Hewlett Packard Enterprise Development LP

Hewlett-packard Development Company, L.p.

Other

Puneet Sharma

Sujata Banerjee

Jean Tourrilhes

Sung-Ju Lee

Diman Zad Tootaghaj

Faraz Ahmed

Lianjie Cao

Jung Gun Lee

Anu Mercian

Sujoy Basu

Junguk Cho

Parthasarathy Ranganathan

Vinay Saxena

Joon-Myung Kang

Huazhe Wang

Ying Zhang

Jeongkeun Lee

Zhikui Wang

Gene Cheung

Praveen Yalagandula

David B Lee

Ali Tariq

Tianxiong Ji

Terrel L Morris

Jeffrey Clifford Mogul

Yoshio Frank Turner

Charles F Clark

Priya Mahadevan

Michael Zayats

Wonho Kim

Vivek Adarsh

Junggun Lee

Derek Spock

Yadi Ma

Arun Raghuramu

Anand Mudgerikar

Bernardo A Huberman

Michael S Schlansker

Ajay Gupta

Kang Geun Shin

Ram Swaminathan

Xiaoyun Zhu

John Thomas Brassil

Jayaram Mudigonda

Alvin AuYoung

Julie Ward Drew

David Andrew Thomas

Raul Hernan Etkin

Taekyoung Kwon

Pramila Mullan

Arun Thulasi

Aisha Syed

Aditya Dhakal

Lucian Popa

Roque Luis Scheer

Vasu Sasikanth Sankhavaram

Amy Csizmar Dalal

Ali Munir

Sonia Fahmy

Mehrnaz Sharifian

Chaithan M Prakash

Alvin Auyoung

Amit Kumar Mondal

Shishir Garg

Sriharsha Gangam

Rita Wouhaybi

Sungjoon Choi

Brandon Heller

Joon Myung Kang

Christian Sieber

Mehdi Malboubi

Amit Kumar Sheoran

Amit Samanta

Badrinath Natarajan

Yunming Xiao

Yang Zhang

Zhen Lin

Freddy Chua

Dominik Stiller

Wenjie Lin

Subramoniam Iyer

Sarbajit Chatterjee

Deepti Sharma

Joseph P Chabarek

Dave Lenrow

Manish Jain

Sukesh Garg

David Lee

David Lenrow

A scheduling platform for scheduling serverless application tasks in persistent memory (PMEM) is provided. A profiler receives application requests from processes of serverless applications. The profiler categorizes the processes as persistent or non-persistent based on the application requests. A read/write batcher creates batches of the persistent requests including the read requests and write requests and assigns the batches to persistent memory banks. A scheduler creates a schedule of the batches to the persistent memory banks in a manner enabling optimization of job completion time.

Enabling persistent memory for serverless applications

Systems and methods are provided for improving autotuning procedures using stateless processing with a remote key-value store. For example, the system can implement a task launcher, a scheduler, and an agent to launch, schedule, and execute decomposed autotuning stages, respectively. The scheduling policy implemented by the scheduler may perform operations beyond a simple scheduling policy (e.g., a FIFO-based scheduling policy), which produces a high queuing delay. Compared to the traditional systems, by leveraging autotuning specific domain knowledge, queueing delay is reduced and resource utilization is improved.

Context-aware and stateless deep learning autotuning framework

Example implementations relate to scheduling of jobs for a plurality of graphics processing units (GPUs) providing concurrent processing by a plurality of virtual GPUs. According to an example, a computing system including one or more GPUs receives a request to schedule a new job to be executed by the computing system. The new job is allocated to one or more vGPUs. Allocations of existing jobs are updated to one or more vGPUs. Operational cost of operating the one or more GPUs and migration cost of allocating the new job are minimized and allocations of the existing jobs on the one or more vGPUs is updated. The new job and the existing jobs are processed by the one or more GPUs in the computing system.

Scheduling jobs on graphical processing units

In certain embodiments, a computer-implemented method includes: receiving, by a caching system plugin, a request to create a persistent volume for a container application instance; configuring, by the caching system plugin, a local cache volume on a host computing device; configuring, by the caching system plugin, a remote storage volume on a remote storage device; selecting, by a policy manager of the caching system plugin, a cache policy for the container application instance; creating, by the caching system plugin and from a cache manager, a virtual block device associated with the local cache volume, the remote storage volume, and the cache policy; and providing the virtual block device for use by the container application instance as the persistent volume.

Accelerating containerized applications with caching

Systems and methods are provided for effectuating a filtering technique that can enable available bandwidth (e.g., on a network path) to be estimated in the presence of moderate losses caused by certain queue management techniques. When packet losses exist on a network path due to certain types of packet queue transmission mechanisms, methods, or models, a bump detection algorithm (BDA) can be used to perform bandwidth estimation. When a pattern of packet loss is identified (from its signature) as being one with which BDA can be performed to accurately estimate available bandwidth on a network path, the BDA-based bandwidth estimate may be used to place/route and load balance network traffic, or otherwise used to engage in network traffic engineering, take other network-related action(s), or reported out. Otherwise, the bandwidth estimation is suppressed and not used.

Bandwidth estimate filtering based on packet loss patterns

A device and corresponding method are provided determining a consumed computing capacity of a first networking device exceeds the threshold for total capacity for processing monitoring data for a monitoring metric. An optimization engine determines a second networking device with unused computing capacity sufficient for processing the monitoring data generated by the first networking device. The optimization engine automatically moves the monitoring data for the monitoring metric generated by the first networking device to the second networking device and causes the second networking device to process the monitoring data.

