Speaker
Description
High-repetition-rate petawatt laser facilities require deterministic and stable large-scale image transfer for online multi-camera diagnostics. Even when cameras and IOCs maintain steady frame updates, complete image-group readout on the client side may experience severe tail latency, introducing uncertainty into online visualization, data logging, and feedback control. In this work, we establish a GigE Vision multi-camera testbed representative of the diagnostic scale of the CLAPA-II petawatt laser proton accelerator and systematically evaluate the end-to-end performance of EPICS Channel Access (CA) and PV Access (PVA) under synchronous serial client readout.
Under 10 GbE networking with ten cameras operating at 20 fps, both CA and PVA achieve an aggregate image production rate of approximately 466 MiB/s on the IOC side. However, significant differences emerge in client-side complete-cycle readout latency. For CA, maximum readout times range from 30.97 to 58.34 s, whereas PVA remains within 1.79 s. Under a bandwidth-constrained 1 GbE configuration with four cameras, the 95th-percentile (p95) readout latency reaches 67.13 s for CA, compared with 3.61 s for PVA. Packet-level tracing reveals that the long-tail latency observed with CA is accompanied by persistent TCP retransmissions, zero-window events, and window-full conditions, indicating complex interactions between serial request scheduling and transport-layer backpressure.
Based on these observations, we propose a production-consumer rate separation evaluation framework, emphasizing that control-system acceptance criteria should consider not only average throughput but also tail-latency metrics, including p95/p99 latency and maximum complete-group readout time. Finally, we discuss practical deployment considerations for EPICS areaDetector-based multi-camera diagnostic systems.
