Distributed resource-aware training of machine learning pipelines

Abstract

A method, structure, and computer system for predicting pipeline training requirements. An example embodiment can include receiving one or more worker node features from one or more worker nodes, extracting one or more pipeline features from one or more pipelines to be trained, and extracting one or more dataset features from one or more datasets used to train the one or more pipelines. An example embodiment can also include predicting, based on one or more models that relate the one or more worker node features, the one or more pipeline features, and the one or more dataset features to one or more resources, an amount of the one or more resources needed by each of the one or more worker nodes to train the one or more pipelines using the one or more datasets. Finally, an example embodiment can include identifying a worker node of the one or more worker nodes that requires the least amount of the one or more resources for training the one or more pipelines.

Description

BACKGROUND

[0001] Exemplary embodiments relate generally to machine learning pipelines, and more particularly to distributed resource-aware training of automated machine learning pipelines.

[0002] Automated machine learning generally involves training several pipelines. These pipelines are generated by a pipeline optimization framework that explores and arranges various transformers and estimators to find the best pipeline. Training these pipelines in a distributed environment is a challenging problem because it is not possible to know beforehand how much time and resources will be spent training a pipeline. As a result, pipelines cannot currently be intelligently assigned to worker nodes. SUMMARY

[0003] Exemplary embodiments disclose a method, structure, and computer system for pipeline training. Exemplary embodiments can include receiving one or more worker node features from one or more worker nodes, extracting one or more pipeline features from one or more pipelines to be trained, and extracting one or more dataset features from one or more datasets used to train the one or more pipelines. Exemplary embodiments can also include predicting, based on one or more models that relate the one or more worker node features, the one or more pipeline features, and the one or more dataset features to one or more resources, an amount of the one or more resources needed by each of the one or more worker nodes to train the one or more pipelines using the one or more datasets. Finally, exemplary embodiments can include identifying a worker node of the one or more worker nodes that requires a least amount of the one or more resources for training the one or more pipelines. BRIEF DESCRIPTION OF DRAWINGS

[0004] The following detailed description will best be understood in conjunction with the accompanying drawings, of which:

[0005] Figure 1 An exemplary schematic diagram of a pipeline training system 100 is depicted in accordance with exemplary embodiments.

[0006] Figure 2 An exemplary flowchart 200 illustrating the operation of a joint optimizer 132 of the pipeline training system 100 is depicted in accordance with exemplary embodiments.

[0007] Figure 3 An example illustrating the operation of the joint optimizer 132 of the pipeline training system 100 is depicted in accordance with exemplary embodiments.

[0008] Figure 4 An exemplary schematic diagram of a pipeline training system 100 is depicted in accordance with exemplary embodiments. Figure 1An exemplary block diagram of hardware components of a pipeline training system 100.

[0009] Figure 5 A cloud computing environment is depicted in accordance with an exemplary embodiment.

[0010] Figure 6 An abstraction model layer is depicted in accordance with an exemplary embodiment.

[0011] The drawings are not necessarily to scale. The drawings are merely schematic representations, not intended to portray specific parameters of the exemplary embodiments. The drawings are intended to depict only typical exemplary embodiments. In the drawings, like numbers refer to like elements. DETAILED DESCRIPTION

[0012] Disclosed herein are specific embodiments of the claimed structures and methods; however, it is understood that the disclosed embodiments are merely illustrative of the claimed structures and methods that can be practiced in various forms. The exemplary embodiments are illustrative only, however, can be practiced in many different forms and should not be construed as limited to the exemplary embodiments set forth herein. Rather, these exemplary embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the exemplary embodiments to those skilled in the art. In the description, details of well-known features and techniques can be omitted to avoid unnecessarily obscuring the presented embodiments.

[0013] References in the specification to "one embodiment," "an embodiment," "exemplary embodiment," etc. indicate that the embodiment described can include a particular feature, structure, or characteristic, but every embodiment can not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Furthermore, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the knowledge of those skilled in the art to effect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described.