Distributed network monitoring

A system maintains a queue for storing packets, which are enqueued at a tail of the queue and dequeued at a head of the queue. The system computes a queue utilization value, based on the packets stored in the queue. The system computes an excess amount value, based on the packets stored in the queue and previously tagged as excess packets. The system receives a first packet at the tail of the queue and determines whether a difference between the queue utilization value and the excess amount value exceeds a predetermined threshold. Responsive to determining that the difference exceeds the predetermined threshold, the system tags the first packet as an excess packet. Responsive to tagging the first packet as an excess packet, the system performs an operation associated with the first packet or a second packet at the head of the queue to reduce congestion.

Excess active queue management (AQM): a simple AQM to handle slow-start

In certain embodiments, a method includes receiving, at an interface of a Smart network interface card (SmartNIC) of a computing device, via a network, a network data unit; processing, by a data allocator of a SmartNIC subsystem of the SmartNIC, the network data unit to make a determination that data included in the network data unit is intended for processing by an accelerator of the computing device, wherein the accelerator is configured to execute a machine learning algorithm; storing, by the data allocator and based on the determination, the data in a local buffer of the SmartNIC subsystem; identifying, by the data allocator, a memory resource associated with the accelerator; and transferring the data from the local buffer to the memory resource.

Smart network interface card control plane for distributed machine learning workloads

Systems and methods are provided for optimally allocating resources used to perform multiple tasks/jobs, e.g., machine learning training jobs. The possible resource configurations or candidates that can be used to perform such jobs are generated. A first batch of training jobs can be randomly selected and run using one of the possible resource configuration candidates. Subsequent batches of training jobs may be performed using other resource configuration candidates that have been selected using an optimization process, e.g., Bayesian optimization. Upon reaching a stopping criterion, the resource configuration resulting in a desired optimization metric, e.g., fastest job completion time can be selected and used to execute the remaining training jobs.

Systems and methods of resource configuration optimization for machine learning workloads

Systems, methods, and computer-readable media are described for employing a machine learning-based approach such as adaptive Bayesian optimization to learn over time the most optimized assignments of incoming network requests to service function chains (SFCs) created within network slices of a 5G network. An optimized SFC assignment may be an assignment that minimizes an unknown objective function for a given set of incoming network service requests. For example, an optimized SFC assignment may be one that minimizes request response time or one that maximizes throughput for one or more network service requests corresponding to one or more network service types. The optimized SFC for a network request of a given network service type may change over time based on the dynamic nature of network performance. The machine-learning based approaches described herein train a model to dynamically determine optimized SFC assignments based on the dynamically changing network conditions.

Machine learning-based approaches for service function chain selection

Systems and methods are provided for updating resource allocation in a distributed network. For example, the method may comprise allocating a plurality of resource containers in a distributed network in accordance with a first distributed resource configuration. Upon determining that a processing workload value exceeds a stabilization threshold of the distributed network, determining a resource efficiency value of the plurality of resource containers in the distributed network. When a resource efficiency value is greater than or equal to the threshold resource efficiency value, the method may generate a second distributed resource configuration that includes a resource upscaling process, or when the resource efficiency value is less than the threshold resource efficiency value, the method may generate the second distributed resource configuration that includes a resource outscaling process. The resource allocation may transmit the second to update the resource allocation.

Network-aware resource allocation

Systems and methods are provided for improving autotuning procedures. For example, the system can implement a task launcher, a scheduler, and an agent to launch, schedule, and execute decomposed autotuning stages, respectively. The scheduling policy implemented by the scheduler may perform operations beyond a simple scheduling policy (e.g., a FIFO-based scheduling policy), which produces a high queuing delay. By leveraging autotuning specific domain knowledge, this may help reduce queuing delay and improve resource utilization that is otherwise found in traditional systems.

Deep learning autotuning task optimization

Systems and methods can be configured to determine a plurality of computing resource configurations used to perform machine learning model training jobs. A computing resource configuration can comprise: a first tuple including numbers of worker nodes and parameter server nodes, and a second tuple including resource allocations for the worker nodes and parameter server nodes. At least one machine learning training job can be executed using a first computing resource configuration having a first set of values associated with the first tuple. During the executing the machine learning training job: resource usage of the worker nodes and parameter server nodes caused by a second set of values associated with the second tuple can be monitored, and whether to adjust the second set of values can be determined. Whether a stopping criterion is satisfied can be determined. One of the plurality of computing resource configurations can be selected.

Example implementations relate to consensus protocols in a stretched network. According to an example, a distributed system includes continuously monitoring network performance and/or network latency among a cluster of a plurality of nodes in a distributed computer system. Leadership priority for each node is set based at least in part on the monitored network performance or network latency. Each node has a vote weight based at least in part on the leadership priority of the node. Each node's vote is biased by the node's vote weight. The node having a number of biased votes higher than a maximum possible number of votes biased by respective vote weights received by any other node in the cluster is selected as a leader node.

Leader election in a distributed system based on node weight and leadership priority based on network performance

Systems and methods are provided for available network bandwidth estimation using a one-way-delay noise filter with bump detection. The method includes receiving one-way delay measurements for each probe packet in a probe train sent over the telecommunications path; grouping the probe packets into a plurality of pairs based on the one-way delay measurements; for each pair, computing a respective noise threshold based on the one-way delay measurements of all the probe packets transmitted after a later-transmitted probe packet of the pair; selecting one of the pairs according to the noise thresholds and the one-way delay measurements for the probe packets of the pairs; and estimating the available bandwidth on the telecommunications path based on transmission times of the probe packets in the selected pair.

Available network bandwidth estimation using a one-way-delay noise filter with bump detection

Growing community of inventors

Palo Alto, CA, United States of America

Puneet Sharma