1 [2] C1 [3]

C1 [4]

C2 [1]

C3 [1]

C4 [1]

C2 [1]

C2 [2] C2 [3] C2 [4]

C1 [2]

C2 [2]

C3 [2]

C4 [2]

C3 [1] C3 [2]

C3 [3] C3 [4]

C1 [3]

C2 [3]

C3 [3]

C4 [3]

C4 [1] C4 [2] C4 [3]

C4 [4]

C1 [4]

C2 [4]

C3 [4]

C4 [4]

channel parallel

image parallel

Figure 2: Data transfers involved during channel parallel to image parallel switch. Channel data is allocated to individual no...