[0014] In order not to obscure the presentation of the exemplary embodiments, in the following detailed description some processing steps or operations that are known in the art can have been combined together in order to present and for illustration purposes and in some cases can not be described in detail. In other cases, some processing steps or operations that are known in the art can not be described at all. It should be understood that the following description focuses on the distinguishing features and elements of the various exemplary embodiments.

[0015] Automated machine learning generally involves training several pipelines. These pipelines are generated by a pipeline optimization framework that explores and arranges various transformers and estimators to find the best pipeline. Training these pipelines in a distributed environment is a challenging problem because it is not possible to know beforehand how much time and resources will be spent training a pipeline. As a result, pipelines cannot currently be intelligently assigned to worker nodes.

[0016] A machine learning pipeline is a sequence of operations (such as data preprocessing, outlier detection, feature engineering, etc.) followed by an estimator. Such a pipeline can be managed by an algorithm called a joint optimizer that is responsible for 1) assembling various pipelines together; 2) training each of the assembled pipelines using cross-validation; 3) analyzing the results; and 4) deciding which pipelines to execute next. The joint optimizer can continue this process until it finds the best pipeline. Among the tasks performed by the joint optimizer, training each of the assembled pipelines using cross-validation is the most computationally expensive step, and is the focus of the invention claimed herein, which discloses a novel approach to resource-aware training of machine learning pipelines.

[0017] There exist methods for training pipelines, however each technique has its own drawbacks. For example, in the current AutoAI® system (AutoAI is a registered trademark of IBM Corporation), all assembled pipelines are executed on a single worker node. In another example, a knapsack / bin-packing technique uses the length of a pipeline to quantify the pipeline. Alternatively, a greedy bin-packing approximation algorithm, for example, the first-fit bin-packing, packs items of different volumes into a finite number of bins or containers, each bin or container having a fixed given volume in a way that minimizes the number of bins used. There also exists a round robin technique in which pipelines are distributed in a certain repeating order.

[0018] However, the aforementioned methods have several problems. First, they do not estimate how much time a particular pipeline training will take because various machine learning algorithms have different complexities. Second, techniques such as round robin do not take into account the busyness of machines and whether they are able to handle the training of a particular pipeline.

[0019] The invention claimed herein overcomes the deficiencies of the aforementioned methods by using a data-driven approach that looks at past performance and builds a predictive model of pipeline training. More specifically, the claimed invention can predict the resource requirements of training a pipeline and continuously learns to improve the prediction using data executed by previous pipelines.

[0020] Figure 1A pipeline training system 100 according to exemplary embodiments is depicted. According to exemplary embodiments, the pipeline training system 100 can include one or more worker nodes 120A-K and a pipeline training server 130, all of which can be interconnected via a network 108. While the programming and data of exemplary embodiments can be stored and accessed remotely via the network 108 across several servers, the programming and data of exemplary embodiments can alternatively or additionally be stored locally on as few as one physical computing device or in other computing devices besides those depicted.

[0021] In exemplary embodiments, the network 108 can be a communication channel capable of transferring data between connected devices. In exemplary embodiments, the network 108 can be the Internet, representing the worldwide collection of networks and gateways that support communications between devices connected to the Internet. Furthermore, the network 108 can utilize various types of connections, such as wireline, wireless, fiber optic, etc., which can be implemented as intranets, local area networks (LANs), wide area networks (WANs), or combinations thereof. In further embodiments, the network 108 can be a Bluetooth network, a Wi-Fi network, or combinations thereof. The network 108 can operate in frequencies including 2.4 GHz and 5 GHz Internet, near field communication, Z-Wave, Zigbee, etc. In yet further embodiments, the network 108 can be a telecommunications network for facilitating two-party or multi-party voice calls, including landline networks, wireless networks, closed networks, satellite networks, or combinations thereof. In general, the network 108 can represent any combination of connections and protocols that will support communications between connected devices.

