Status Talk - Coordinated Polling in User Space

BBB-Link: https://bbb.tu-dresden.de/b/mat-xin-oyh-xzn

Presentation Language: English

Modern high-performance computing systems rely on asynchronous hardware resources whose events must be delivered to the CPU with minimal latency. Common approaches are either busy polling or interrupt-driven mechanisms. While busy polling achieves the lowest latency, it requires dedicated CPU cores, wastes cycles when no communication occurs, and degrades under oversubscription. Interrupt-based approaches are more resource-efficient but introduce overhead from interrupt handling and scheduling, which is significant for latency-critical networks like InfiniBand.
 
Coordinated polling finds a middle ground by redirecting busy polling to a global daemon while the individual applications block. The daemon continuously monitors registered resources and wakes the corresponding user processes when progress is detected.
 
This presentation reports on the status of my Diploma thesis, which investigates implementing coordinated polling entirely in userspace without kernel modifications, enabling transparent deployment on existing HPC systems. The talk covers key design decisions, implementation challenges across the InfiniBand software stack, and early evaluation results on latency and energy efficiency