Figure 3

The 'stride' as a control parameter. PaPy's adjustable stride modulates the trade-off between high memory consumption and parallelism (high stride) versus less aggressive paralellism and lower memory consumption (lower stride). This diagram un-winds PaPy's parallelism to show the interplay between the stride and item-level processing as pipeline execution proceeds (main blue arrow directed rightward). The relevent pipers are shown to the left (A), and traversal of the workflow graph by data-batches is shown in (B). Execution progress is also indicated by broken arrows progressing to the right, each arc representing equal incremenets of time (t₀, t₁,.... assuming a uniform processing time per data-item) for strides of 3 (orange) or 5 (green).