[0022] In exemplary embodiments, the one or more worker nodes 120A-K can each be an enterprise server, a laptop computer, a notebook, a tablet computer, a netbook computer, a personal computer (PC), a desktop computer, a server, a personal digital assistant (PDA), a flip phone, a key phone, a smart phone, a mobile phone, a virtual device, a thin client, an IoT device, or any other electronic device or computing system capable of sending data to and receiving data from other computing devices. Each of the one or more worker nodes 120A-K can be configured to train one or more machine learning pipelines. In exemplary embodiments, it is assumed that each of the one or more worker nodes 120A-K has access to the same dataset, and each pipeline can be trained on a single worker node 120 of the one or more worker nodes 120A-K. Reference is made to Figure 4 The one or more worker nodes 120A-K are described in greater detail as hardware implementations, reference is made to Figure 5 described as part of a cloud implementation, and / or reference is made to Figure 6 described as being processed with a functional abstraction layer.

[0023] In example embodiments, the pipeline training server 130 includes a joint optimizer 132, a performance predictor 134, and a load balancer 136. The pipeline training server 130 can be an enterprise server, a laptop, a notebook, a tablet, a netbook, a personal computer (PC), a desktop computer, a server, a personal digital assistant (PDA), a rotary telephone, a push-button telephone, a smart phone, a mobile phone, a virtual appliance, a thin client, an IoT device, or any other electronic device or computing system capable of sending data to and receiving data from other computing devices. While the pipeline training server 130 is shown as a single device, in other embodiments, the pipeline training server 130 can be composed of clusters or multiple computing devices working together or independently in a modular fashion, etc. Reference is made to Figure 4 The pipeline training server 130 is described in more detail as a hardware implementation, reference is made to Figure 5 as part of a cloud implementation, and / or reference is made to Figure 6 as being processed with a functional abstraction layer.

[0024] In embodiments, the joint optimizer 132 can be a software and / or hardware program capable of assembling various pipelines together, training each assembled pipeline using cross-validation, analyzing the results, and deciding which pipelines to execute next. Specifically, and with respect to training each of the assembled pipelines using cross-validation, the joint optimizer 132 can also be capable of receiving one or more worker node heartbeat features and extracting one or more pipeline features and one or more dataset features. The joint optimizer 132 can also be capable of predicting the required pipeline training resources by producing one or more models and selecting one or more worker nodes based on the models. Finally, the joint optimizer 132 can be capable of adjusting the models based on a feedback loop. Reference is made to Figures 2-6 The joint optimizer 132 is described in more detail.

[0025] The performance predictor 134 can be a software and / or hardware program capable of predicting the amount of resources required (e.g., the required time) to train a pipeline. Reference is made to Figures 2-6 The performance predictor 134 is described in more detail.

[0026] The load balancer 136 can be a software and / or hardware program capable of distributing a set of tasks over a set of resources to make their overall processing more efficient. Reference is made to Figures 2-6 The performance predictor 134 is described in more detail.

[0027] Figure 2An exemplary flowchart 200 depicts the operation of the joint optimizer 132 of a pipeline training system 100 according to an exemplary embodiment. In the following detailed description, it is assumed that each of one or more worker nodes 120A-K has access to a dataset and can train any pipeline. Although only one dataset is referenced, those skilled in the art will understand that the claimed invention is applicable to any number of datasets.

[0028] The performance predictor 134 may receive heartbeat features from one or more worker nodes 120A-K (step 202). In an embodiment, the heartbeat features quantify the busyness and power of each of the worker nodes 120A-K and may include features extracted from worker node data, such as the number of CPUs and their cores, the number of GPUs and their cores, CPU and GPU utilization, CPU and GPU memory and swap usage, and the output of the vmstat and iostat commands. In an example embodiment, the heartbeat features may be received by the load balancer 136 of the joint optimizer 132 via network 108 at predefined intervals. However, if the joint optimizer 132 fails to receive heartbeat features from any of the one or more worker nodes 120A-K, the joint optimizer 132 marks the one or more unresponsive worker nodes 120A-K as unresponsive and omits training predictions for them until heartbeat features are received again. In an embodiment, the joint optimizer 132 may be trained based on heartbeat features collected herein, as well as pipeline features and dataset features described below, to determine which of the worker nodes 120A-K can train the pipeline model with the least amount of resources.

