Visualization and data reordering using EcoRAM

Introduction

A growing number of seismic applications are becoming IO bound. In some cases, an application can be IO bound because the amount of computation per input sample is low, such as with Normal Move-Out. Another category is applications where the input, intermediate, and output space which are to large to simultaneously reside in physical Random Access Memory (RAM), such as conventional 3-D Reverse Time Migration (RTM) (Baysal et al., 1983).

Several different methods exist to improve IO performance. Raid based disk arrays stripe data over many different disk drives. Network storage uses a collection of nodes as temporary storage. Large shared-memory and distributed-memory machines can distribute the required memory storage over 100s to 1000s of nodes, therefore allowing the entire problem to sit in RAM. Recently, solid state disks have offered promise. These disks, when used in a raid system, can offer 2-5x the IO performance of conventional disks. A new weapon in the IO bottleneck battle is EcoRAM from Spansion. EcoRAM sits in conventional RAM slots but is composed of a series of solid state disks. An accelerator replaces one of the CPU motherboard slots to facilitate low latency, read speeds of $30$x conventional disk and $2-10$x write speeds.

I show that applications that are, or can be made, read dominant can benefit dramatically from EcoRAM with minimal programming changes. I begin by comparing IO characteristic of EcoRAM versus conventional RAM, disk, network, and solid state disk. I then describe how, and the results of, writing a visualization and transpose program to maximize EcoRAM performance.

Visualization and data reordering using EcoRAM