International Business Machines Corporation

University of Michigan

Koushik Kumar Das

Valentina Salapura

Ruchi Mahindru

John Alan Bivens

Min Li

Eugen Schenfeld

Harigovind Venkatraj Ramasamy

Yaoping Ruan

Ching-Te Kent Chuang

Richard B Brown

Shih-Hsien Lo

Rajiv V Joshi

Jeffrey W Sleight

Keunwoo Kim

Peter William Cook

For power management in a disaggregated computing system, upon detecting an electrical power reduction to the disaggregated computing system, respective workloads being performed by respective processors within a set of processors are prioritized according to a service level agreement (SLA) of the respective workloads. Backup electrical power is dynamically allocated to the respective processors performing the respective workloads based upon a priority of the SLA of the respective workloads.

SLA-based backup power management during utility power interruption in disaggregated datacenters

Server resources in a data center are disaggregated into shared server resource pools. Servers are constructed dynamically, on-demand and based on workload requirements, by allocating from these resource pools. A disaggregated compute system of this type keeps track of resources that are available in the shared server resource pools, and it manages those resources based on that information. Each server entity built is assigned with a unique server ID, and each resource that comprises a component thereof is tagged with the identifier. As a workload is processed by the server entity, its composition may change, e.g. by allocating more resources to the server entity, or by de-allocating resources from the server entity. Workload requests are associated with the unique server ID for the server entity. When a workload request is received at a resource, it matches its unique server ID to that of the request before servicing the request.

Specifying a disaggregated compute system

For measuring component utilization in a computing system, a server energy utilization reading of a statistical significant number of servers out of a total number of servers located in the datacenter is obtained by measuring, at predetermined intervals, a collective energy consumed by all processing components within each server. The collective energy is measured by virtually probing thereby monitoring an energy consumption of individual ones of all the processing components to each collect an individual energy utilization reading, where the individual energy utilization reading is aggregated over a predetermined time period to collect an energy consumption pattern associated with the server utilization reading.

Measuring utilization of resources in datacenters

For power management in a disaggregated computing system, initial electrical power levels are distributed thereby allocating a voltage and a clock speed to each one of a set of processor cores in the disaggregated computing system. The voltage and the clock speed of respective processor cores within the set of processor cores are adjusted according to a workload priority of respective workloads performed by each respective one of the processor cores, wherein the workload priority is assigned based upon a service level agreement (SLA).

SLA-based power management in disaggregated computing systems

For measuring component utilization in a system having a plurality of subsystems, an energy consumption of each of the plurality of subsystems is monitored whether or not each subsystem performs at least a portion of an overall computation. Respective workloads are classified based upon an energy consumption pattern associated with the monitored energy consumption of each of the plurality of subsystems.

Energy consumption as a measure of utilization and work characterization in a system

For power management in a computing system, component utilization is dynamically managed within the computing system according to a calculated aggregate energy consumed by each one of a set of processors. Each of a plurality of energy factors are measured individually between each one of the set of processors to accumulate the calculated aggregate energy in real time.

Power management and utilization detection of computing components

A memory management service occupies a configurable portion of an overall memory system in a disaggregate compute environment. The service provides optimized data organization capabilities over the pool of real memory accessible to the system. The service enables various types of data stores to be implemented in hardware, including at a data structure level. Storage capacity conservation is enabled through the creation and management of high-performance, re-usable data structure implementations across the memory pool, and then using analytics (e.g., multi-tenant similarity and duplicate detection) to determine when data organizations should be used. The service also may re-align memory to different data structures that may be more efficient given data usage and distribution patterns. The service also advantageously manages automated backups efficiently.

Multi-tenant memory service for memory pool architectures

For power management in a disaggregated computing system, a set of initial electrical power levels are allocated to a set of processor cores according to a predicted desired workload, where the set of initial power levels aggregate to an initial collective contracted power level. Electrical power is dynamically allocated to respective processor cores within the set of processor cores to produce a capacity to execute a collective demanded workload while maintaining the electrical power to the set of processor cores to an approximately constant electrical power level within a threshold of the initial collective contracted electrical power level.

Power management in disaggregated computing systems

Various embodiments for allocating resources in a disaggregated cloud computing environment, by a processor device, are provided. Respective members of a pool of hardware resources are assigned to each one of a plurality of tenants based upon a classification of the respective members of the pool of hardware resources. The respective members of the pool of hardware resources are assigned to each one of the plurality of tenants independently of a hardware enclosure in which the respective members of the pool of hardware resources are physically located.

Constructing computing systems with flexible capacity of resources using disaggregated systems

Server resources in a data center are disaggregated into shared server resource pools. Servers are constructed dynamically, on-demand and based on workload requirements and a tenant's resiliency requirements (e.g., as specified in an SLA), by allocating from these resource pools. A disaggregated compute system of this type keeps track of resources that are available in the shared server resource pools, and it manages those resources based on that information and the health of the resources. As a workload is processed by the server entity and component resources fail, the server entity composition is changed, e.g. by allocating other resources to the server entity, or by transitioning to other server entities, to ensure that a resiliency requirement is maintained.

Specifying a highly-resilient system in a disaggregated compute environment

Server resources in a data center are disaggregated into shared server resource pools, including a graphics processing unit (GPU) pool. Servers are constructed dynamically, on-demand and based on workload requirements, by allocating from these resource pools. According to this disclosure, GPU utilization in the data center is managed proactively by assigning GPUs to workloads in a fine granularity and agile way, and de-provisioning them when no longer needed. In this manner, the approach is especially advantageous to automatically provision GPUs for data analytic workloads. The approach thus provides for a 'micro-service' enabling data analytic workloads to automatically and transparently use GPU resources without providing (e.g., to the data center customer) the underlying provisioning details. Preferably, the approach dynamically determines the number and the type of GPUs to use, and then during runtime auto-scales the GPUs based on workload.

Dynamically provisioning and scaling graphic processing units for data analytic workloads in a hardware cloud

A new power gating structure with robust data retention capability using only one single double-gate device to provide both power gating switch and virtual supply/ground diode clamp functions. The scheme reduces the transistor count, area, and capacitance of the power gating structure, thus improving circuit performance, power, and leakage. The scheme is compared with the conventional power gating structure via mixed-mode physics-based two-dimensional numerical simulations. Analysis of virtual supply/ground bounce for the proposed scheme is also presented.

High-density low-power data retention power gating with double-gate devices

Disclosed are a multi-threshold CMOS circuit and a method of designing such a circuit. The preferred embodiment combines an MTCMOS scheme and a hybrid SOI-epitaxial CMOS structure. Generally, the logic transistors (both nFET and pFET) are placed in SOI, preferably in a high-performance, high density UTSOI; while the headers or footers are made of bulk epitaxial CMOS devices, with or without an adaptive well-biasing scheme. The logic transistors are based on (100) SOI devices or super HOT, the header devices are in bulk (100) or (110) pFETs with or without an adaptive well biasing scheme, and the footer devices are in bulk (100) NFET with or without an adaptive well biasing scheme.

Growing community of inventors

Yorktown Heights, NY, United States of America

Koushik Kumar Das