Skip to content

Cycles 5.2.0 / OpenUSD 26.03 hangs after completing the Moana Island render #30

Description

@nicolaspopravka

Summary

On a large-memory pod, the Moana Island render completes and writes its image, but the Cycles render process never exits. The main thread spins at 100% CPU with all worker threads parked, and the last renderer log line is a SyncAovBindings warning about an unknown pass elementId. The hang happens after a successful render, so it is distinct from both published failure modes: it is not the status-139 material-sync crash, and it is not a memory-limit result.

Configuration

  • Cycles 5.2.0 (Hydra delegate), with the Blender cycles PR #75 material-sync change (the build validated for OpenChessSet on test/cy2026-cycles-only)
  • OpenUSD 26.03
  • Container: ghcr.io/nicolaspopravka/usd-render-benchmark-cycles@sha256:bad67a76cbfa5e342ee06049864555b8280630258a446a1db6d8f6fcba7c8d50
  • NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB VRAM, 188 GB system RAM
  • NVIDIA driver 580.159.04
  • Rocky Linux 8.10 (Green Obsidian), Linux 6.8.0-117
  • Headless NVIDIA EGL through the release-matched usdrecord wrapper
  • Scene: scenes/MoanaIsland/usd/island.usda, camera /island/cam/shotCam, --purposes render

Observed behavior

The render runs for roughly 14 minutes and writes island.jpg (960×402, 9,127 unique colors, mean RGB (218, 218, 250)). The process then does not terminate:

  • About 19 minutes after the image was written, the process was still alive with the main thread spinning at exactly 100% CPU (201 ticks over a 2 s sample) and all 252 worker threads parked in futex_wait_queue.
  • render.log is frozen. Its last line is:
Warning: in SyncAovBindings at line 154 of /opt/build/cycles/src/hydra/session.cpp -- Unknown pass elementId
  • Peak resident memory was ~96.5 GiB of the 188 GiB available, so this is not memory pressure.
  • gdb could not attach to the live process (the container lacks CAP_SYS_PTRACE), so the spin site is not symbolized.
  • A stock usdrecord invocation would therefore never return. The pod was deleted after ~35 minutes of total runtime.

On the smaller pods this defect is never reached: the unpatched build segfaults during material sync (GH #25 class, fixed by PR #75), and the patched build on a ~29 GiB pod is killed by the cgroup OOM killer before completing (GH #9). Only a pod with enough memory to finish the render exposes the hang.

Logs and outputs

Follow-up

This is an hdCycles defect candidate: after the finished render, session work appears to wait on an unknown pass elementId while the main thread spins. When the next tagged Cycles release containing PR #75 lands and the full-scope rerun documented on GH #9 happens, capture a symbolized backtrace of the spin site (a debug build, or a runtime with CAP_SYS_PTRACE) and report upstream with this issue as the benchmark reference.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    benchmark-findingEvidence-backed benchmark result worth trackingupstream-candidateCandidate for an upstream (ASWF/Pixar/USD-WG) report

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions