Advanced Micro Devices Corporation

Leonardo Piga

Indrani Paul

Mauricio Breternitz, Jr

Wei T Huang

Larry David Hewitt

Anil Harwani

Karthik Rao

Sriram Sambamurthy

Aaron Joseph Grenat

Dana Glenn Lewis

Wonje Choi

Sriram Sundaram

Brian J Kocoloski

Yasuko Eckert

Manish Arora

Patryk Kaminski

Abhinandan Majumdar

Samuel Lawrence Wasmundt

Md Abdullah Shahneous Bari

A system and method for updating power supply voltages due to variations from aging are described. A functional unit includes a power supply monitor capable of measuring power supply variations in a region of the functional unit. An age counter measures an age of the functional unit. A control unit notifies the power supply monitor to measure an operating voltage reference. When the control unit receives a measured operating voltage reference, the control unit determines an updated age of the region different from the current age based on the measured operating voltage reference. The control unit updates the age counter with the corresponding age, which is younger than the previous age in some cases due to the region not experiencing predicted stress and aging. The control unit is capable of determining a voltage adjustment for the operating voltage reference based on an age indicated by the age counter.

Operating voltage adjustment for aging circuits

A cluster compute server stores different types of data at different storage volumes in order to reduce data duplication at the storage volumes. The storage volumes are categorized into two classes: common storage volumes and dedicated storage volumes, wherein the common storage volumes store data to be accessed and used by multiple compute nodes (or multiple virtual servers) of the cluster compute server. The dedicated storage volumes, in contrast, store data to be accessed only by a corresponding compute node (or virtual server).

Storage location assignment at a cluster compute server

Systems, apparatuses, and methods for achieving balanced execution in a multi-node cluster through runtime detection of performance variation are described. During a training phase, performance counters and an amount of time spent waiting for synchronization is monitored for a plurality of tasks for each node of the multi-node cluster. These values are utilized to generate a model which correlates the values of the performance counters to the amount of time spent waiting for synchronization. Once the model is built, the values of the performance counters are monitored for a period of time at the start of each task, and these values are input into the model. The model generates a prediction of whether a given node is on the critical path. If the given node is predicted to be on the critical path, the power allocation of the given node is increased.

Achieving balanced execution through runtime detection of performance variation

Systems, apparatuses, and methods for performing temperature-aware task scheduling and proactive power management. A SoC includes a plurality of processing units and a task queue storing pending tasks. The SoC calculates a thermal metric for each pending task to predict an amount of heat the pending task will generate. The SoC also determines a thermal gradient for each processing unit to predict a rate at which the processing unit's temperature will change when executing a task. The SoC also monitors a thermal margin of how far each processing unit is from reaching its thermal limit. The SoC minimizes non-uniform heat generation on the SoC by scheduling pending tasks from the task queue to the processing units based on the thermal metrics for the pending tasks, the thermal gradients of each processing unit, and the thermal margin available on each processing unit.

Temperature-aware task scheduling and proactive power management

Systems, apparatuses, and methods for managing variations among nodes in parallel system frameworks. Sensor and performance data associated with the nodes of a multi-node cluster may be monitored to detect variations among the nodes. A variability metric may be calculated for each node of the cluster based on the sensor and performance data associated with the node. The variability metrics may then be used by a mapper to efficiently map tasks of a parallel application to the nodes of the cluster. In one embodiment, the mapper may assign the critical tasks of the parallel application to the nodes with the lowest variability metrics. In another embodiment, the hardware of the nodes may be reconfigured so as to reduce the node-to-node variability.

Managing variations among nodes in parallel system frameworks

Systems, apparatuses, and methods for managing cluster-level performance variability without a centralized controller are described. Each node of a multi-node cluster tracks a maximum and minimum progress across the plurality of nodes for a workload executed by the cluster. Each node also tracks its local progress on its current task. Each node also utilizes a comparison of the local progress to reported maximum and minimum progress across the cluster to identify a critical, or slow, node and whether to increase or reduce an amount of power allocated to the node. The nodes append information about the maximum and minimum progress to messages sent to other nodes to report their knowledge of maximum and minimum progress with other nodes. A node updates its local information if the node receives a message from another node with more up-to-date information about the state of progress across the cluster.

Managing cluster-level performance variability without a centralized controller

Systems, apparatuses, and methods for accelerating page migration using a two-level bloom filter are disclosed. In one embodiment, a system includes a GPU and a CPU and a multi-level memory hierarchy. When a memory request misses in a first memory, the GPU is configured to check a first level of a two-level bloom filter to determine if a page targeted by the memory request is located in a second memory. If the first level of the two-level bloom filter indicates that the page is not in the second memory, then the GPU generates a page fault and sends the memory request to a third memory. If the first level of the two-level bloom filter indicates that the page is in the second memory, then the GPU sends the memory request to the CPU.

Page migration acceleration using a two-level bloom filter on high bandwidth memory systems

Systems, apparatuses, and methods for implementing performance estimation mechanisms are disclosed. In one embodiment, a computing system includes at least one processor and a memory subsystem. During a characterization phase, the system utilizes a memory intensive workload to detect when the memory subsystem reaches its saturation point. Then, the system collects performance counter values during a sampling phase of a target application to determine the memory bandwidth. If the memory bandwidth is greater than the saturation point, then the system generates a prediction of the memory time which is based on a ratio of the memory bandwidth over the saturation point. Otherwise, if the memory bandwidth is less than the saturation point, the system assumes memory time is constant versus processor frequency. Then, the system uses the memory time and an estimate of the compute time to estimate a phase time for the target application at different processor frequencies.

Bandwidth-aware multi-frequency performance estimation mechanism

Systems, apparatuses, and methods for balancing computation and communication power in power constrained environments. A data processing cluster with a plurality of compute nodes may perform parallel processing of a workload in a power constrained environment. Nodes that finish tasks early may be power-gated based on one or more conditions. In some scenarios, a node may predict a wait duration and go into a reduced power consumption state if the wait duration is predicted to be greater than a threshold. The power saved by power-gating one or more nodes may be reassigned for use by other nodes. A cluster agent may be configured to reassign the unused power to the active nodes to expedite workload processing.

Balancing computation and communication power in power constrained clusters

Various datacenter or other computing center control apparatus and methods are disclosed. In one aspect, a method of computing is provided that includes defining plural processor performance bins where each processor performance bin has a processor performance state. At least one processor is assigned to each of the plural processor performance bins. Processor performance metrics of at least one of the processors are monitored while the at least one of the processors executes an incoming task. Processor power is modeled based on the monitored performance metrics. Future incoming tasks are assigned to one of the processor performance bins based on the modeled processor power.

Combined dynamic and static power and performance optimization on data centers

A method is performed in a computing system that includes a plurality of processing nodes of multiple types configurable to run in multiple performance states. In the method, an application executes on a thread assigned to a first processing node. Power and performance of the application on the first processing node is estimated. Power and performance of the application in multiple performance states on other processing nodes of the plurality of processing nodes besides the first processing node is also estimated. It is determined that the estimated power and performance of the application on a second processing node in a respective performance state of the multiple performance states is preferable to the power and performance of the application on the first processing node. The thread is reassigned to the second processing node, with the second processing node in the respective performance state.

Thread assignment for power and performance efficiency using multiple power states

A device may receive information that identifies a first task to be processed, may determine a performance metric value indicative of a behavior of a processor while processing a second task, and may assign, based on the performance metric value, the first task to a bin for processing the first task, the bin including a set of processors that operate based on a power characteristic.

Growing community of inventors

Austin, TX, United States of America

Leonardo Piga