International Business Machines Corporation

Min Li

Valentina Salapura

John Alan Bivens

Ruchi Mahindru

Eugen Schenfeld

Harigovind Venkatraj Ramasamy

Yaoping Ruan

Minkyong Kim

Clifford A Pickover

Koushik Kumar Das

Li Feng Zhang

Jian Tan

Liangzhao Zeng

Shicong Meng

Yandong Wang

Nicholas C Fuller

Jenny S Li

Yu Deng

Rui Zhang

Yunyao Li

Lucian Popa

Prithviraj Sen

Theresa Y Tai

Kun Qian

Luna Xu

Quang C Duong

Ashish Kundu

Ramani Ranjan Routray

Yang Song

Richard Harold Boivie

Prasenjit Sarkar

Zhenhua Hu

Dinesh Kumar Subhraveti

Chung-hao Tan

Rakesh Jain

Mauricio Antonio Hernandez-Sherrington

Marina Danilevsky Hailpern

Sara A Noeman

Nikita Bhutani

Oluwatosin Alabi

Embodiments are provided herein for efficient component communication and resource optimization in a disaggregated computing system. A first set of computing elements are used as in line accelerators and a second set of the computing elements are used as block accelerators within the disaggregated computing system. A switching operation is dynamically performed between the first set of computing elements and the second set of computing elements to perform a workload by rewiring one of a plurality of links associated with respective ones of the first set of computing elements and the second set of computing elements.

Efficient component communication through accelerator switching in disaggregated datacenters

Methods, systems, and computer program products for learning models for entity resolution using active learning are provided herein. A computer-implemented method includes determining a set of data items related to a task associated with structured knowledge base creation, and outputting the set of data items to a user for labeling. Such a method also includes generating, based on a user-labeled version of the set of data items, a candidate model for executing the task, and one or more generalized versions of the candidate model. Additionally, such a method can also include generating a final model based on one or more iterations of analysis of the candidate model and analysis of the one or more generalized versions of the candidate model, and performing the task by executing the final model on one or more datasets.

Learning models for entity resolution using active learning

Various embodiments for optimizing memory bandwidth in a disaggregated computing system, by a processor device, are provided. Respective memory devices are assigned to respective processor devices in the disaggregated computing system, the disaggregated computing system having at least a pool of the memory devices and a pool of the processor devices. An analytic function is performed on data resident in the pool of the memory devices using memory bandwidth not currently committed to a primary compute task.

Opportunistic data analytics using memory bandwidth in disaggregated computing systems

Embodiments for optimizing dynamic resource allocations for storage-dependent workloads in a disaggregated computing environment. A new workload is assigned to a subset of a plurality of processors, the subset of processors assigned a subset of a plurality of memory devices associated with a plurality of storage devices. A determination is made that the new workload is categorized as a storage-dependent workload having a storage need which can be met primarily by a subset of the storage devices after having identified whether data requests associated with the new workload can be satisfied by the subset of memory devices. Pursuant to determining the new workload is the storage-dependent workload, a storage related action is proactively performed to increase efficiency of the new workload prior to commencement of a performance of the new workload.

Optimizing dynamic resource allocations for storage-dependent workloads in disaggregated data centers

Respective memory devices are assigned to respective processor devices in a disaggregated computing system, the disaggregated computing system having at least a pool of the memory devices and a pool of the processor devices. An iterative learning algorithm is used to define data boundaries of a dataset for performing an analytic function on the dataset simultaneous to a primary compute task, unrelated to the analytic function, being performed on the dataset in the pool of memory devices using memory bandwidth not currently committed to the primary compute task, thereby efficiently employing the unused memory bandwidth to prevent underutilization of the pool of memory devices.

Server resources in a data center are disaggregated into shared server resource pools, including an accelerator (e.g., FPGA) pool. Servers are constructed dynamically, on-demand and based on workload requirements, by allocating from these resource pools. According to this disclosure, accelerator utilization in the data center is managed proactively by assigning accelerators 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 accelerators for data analytic workloads. The approach thus provides for a 'micro-service' enabling data analytic workloads to automatically and transparently use FPGA 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 FPGAs to use, and then during runtime auto-scales the FPGAs based on workload.

