Closed-Loop Thermal-Power Matched Workload Scheduling System for Immersion Cooled Computing Infrastructure

Abstract

Closed-Loop Thermal-Power Matched Workload Scheduling System for Immersion Cooled Computing Infrastructure In existing high-density computing infrastructures, workload scheduling systems operate independently of the thermal management and power distribution layers. This structure leads to hotspot formation, inefficient controlled power reduction, and inefficient use of cooling capacity. Especially in immersion cooled systems, the exclusion of the flow infrastructure from scheduling decisions creates a significant performance and reliability gap. This invention presents a workload scheduling system that combines thermal status, power consumption, and immersion flow data in a closed loop. The system derives zonal trends by receiving power, temperature, and flow telemetry from computing zones, predictively redistributes workloads according to these trends, and sends control signals directly to the immersion flow subsystem.The timing system both monitors and manages the immersion flow infrastructure; this bidirectional relationship forms the fundamental architectural feature of the system. Unlike instantaneous threshold-based reactive approaches, the invention incorporates a slope analysis-based predictive decision-making mechanism, includes power supply unit conditions in timing decisions, and preferably includes a cluster-level orchestration layer supporting multi-rack coordination.

Description

1 TARIFF Closed-Loop Thermal-Power for Immersion Cooled Computing Infrastructure Matched Workload Scheduling System TECHNICAL FIELD 0 This invention is suitable for high-density computing infrastructures, especially liquid immersion. with cooling systems in equipped artificial intelligence and high-performance computing (HPC) rack assemblies for use The invention relates to an improved workload scheduling system. The invention concerns scheduling. their decisions thermal status, power consumption and immersion flow data are combined in a closed loop. an infrastructure It relates to orchestration architecture. STATE OF THE ART 0. Workload scheduling in today's high-density computing infrastructures. thermal systems Management systems and power distribution infrastructure are independent layers. It is being designed. In this architecture, the timer only monitors the availability of computing resources. in evaluation; cooling capacity, immersion flow condition, and power supply unit (PSU) conditions timing They are not included in the decisions. Known thermal awareness timing systems typically use instantaneous temperature thresholds. based on its values It adopts a reactive model. In this approach, the processor or accelerator temperature of its components Frequency reduction or load transfer when exceeding a predetermined threshold of 35. interventions such as 2 40 triggers are encountered. However, this method cannot predict thermal trends in advance. cooling infrastructure 45 is unable to be actively managed and is being used with the power supply unit capacity. timing the situation This is not reflected in the 50 decisions. 55 This deficiency becomes even more pronounced in immersion cooling environments. Immersion systems, Compared to 60 air-cooled systems, it exhibits a much more dynamic thermal response; flow rate, liquid Temperature and regional heat density increase rapidly depending on the workload profile. It can vary. 70 In contrast, existing timing systems allow for the controllable immersion flow infrastructure. a system component It considers it not as a single point (75), but as an independent external system. This situation refers to a hotspot. formation, inefficient 80% controlled power reduction, unbalanced power distribution, and effective cooling capacity. unusable It leads to 85 problems. TECHNICAL PROBLEMS THAT THE INVENTION AIMS TO SOLVE 0 The fundamental technical problem solved by this invention is immersion-cooled computing. working in their infrastructure load timer with thermal status, power consumption and immersion flow data in a closed loop It is a performance and reliability gap resulting from their failure to work together. Within the scope of the invention, the timing system is solely responsible for managing computational loads. not remaining but immersion by also monitoring the operating parameters of the flow infrastructure, including thermal, power and operational. between the load It forms a closed-loop coupling. EXPLANATION OF THE IMAGES 3 Figure 1: Shows the general system architecture. Timing control layer. (100), telemetry interface (110), analysis engine (120), flow control interface (130), calculation zones (300), immersion flow sub system (200), power supply units (400) and cluster-level orchestration layer (500) The data and control flows between them are shown. Figure 2: Shows the closed-loop data stream (600). Telemetry data to the timing system transmission and sending of control signals to the immersion flow subsystem schematically It is presented. 35 Figure 3: Predictive intervention mechanism and workload redistribution It shows. 40 Zonal temperature slope curves (310, 320), threshold value (Tm), predictive intervention trigger point Comparative load distribution before (F1) and after (F2) intervention with 45 (M1) It is presented. Figure 4: Shows multi-rack coordination. Multiple racks within timing control between the 55 layers (100) and the cluster-level orchestration layer (500) policy and resources The coordination of the 60 envelopes is presented schematically. ANNOUNCEMENT OF THE INVENTION Timing system within the scope of the invention (100), high-intensity immersion cooled computing load distribution in their infrastructures, thermal conditions and power conditions, and enclosed online It is structured to manage. Telemetry interface (110), power from one or more computing zones (300) consumption, temperature and 4 It continuously receives at least one piece of immersion stream data. That data is: instantaneous measurement The values ​​can be in the form of time series recordings or a combination thereof. The analysis engine (120) processes the data it receives from the telemetry interface to produce zonal thermal and power trends It derives 35 trends. Trend derivation involves time series trend analysis and moving averages. account, 40 derivative-based methods or functionally equivalent approaches through 45 can be implemented. Preferably an analysis engine, a machine learning model. may include; however, invention 50 is not tied to a specific model architecture. 55 Timing control layer (100), receives trend outputs from the analysis engine working based on It dynamically redistributes the loads between the calculation zones. Predictive intervention 65 trigger points (M1), before critical thermal or power conditions occur is being determined and the work The redistribution of the 70 loads is initiated at this point. 75 Flow control interface (130), receives decisions from the timing control layer immersion flow sub It transmits the control signal to the 80 system (200). These signals are the pump flow rate. (210), valve position affecting at least one of the 85 (220) and bypass line routing (230) It is configurable. 90 The structure in question monitors both the timing system's immersion flow subsystem. and the one he managed It consists of 95 closed loops (600). 0 Ideally, the system should have multiple racks with varying workload, cooling capacity, and power envelope a cluster-level orchestration layer that manages in a coordinated manner (500) may include. HOW THE INVENTION WAS APPLIED TO INDUSTRY 0 Invention; immersion-cooled artificial intelligence computing racks, high-performance computation sets It is applicable in data center infrastructures. The scheduling system is card-based. management controller firmware, orchestration software, or data center control plane It can be deployed. Thanks to its hardware-independent structure, the system supports immersion cooling from different manufacturers. with its products They can work together.