[0029] To better illustrate the operation of the joint optimizer 132, reference is now made to... Figure 3 The illustrated example depicts a joint optimizer 132 configured to train two pipelines: 1) Principal Component Analysis (PCA) on a Random Forest (RF); and 2) Outlier Detection (OD) on a Support Vector Machine (SVM) over any of four worker nodes 120A, 120B, 120C, and 120D. The joint optimizer 132 receives corresponding heartbeat features H1, H2, H3, and H4, which include the number of CPUs and their cores, the number of GPUs and their cores, CPU and GPU utilization, CPU and GPU memory and swap usage, and the outputs of the vmstat and iostat commands.

[0030] The joint optimizer 132 can extract pipeline features (step 204). In embodiments, the pipeline features can include the types of estimators, the types of pre-processors, the types of feature engineering, and their parameter settings, and the joint optimizer 132 can extract the pipeline features via the performance predictor 134 by performing a sparse encoding of the one or more pipelines to be trained. Specifically, the joint optimizer 132 can set the entries for the used transformers and estimators to 1, and the unused transformers and estimators to 0. The joint optimizer 132 can also encode the parameters of each individual pipeline component to create an encoded version of the pipeline. As described above, the joint optimizer 132 can utilize the pipeline features along with the heartbeat features and the dataset features to build one or more models that can predict which of the worker nodes 120A-K can train the pipeline using the least amount of resources.

[0031] Returning to the illustrative example introduced above and depicted by Figure 3 the first pipeline and the second pipeline described above.

[0032] The joint optimizer 132 can extract one or more dataset features (step 206). In embodiments, the dataset features capture various high-level dimensions of the dataset, and the extracted dataset features can include the number of data points, the number of features, the number of cross-validation folds, the number of features with categorical values, the number of features with real values, the number of missing values, and the number of sparse values. As described above, the joint optimizer 132 can utilize the dataset features along with the heartbeat features and the pipeline features to build one or more models that can predict which of the worker nodes 120A-K can train the pipeline using the least amount of resources.

[0033] Continuing with the example introduced above and depicted by Figure 3 the joint optimizer 132 can extract dataset features from the dataset, including the number of data points from the dataset, the number of features, the number of CV folds, the number of features with categorical values, the number of features with real values, the number of missing values, and the number of sparse features.

[0034] The joint optimizer 132 can predict the required pipeline training resources (step 208). In embodiments, the joint optimizer 132 can predict, via the performance predictor 134, a performance measure required by each of the one or more worker nodes 120A-K to train the respective pipeline. Such required performance measures can include time, power consumption, peak memory usage, peak CPU utilization, etc. In embodiments, the joint optimizer 132 can use a machine learning or deep learning model to predict the required pipeline training resources. In such embodiments, the joint optimizer 132 can receive the pipeline features, dataset features, and heartbeat features as input, train a model from these inputs using algorithms such as random forest, GBM, logistic regression, deep neural networks, and autoencoders. The model captures how each of the aforementioned features affects the amount of resources required by each of the worker nodes 120A-K to train the pipeline. Once the model is trained, the joint optimizer 132 is able to apply the trained model to new sets of heartbeat features, pipeline features, and dataset features in order to output predicted performance measures regarding each worker node training the pipeline. The joint optimizer 132 can then adjust the model based on comparing the predicted performance measures to actual performance measures, as will be described in more detail below.

[0035] Further Description Figure 3 With the previously introduced example illustrated, the joint optimizer 132 predicts that the worker nodes 120A-D can train the first pipeline in A, B, C, and D minutes, respectively, and train the second pipeline in W, X, Y, and Z minutes, respectively. The joint optimizer 132 can also predict peak CPU usage, etc.