Utilizing accelerators to accelerate data analytic workloads in disaggregated systems

Embodiments for replicating data in a disaggregated computing system. A memory pool is allocated, where the memory pool includes allocated memory elements at a first site and allocated memory elements at a second site. The allocated memory elements are mapped at the first site to the allocated memory elements at the second site. A replication operation is initiated to mirror data stored within the allocated memory elements at the first site to the allocated memory elements at the second site. The allocated memory elements at the first site are directly connected through an independent networking connection to the allocated memory elements at the second site such that the replication operation is processed exclusively through compute resources at the first site.

Replicating workload and state data for disaster recovery in disaggregated datacenters

A computing problem management method, system, and non-transitory computer readable medium, include detecting an impending problem of the computing system, spawning a plurality of replicas when detecting the impending problem, introducing a plurality of new impending problems on the plurality of replicas for the launching to launch versions of an action to resolve, and learning a version that resolves each of the plurality of new impending problems.

System, method and recording medium for antifragile computing problem management

Embodiments for optimizing dynamic resource allocations in a disaggregated computing environment. A new workload is assigned to a subset of a plurality of processors, the subset of processors assigned a subset of a plurality of cache devices. A determination is made that the new workload is categorized as a cache-friendly workload having a memory need which can be met primarily by the subset of cache devices by identifying that underlying data necessitated by the new workload resides primarily within the subset of cache devices. Pursuant to determining the new workload is the cache-friendly workload, a cache related action is performed to increase performance of the new workload executed by the subset of processors and commensurately executes additional workloads performed by other ones of the plurality of processors within the disaggregated computing environment.

Optimizing dynamical resource allocations for cache-friendly workloads in disaggregated data centers

A new approach to resiliency management is provided in a data center wherein servers are constructed dynamically, on-demand and based on workload requirements and a tenant's resiliency requirements by allocating resources from these pools. In this approach, a set of functionally-equivalent 'interchangeable compute units' (ICUs) are composed of resources from resource pools that have been extended to include not only different resource types (CPU, memory, accelerators), but also resources of different specifications (specs) and flavors. As a workload is being processed, the health or status of the resources are monitored. Upon a performance issue or failure event, a resiliency manager can swap out a current ICU and replace it with a functionally-equivalent ICU. Preferably, individual ICUs are hosted on one of: resources of a same type each with different specifications, and resources of a same type and specification and different flavors. The approach enables failure independence in a disaggregated environment.

Building a highly-resilient system with failure independence in a disaggregated compute environment

Embodiments for preemptive substitution of resources in a disaggregated computing environment. Failure patterns and mitigation actions are analyzed for specific failures of respective resources within the disaggregated computing environment. Responsive to determining a failure threshold has been reached for a first resource of a first type of the respective resources, a mitigation action is performed according to the analyzed failure patterns. A result of the mitigation action is determined and the result is used to improve the failure pattern analyzation.

Preemptive resource replacement according to failure pattern analysis in disaggregated data centers

Embodiments are provided herein for efficient component communication and resource optimization in a disaggregated computing system. A general purpose link is provided between a plurality of devices in the disaggregated computing system. The general purpose link is used to connect the plurality of devices which respectively comprise different types of devices. A first data object is stored within one of the plurality of devices according to a past use or an expected use of the data object by at least a respective one of a set of processors within the disaggregated computing system, and the general purpose link is dynamically rewired according to a location of the first data object and a location of the at least one of the respective set of processors.

Efficient component communication through resource rewiring in disaggregated datacenters

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

Embodiments are provided herein for efficient component communication and resource optimization in a disaggregated computing system. A general purpose link is provided to connect a computing element to a plurality of other computing elements of the disaggregated computing system. The general purpose link is dynamically switched between a plurality of different hardware protocols to communicate with the other computing elements, where respective ones of the other computing elements comprise different types of hardware elements.

Efficient component communication through protocol switching in disaggregated datacenters

Various embodiments for performing hardware upgrades in a disaggregated computing environment. A workload is run on a disaggregated computing system while providing a new component to at least one of a plurality of component pools used by the disaggregated computing system. Point-to-point circuit wire level switching is used to switch the disaggregated system from an assigned component residing in a first of the plurality of component pools to the new component residing in a second of the plurality of component pools without interrupting the running workload.

Growing community of inventors

San Jose, CA, United States of America

Min Li