Sim Multi-Device Isolation¶

Problem¶

When 2+ ChipWorkers run concurrently on the sim platform, host_runtime.so global state causes crashes. Root causes identified:

Task cookie core-ID collision: platform_set/get_cpu_sim_task_cookie() keys by (core_id, reg_task_id). Both instances allocate cores 0-8, so one instance consumes the other's cookies.
Affinity gate counter collision: platform_aicpu_affinity_gate() uses process-global atomic counters shared across all instances.
clear_cpu_sim_shared_storage() destroys the process-global pthread_key, breaking other running instances.

Solution: Fork+SHM Process Isolation¶

Each ChipWorker runs in its own forked Python child process. The OS provides full address-space isolation — all global state in host_runtime.so is automatically separate per process.

Main process (Scheduler)
  ├── WorkerThread[0] → shm mailbox → Child process 0 (ChipWorker[0])
  └── WorkerThread[1] → shm mailbox → Child process 1 (ChipWorker[1])
                                            └── dlopen(host_runtime.so)
                                                ← isolated globals

Communication uses a fixed-size (MAILBOX_SIZE-byte) shared-memory mailbox per chip — the same layout used for SUB-type workers. The parent copies ChipStorageTaskArgs into the mailbox; the child copies it to heap before calling run_runtime (sim runtime requires heap-backed args, not mmap-backed).

Why Not Fix the Globals¶

The global state in host_runtime.so spans multiple files (cpu_sim_context.cpp, platform_aicpu_affinity.cpp, l2_swimlane_collector_aicpu.cpp, device_log.cpp) and is deeply embedded in the AICPU/AICore thread model. Fixing each one individually is fragile. Process isolation solves all of them at once with zero platform code changes.

Files¶

File	Role
`src/common/hierarchical/worker_manager.{h,cpp}`	C++ `WorkerThread`: mailbox-IPC dispatcher for the forked chip child
`python/simpler/worker.py` (`_chip_process_loop`)	Python child: init ChipWorker, poll mailbox, run tasks