Skip to content

P2P Thrasher

Overview

The p2p_thrasher is designed to overwhelm the physical interconnect bridges between multiple GPUs. It hammers the fabric with DMA operations using bidirectional transfers to force continuous cross-die traffic.

Execution Mechanics

The kernel bypasses the host CPU entirely to perform massive, simultaneous asynchronous peer-to-peer (P2P) memory copies.

  • Initiates bidirectional Device-to-Device transfers without staging through system RAM.
  • Floods the PCIe Root Complex and specialized bridges like NVLink or Infinity Fabric.
  • Runs concurrently across all available links to maximize cross-die traffic and synchronization overhead.

Target Subsystems

  • Primary Target: NVLink, AMD Infinity Fabric, and PCIe Root Complex.
  • Secondary Target: DMA/Copy Engines.

Failure Symptoms

Critical Failures

  • Degraded Routing: Severe drops in throughput (GB/s) indicate fabric routing bottlenecks or thermal throttling on the interconnect bridges.
  • Physical Link Crashes: Complete driver timeouts caused by poorly seated PCIe risers or degraded NVLink connectors.