Accelerated workflow execution
Available in: 
pSeven Enterprise v2025.07
When executing a workflow in pSeven Enterprise, each block should ideally run immediately once its input data becomes available - meaning all prerequisite blocks linked to it have completed and produced their outputs. However, block startup often involves initialization delays due to resource allocation for containers, network latencies and other overhead.
pSeven Enterprise offers multiple block execution strategies, each designed to optimize workflow runtime and cluster resource consumption differently. Let's explore how these strategies work using a simple workflow with three sequential blocks.
1. All-at-once: All blocks start simultaneously with the workflow run. Most blocks become ready to process input data by the time it's available, which minimizes inter-block delays. Total workflow runtime then approaches the sum of individual block execution times. Conversely, this leads to excessive cluster load as all blocks reserve required resources upfront.
2. On-demand: Blocks start only when their input data becomes available. Due to initialization delays, blocks cannot immediately start processing input data during the workflow run, which can significantly increase inter-block delays and total workflow runtime. However, this approach minimizes the load by allocating resources only when needed and is recommended for overloaded clusters.
3. On-demand with predictive initialization: The new optimized scheduling strategy (now enabled by default) builds on the on-demand strategy. Using workflow run statistics, it predicts and compensates for block initialization delays. Given accurate timing estimates from previous runs, this strategy ensures each block is ready to execute precisely when its input data (outputs from all prerequisite blocks) becomes available. The updated run strategy maintains low computational overhead comparable to the original on-demand strategy, while achieving runtimes that nearly match the all-at-once strategy. Performance improvements are most significant in workflows with long chains of sequential blocks, while a little less noticeable in highly parallel workflows.
