Daemon-Assisted Batched Polling for InfiniBand

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

Presentation Language: English

Modern high-performance computing systems and data centers rely on RDMA to provide low-latency communication between nodes. Common approaches for waiting on asynchronous network events are either busy polling or interrupt-driven mechanisms. While busy polling achieves the lowest latency, it requires dedicated CPU cores, wastes energy and degrades under oversubscription. Interrupt-based waiting yields the CPU while blocking, but kernel interrupt handling and process rescheduling push average latency into the multi-microsecond range, which is significant for latency-critical networks like InfiniBand.

Daemon-Assisted Batched Polling (DABP) occupies the middle ground by moving busy polling to a central daemon. Applications register with the daemon and block while the daemon continuously monitors all registered queues and wakes the corresponding application upon detecting progress.

This presentation covers my Diploma thesis, which designs and implements DABP for InfiniBand in user space on GNU/Linux without requiring kernel modifications. The talk walks through key design decisions, implementation challenges across the software stack and extended evaluation results on latency and energy efficiency.