[0036] The federated optimizer 132 can select worker nodes (step 210). In embodiments, the federated optimizer 132 can select at least one of the one or more worker nodes 120A-K for execution of the pipelines based on the predicted required pipeline training resources. In embodiments, the federated optimizer 132 can select the one or more worker nodes 120A-K based on an epsilon-greedy or multi-armed bandit problem approach. In the epsilon-greedy or epsilon-greedy approach, the federated optimizer 132 selects the best predicted worker nodes 120A-K with a probability of (1 - e) and a random worker node 120A-K with a probability of e. For example, the federated optimizer 132 can then select the top three predicted workers (1 - e), where e is set between 0.05 and 0.1, and select three random workers with a probability of e. In most cases, the federated optimizer 132 will select the best predicted worker nodes 120A-K, however, there is a small chance that the federated optimizer 132 selects a random worker node 120A-K. In the multi-armed bandit approach, the federated optimizer 132 can train the model by first selecting three random workers and executing the pipelines for n iterations, e.g., n = 1000. Once the federated optimizer 132 has made enough evaluations to identify the best performing worker nodes 120A-K, the federated optimizer 132 can then send all of the pipelines to the best performing worker nodes 120A-K. If the federated optimizer 132 determines that the performance of the best performing worker nodes 120A-K is thus reduced, the federated optimizer 132 can then return to randomly identifying the best performing worker nodes 120A-K and repeat the process.

[0037] With reference to the previously introduced example, the federated optimizer 132 selects worker node 120A to train the first pipeline and selects worker node 120B to train the second pipeline.

[0038] The federated optimizer 132 can adjust the model based on a feedback loop (step 212). In embodiments, the federated optimizer 132 can adjust the model by determining a loss after training the one or more pipelines by the one or more worker nodes 120A-K. Specifically, the federated optimizer 132 can receive the predicted required pipeline training resources and the actual pipeline training resources consumed and then output a loss based on equation 1:

[0039] Loss = (Predicted Resources - Actual Resources) 2 Equation 1

[0040] The federated optimizer 132 can then backpropagate the loss using gradient descent.

[0041] Summarized by Figure 3In the depicted example, the joint optimizer 132 adjusts the model based on identifying a loss while the first pipeline is trained at worker node 120A and the second pipeline is trained at worker node 120B.

[0042] Key features of the claimed invention include performance improved over time through backpropagation of loss, generation of various training data using a multi-armed bandit approach, and use of a random forest system that uses continuous prediction, collection of training data, learning, and better prediction.

[0043] Those skilled in the art will appreciate that the foregoing invention can be implemented within the field of, for example, automated machine learning and artificial intelligence, distributed pipeline training, fast and continuous data science pipeline joint optimization, and load balancing of machine learning training pipelines. The claimed system can be implemented in, for example, Kubernetes and Docker platforms, where one or more worker nodes 120A-K are Docker containers, and the heartbeat feature can be obtained using kubectl. Further, the machine learning / deep learning model and the e-greedy model can run on a primary or microservice in the cluster. Further, the system can be scaled using an automated scanner or manually created containers using the output of the ML / DL model, and the containers can write the output of each pipeline training to a shared cloud or network storage, such that the model can pick up the data from the storage and gradient descent their loss.

[0044] Figure 4 An example illustrating the operation of the joint optimizer 132 of the pipeline training system 100 is depicted in accordance with example embodiments.

[0045] Figure 1 A block diagram of a device used within the pipeline training system 100 in accordance with example embodiments is depicted. Figure 4 It should be understood that Figure 5 Only a high-level description of one implementation is provided for the sake of brevity. Many modifications can be made to the described environment.

