The Giant Cube Transposer

The talk will focus on the difficulty in modelling strong hydrodynamic turbulence on a large loosely-coupled cluster. 
It will be shown that at the core of the problem lies the cube transposition pattern which can be thought of as
transposition of a large distributed matrix with a very large size of the matrix element (in the order of 100Mb).
This pattern is common to all 3d FFT-based methods and occurs in a range of 2d applications as well where
a distributed matrix has to be transposed and updated iteratively.

The talk will present the idea of an electronic implementation of a cube transposer for which some initial funding
is available. There is also some preliminary industrial interest. The transposer will be a combination of
global shared storage and a hardware data mover with a very large cross-section bandwidth (in the order of
tens of terrabytes per second) implementing one specific pattern at one or more specific cube sizes. The expectation
is to be able to transpose a (2K)3 cube within 10ms for a 64-node cluster.