Claims

6 REQUESTS 1. A study for a high-density immersion-cooled computing infrastructure. The load is a timing system, and the system in question is; — power consumption, temperature, and flow from multiple computing zones at least one telemetry interface that receives at least one of its data (110), — at least deriving zonal thermal or power trends from the data in question an analysis engine (120), — dynamically adjusting workloads according to these trends at least one timing control layer (100) that distributes — to transmit control signals to the immersion flow subsystem (200) At least one structured flow control interface (130) characterized by its inclusion of the said timing control layer (100) It receives telemetry from the immersion flow subsystem (200) and sends it to the immersion flow subsystem. (200) closed loop (600) thermal-power-working load to which it sends control signals Matched timing system.

2. According to claim 1, the system and the said analysis engine (120) is zonal thermal or Time series slope analysis, moving average, instead of instantaneous threshold comparison of power trends. Characterized by deriving it based on at least one of the methods of averaging or derivative calculations. the system that was used.

3. The system according to claim 1 or 2 and the flow control interface in question (130), pump flow rate (210), valve position (220) and bypass line routing (230) at least characterized by being structured to control one independently system.

4. The system is based on any of claims 1 to 3, and the timing control in question is... layer (100), zonal temperature or strength trend predictive intervention trigger to initiate workload redistribution before reaching point (M1) A predictive workload handling system characterized by its configuration. 7 5. The system is based on any of claims 1 to 4, and the timing control in question is... instantaneous load capacity of the layer (100), power supply units (400), efficiency curve and include at least one instance of usage history in workload deployment decisions a system characterized by its structure.

6. The system is based on any of claims 1 to 5, and the timing control in question is... layer (100) to balance the in-rack thermal distribution of the workloads Simultaneously recalculating between multiple calculation zones (300) A system characterized by its structure designed for positioning.

7. The system according to any of claims 1 to 6, and that system has more than one among the multiple racks, at least one of the following: workload, cooling capacity and power envelope a cluster-level orchestration structured to manage in a coordinated manner The system is characterized by containing more than (500) layers.

8. The system is defined according to any of claims 1 to 7, and the analysis engine in question... (120), at least one machine learning model that learns from past thermal and power data. including and timing zonal thermal or power trend predictions of the model in question. The system is characterized by its configuration to provide control to the layer (100).

9. The system is based on any of requests 1 through 8, and the telemetry in question is... the pressure, volumetric flow rate and inlet of the interface (110) in the immersion flow subsystem (200). to obtain sensor outputs (240) measuring at least one of the outlet temperature differences structuring and timing control layer (100) decisions of the said outputs A system characterized by its input generation.

Images