mixed-models machine

Overview

mixed-models uses two device adapters in one machine: two gem5 MI300X performance models and two QEMU edu functional devices exported over vfio-user, sharing a 40 GiB memory pool. It shows that one machine can combine PCI devices from different sources.

Topology

# resources/machines/mixed-models.yaml
version: 1
imports:
  - vhosts/qemu.yaml
  - devices/gpu/mi300x.yaml
  - devices/pci/qemu-edu.yaml
machines:
  - id: m0
    vhost: {use: vhost-x86-rocm}
    devices:
      - {use: mi300x, count: 2}
      - {use: qemu-edu, count: 2}

  • Machine count: one (m0); the host uses vhost-x86-rocm.

  • Device identifiers: mi300x-0, mi300x-1, qemu-edu-0, qemu-edu-1.

  • qemu-edu declares no private memory region; the pool accumulates only 2 × 16 GiB of VRAM.

Hardware resources

Resource

Value

Source

Machine

1 (m0)

resources/machines/mixed-models.yaml

vHost profile

vhost-x86-rocm: q35, kvm, 8 GiB RAM, 4 vCPU

resources/vhosts/qemu.yaml

gem5 devices

2 × mi300x, 16 GiB VRAM and 40 CU each

resources/devices/gpu/mi300x.yaml

QEMU devices

2 × qemu-edu with args.device set to {driver: edu}

resources/devices/pci/qemu-edu.yaml

Memory pool

40 GiB = 8 GiB RAM + 32 GiB VRAM

crates/cosim-topology/src/layout.rs

Region offsets

mi300x-0.vram at 8 GiB; mi300x-1.vram at 24 GiB

Declaration order

Processes

1 host process + 2 gem5 processes + 2 QEMU device processes

Session plan

Software stack

Layer

Component

Description

Guest OS

Ubuntu 24.04.2 LTS, kernel 6.8.0-124-generic

vhost-x86-rocm selects the image recipe through vhost.image

GPU driver

amdgpu with fixed parameters

Applies to the gem5 MI300X devices only

Functional device

QEMU edu

Test PCI device with no matching guest driver

Device models

gem5 performance model + QEMU functional model

QEMU devices start with -machine x-remote and x-vfio-user-server

Guest initialization

cosim-gpu-initialize.service

Handles AMD PCI devices only

Running this machine

# Validate and preview
just run-cli validate resources/machines/mixed-models.yaml  --assets
just run-topology resources/machines/mixed-models.yaml --dry-run
# Start a session
just run-topology resources/machines/mixed-models.yaml

  • QEMU device processes need the QEMU built by just build-qemu with vfio-user support.

  • The vHost starts only after all four devices are ready and have handed over their leases; on shutdown the vHost stops first and the devices exit afterwards.

  • Lifecycle and readiness probes are described in Session lifecycle.

Verification

  • The plan output is one host process and four device processes, a 40 GiB pool and four vfio-user sockets.

  • The sample is loaded by the render test; the MI300X model and the vfio-user device path each have separate test jobs.

  • This particular sample, which combines both adapters, has no dedicated end-to-end CI job.

Limitations

  • Both adapters have their own tests, but end-to-end guest behavior of this composition is established by local runs.

  • qemu-edu is a test device with no guest driver; the guest only sees an unbound PCI device.

  • Running needs Linux glibc and setpriv; booting a guest needs KVM.

  • The 40 GiB pool plus four device processes are resource-heavy; check host memory and /dev/shm capacity before running.