[0046] The device used herein can include one or more processors 02, one or more computer-readable RAMs 04, one or more computer-readable ROMs 06, one or more computer-readable storage media 08, device drivers 12, read / write drives or interfaces 14, network adapters or interfaces 16, all of which are interconnected by a communication structure 18. The communication structure 18 can be implemented using any architecture designed for passing data and / or control information between processors (such as microprocessors, communications and network processors, etc.), system memory, peripheral devices, and any other hardware components within a system.

[0047] One or more operating systems 10 and one or more application programs 11 are stored on one or more computer-readable storage media 08 for execution by one or more processors 02 via one or more respective RAMs 04 (which typically include cache memory). In the illustrated embodiment, each computer- readable storage medium 08 can be a magnetic disk storage device of internal hard disk drive, a CD-ROM, DVD, memory stick, magnetic tape, magnetic disk, optical disk, semiconductor memory device such as RAM, ROM, EPROM, flash memory, or any other computer-readable tangible storage device that can store the computer programs and digital information.

[0048] The devices used herein can also include a read / write drive or interface 14 to read from and write to one or more portable computer-readable storage media 26. Application programs 11 on the devices can be stored on one or more portable computer-readable storage media 26, read via the respective read / write drive or interface 14, and loaded into the respective computer-readable storage media 08.

[0049] The devices used herein can also include a network adapter or interface 16, such as a TCP / IP adapter card or wireless communication adapter (such as a 4G wireless communication adapter using OFDMA technology). Application programs 11 on the computing devices can be downloaded to the computing devices from an external computer or external storage device via a network (for example, the Internet, a local area network or other wide area network or wireless network) and the network adapter or interface 16. From the network adapter or interface 16, the programs can be loaded into the computer-readable storage media 08. The network can comprise copper wires, optical fibers, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers.

[0050] The devices used herein can also include a display screen 20, a keyboard or keypad 22, and a computer mouse or touchpad 24. Device drivers 12 interface to the display screen 20 for imaging, to the keyboard or keypad 22, to the computer mouse or touchpad 24, and / or to the display screen 20 for pressure-sensing alphanumeric character input and user selection. The device drivers 12, read / write drives or interfaces 14, and network adapters or interfaces 16 can include hardware and software (stored on computer-readable storage media 08 and / or ROM 06).

[0051] The programs described herein are identified based upon the application for which they are implemented in a particular one exemplary embodiment. It should be understood, however, that any particular program term(s) used herein is / are for convenience, and thus is / are not intended to limit the example embodiments to a specific application wherein the program term(s) is / are used. All such embodiments are intended to mean "any of the example embodiments."

[0052] Based on the foregoing, a computer system, method, and computer program product are disclosed. Numerous modifications and alternatives can be made, however, without departing from the scope of the exemplary embodiments. Accordingly, the exemplary embodiments have been disclosed by way of example and not limitation.

[0053] It should be appreciated that while the present disclosure includes a detailed description of cloud computing, implementation of the teachings recited herein are not limited to a cloud computing environment. Rather, exemplary embodiments are capable of implementation in conjunction with any other type of computing environment now known or later developed.

[0054] Cloud computing is a model of service delivery for enabling convenient, on-demand network access to a shared pool of configurable computing resources (e.g. networks, network bandwidth, servers, processing, memory, storage, applications, virtual machines, and services) that can be rapidly provisioned and released with minimal management effort or interaction with a provider of the service. This cloud model can include at least five characteristics, at least three service models, and at least four deployment models.

[0055] The characteristics are as follows:

[0056] On-demand self-service: cloud consumers can unilaterally provision computing capabilities, such as server time and network storage, as needed automatically without requiring human interaction with the service's provider.

[0057] Broad network access: capabilities are available over a network and accessed through standard mechanisms that promote use by heterogeneous thin or thick client platforms (e.g., mobile phones, laptops, and PDAs).

[0058] Resource pooling: the provider's computing resources are pooled to serve multiple consumers using a multi-tenant model, with different physical and virtual resources dynamically assigned and reassigned according to demand. There is a sense of location independence in that the consumer generally has no control or knowledge over the exact location of the provided resources but can be able to specify location at a higher level of abstraction (e.g., country, state, or datacenter).

