Boosting the performance of computational fluid dynamics codes for interactive supercomputing

Abstract

An extreme form of pipelining of the Piecewise-Parabolic Method (PPM) gas dynamics code has been used to dramatically increase its performance on the new generation of multicore CPUs. Exploiting this technique, together with a full integration of the several data post-processing and visualization utilities associated with this code has enabled numerical experiments in computational fluid dynamics to be performed interactively on a new, dedicated system in our lab, with immediate, user controlled visualization of the resulting flows on the PowerWall display. The code restructuring required to achieve the necessary CPU performance boost, as well as the parallel computing methods and systems used to enable interactive flow simulation are described. Requirements for these techniques to be applied to other codes are discussed, and our plans for tools that will assist programmers to exploit these techniques are briefly described. Examples showing the capability of the new system and software are given for applications in turbulence and stellar convection.

A Study of the Performance of Multifluid PPM Gas Dynamics on CPUs and GPUs

“…To fully realize these benefits, massive pipelining of numerical algorithms to achieve the highest possible amount of reuse of cached data is extremely important. The briquette data structure, massive code pipelining, and short, aligned vector operands all are features of the code restructuring and transformation that our team has been advocating as a result of our experience with the IBM Cell processor [2][3][4][5]. We have built and are refining automatic code translators that perform these code restructuring transformations as a precompilation step.…”

“…It also shares with our full application a very important feature: it has a large difference stencil, and the computation proceeds in phases that alternate between computations of cell and cell interface quantities. Each of these alternating phases of computation becomes transformed in our high performance code expression into a separate stage in a computation pipeline (see [2][3][4][5]). In this way, this PPM advection kernel reflects the overall character of our full application.…”

“…It has been extensively transformed so that the entire algorithm is performed in a pipelined fashion exclusively using aligned quadword operands. These code transformations have been described elsewhere [2][3][4][5]. We are building automatic code translators (see [5]) to produce our pipelined Fortran code from a much simpler form that is more easily written, read, modified, and maintained.…”

“…The Intel Westmere CPU is able to prefetch all the data needed to update a single grid briquette of 4 3 cells for each of 2 threads running on each of its 6 CPU cores during only 43% of the time it takes that thread to update the previous grid briquette in a sequence along the direction of the 1-D pass. The massive pipelining of our fluid dynamics algorithm that makes this possible has been described in detail elsewhere [2][3][4][5]. It results in an overall computational intensity of 33.5 flops per main device memory word read or written, which is sufficient to allow the CPU core to run without waiting on data transfers from or to its main device memory.…”

ExaSAT: An exascale co-design tool for performance modeling

“…In the traditional blocking scheme, each loop runs over the entire domain before proceeding to the next loop. In an alternative scheme (Woodward et al, 2010) all of the loops are run on a block before moving to the next block, as illustrated in Figure 5. Each large rectangle represents the iteration space at different points of progress (indicated by shading), and each subrectangle represents a block of the iteration space that fits into local memory.…”