New firmware is out with OpenMIOP v4 support

I wrote an OpenMIOP driver for Talos Linux and Debian/Ubuntu, enabling Ethernet communication between Mixtile Blade 3 nodes directly over the ClusterBox internal PCIe fabric.

Important: OpenMIOP requires the compatible ClusterBox BMC/OpenWrt firmware v0.1.0-rc.1 to be installed.

Talos v1.14.2 + OpenMIOP:

ClusterBox OpenWrt firmware:
https://github.com/roysbike/mixtile-clusterbox-mt7620a-openwrt/releases/download/v0.1.0-rc.1/mixtile-clusterbox-v120-v0.1.0-rc.1-sysupgrade-sd.bin

Tested on a 4-node Mixtile Blade 3 ClusterBox.

If anyone needs it — feel free to use it, test it, and contribute.



1 Like

Hi everyone!

I’ve released the first stable version of the complete OpenMIOP stack for Mixtile Blade 3 / Cluster Box!

This release includes updates for all three components:

  • OpenMIOP PCIe networking driver v0.3.0
  • Mixtile Talos v0.3.0
  • ClusterBox BMC OpenWrt firmware v0.1.0

What’s new and fixed

  • PCIe root complex hang fixed: Introduced a new push-mode architecture. The BMC no longer reads blade memory, preventing PCIe fabric lockups when a blade disconnects or its link drops.
  • Independent P2P recovery: Blades can now restore PCIe connectivity automatically after a Cluster Box reboot, typically within ~12 seconds, without waiting for the BMC.
  • NVMe recovery fixed: Cluster Box reboots previously caused NVMe drives to disappear due to shared PCIe reset signals. NVMe devices now recover automatically, preventing storage failures affecting ZFS and DRBD.
  • New BMC kernel module ( omi-rc ): Replaces the userspace helper with kernel-level push-mode support. The old helper remains available as a fallback.
  • Faster initialization: Fixed a timing issue that could delay blade activation by approximately 5 minutes.
  • Stalled peer recovery: Connections stalled immediately after establishment are automatically retried.
  • Improved blade management: Updated nodectl to properly release PCIe resources before reset, poweroff, or hard reboot.
  • Improved performance: Aggregate throughput increased from 22.2 Gbit/s to 25.0 Gbit/s across four blades.

Performance results

Tested with 4 × Mixtile Blade 3 (RK3588), Talos Linux v1.14.2, Kubernetes, Cozystack, and PCIe Gen3 x2:

  • Single TCP stream: 4.3–5.1 Gbit/s
  • 4 parallel streams: 8.1 Gbit/s
  • Bidirectional: 6.9 + 7.0 Gbit/s
  • 4-node simultaneous traffic: 25.0 Gbit/s aggregate

The Kubernetes cluster successfully recovered after BMC reboots, with NVMe storage operational, etcd 3/3 healthy, and no faulty DRBD resources.

Downloads

OpenMIOP v0.3.0 – PCIe networking driver

Mixtile Talos v0.3.0 – Talos Linux for Blade 3

ClusterBox BMC Firmware v0.1.0 – OpenWrt with OpenMIOP push-mode support

Important upgrade instructions

Upgrade the blades first, one at a time, and then update the Cluster Box BMC firmware.

The new BMC push-mode module requires OpenMIOP v0.3 on all blades.

BMC firmware updates are available through the LuCI web interface under System → Firmware Update.

The release maintains protocol v4 compatibility between v0.2 and v0.3 blades.

This is a major stability milestone for running Kubernetes and distributed storage over the Cluster Box PCIe fabric.

Feedback, testing, and contributions are welcome!

1 Like