Redfish Practices and Learnings in Immersion Cooling Management

Author: Mohan Kumar, Intel Fellow, Data Center AI Group Collection: 2023 Global Summit - Cooling Environments

Summary

Mohan Kumar (Intel Fellow) presented Intel's practical experience building a Redfish-based management interface for an actual immersion cooling tank deployment. Unlike the companion presentation on the Redfish data model framework (Jeff Otter), this talk focused on the two specific engineering challenges that had to be solved when applying the model to real immersion cooling hardware: CDU redundancy modeling and sensor location disambiguation. Both solutions were implemented, tested on a real Intel immersion tank, and contributed back to the DMTF Redfish Forum.

The core challenge of immersion cooling management is that the physical reality of the system is topologically different from the logical view exposed by Redfish. Sensors physically sit inside the immersion tank (submerged in coolant), but they are electrically connected to and logically managed by the PLC inside the CDU — so from a Redfish perspective, the sensors appear to be "attached" to the CDU, not the tank. A service technician dispatched based on the Redfish data would be directed to the CDU when the sensor is actually at the tank. Similarly, mission-critical deployments require multiple CDUs providing redundant cooling to a single tank, but Redfish's existing redundancy model (originally built for redundant power supplies) needed to be adapted for cooling connector redundancy.

For redundancy, Kumar borrowed the existing Redfish redundancy construct used for power supplies and applied it to CoolantConnectors. The new property CoolantConnectorRedundancy in ThermalSystem describes which CoolantConnectors form a redundancy set (with properties: RedundancyEnabled: true, RedundancySet listing the odata IDs of the redundant connectors, Mode: Sharing, MinNumNeeded: 1, MaxNumSupported: 2). For sensor location, Kumar used two mechanisms: (1) RelatedItem — a pointer in the sensor's JSON record from the CDU's PLC to the immersion tank chassis where the sensor physically resides; and (2) Sensor Excerpts — a Redfish mechanism that allows a sensor defined in one location (the CDU) to project a mirrored copy of its reading into another resource (the Tank), so that programmatic queries on the Tank resource see the sensor data as if it were locally attached.

Intel verified the complete solution on an actual immersion tank and confirmed the model works. The extensions were submitted to and accepted by the DMTF Redfish Forum. A white paper with full model details was in preparation at the time of the presentation.