Wait-Free Epoch-Based Reclamation Without Tail Latency Spikes From Stalled Readers

Combining QSBR / Crossbeam-style epoch pinning with Hazard Eras to prevent a single preempted reader thread from halting memory reclamation.

1. The Achilles' Heel of Pure Epoch-Based Reclamation (EBR)

Epoch-Based Reclamation (used in `crossbeam-epoch` and RCU-like userspace structures) is blazingly fast because readers only issue a single atomic store-load barrier to pin the global epoch. However, if a single reader thread blocks on a page fault or gets preempted while holding an epoch guard, the global garbage limbo list cannot advance—growing at hundreds of megabytes per second under heavy write churn.

2. Bounded Limbo Ejection & Reader Yield Signaling

By tagging retired nodes with birth and retirement epochs (Hazard Era style) and automatically splitting long-running scan iterators across fine-grained 256-key cursor checkpoints that unpin and re-pin the guard every 50 µs, we bounded maximum unreclaimed limbo memory to 16 MiB even under artificial `SIGSTOP` thread injection.

← Older Note (2025-01-09)Measuring PCIe Peer-to-Peer DMA Throughput Between NVMe Drives and RDMA NICsNewer Note (2025-04-02) →Tracing Ephemeral Port Exhaustion and TIME_WAIT Reuse Anomalies Behind L4 Load Balancers

← Back to all notes · View full chronological archive (24)