[0059] Rapid elasticity: capabilities can be rapidly and elastically provisioned, in some cases automatically, to quickly scale out and rapidly scale in. To the consumer, the provider's ability to provision capabilities on-demand and in near real-time allows for rapidly adjusting to changing business demand.

[0060] Measured service: cloud systems automatically control and optimize resource use by leveraging utilization of resources in an efficient manner, for example, in response to varying demand for the services. Resource usage can be monitored, controlled, and reported, providing transparency for both the provider and the consumer of the utilized services.

[0061] Service models are as follows:

[0062] Software as a Service (SaaS): the capability provided to the consumer is to use the provider's applications running on a cloud infrastructure. The applications are accessible from various client devices through a thin client interface such as a web browser (e.g., web-based e-mail). The consumer does not manage or control the underlying cloud infrastructure including network, servers, operating systems, storage, or even individual application capabilities, with the possible exception of limited user-specific application configuration settings.

[0063] Platform as a Service (PaaS): the capability provided to the consumer is to deploy onto the cloud infrastructure consumer-created or acquired applications created using programming languages and tools supported by the provider. The consumer does not manage or control the underlying cloud infrastructure including networks, servers, operating systems, or storage, but has control over the deployed applications and possibly application hosting environment configurations.

[0064] Infrastructure as a Service (IaaS): the capability provided to the consumer is to provision processing, storage, networks, and other fundamental computing resources where the consumer is able to deploy and run arbitrary software, which can include operating systems and applications. The consumer does not manage or control the underlying cloud infrastructure but has control over operating systems, storage, deployed applications, and possibly limited control of select networking components (e.g., host firewalls).

[0065] Deployment models are as follows:

[0066] Private cloud: the cloud infrastructure is operated solely for the organization. It can be managed by the organization or a third party and can exist on-premises or off-premises.

[0067] Community cloud: the cloud infrastructure is shared by several organizations and supports mission-critical production workloads. It can be managed by the organizations or a third party and can exist on-premises or off-premises.

[0068] Public cloud: the cloud infrastructure is made available to general public or a large industry group and is owned by an organization selling cloud services.

[0069] Hybrid cloud: the cloud infrastructure is a composition of two or more types of cloud (private, community, or public) that remain unique clouds but are bound together by standardized or proprietary technologies that enable data and application portability.

[0070] Cloud computing environments are service-oriented, focusing on statelessness, loose coupling, modularity, and semantic interoperability. At the core of cloud computing is an infrastructure comprising a network of interconnected nodes.

[0071] Now for reference Figure 5 The illustration depicts a cloud computing environment 50. As shown, the cloud computing environment 50 includes one or more cloud computing nodes 40 to which local computing devices used by cloud consumers can communicate, such as personal digital assistants (PDAs) or cellular phones 54A, desktop computers 54B, laptop computers 54C, and / or automotive computer systems 54N. The nodes 40 can communicate with each other. They can be physically or virtually grouped (not shown) in one or more networks, such as private clouds, community clouds, public clouds, or hybrid clouds, or combinations thereof, as described above. This allows the cloud computing environment 50 to provide infrastructure, platform, and / or software as a service, without requiring cloud consumers to maintain resources on their local computing devices. It should be understood that... Figure 6 The types of computing devices 54A-N shown are for illustrative purposes only, and computing node 40 and cloud computing environment 50 can communicate with any type of computerized device over any type of network and / or network-addressable connection (e.g., using a web browser).

[0072] Now for reference Figure 5 This demonstrates a cloud computing environment of 50 ( Figure 6 This provides a set of functional abstractions. It should be understood beforehand that... ​ The components, layers, and functions shown are for illustrative purposes only, and exemplary embodiments are not limited thereto. As depicted, the following layers and corresponding functions are provided:

[0073] The hardware and software layer 60 includes hardware and software components. Examples of hardware components include: a host 61; a server 62 based on a RISC (Reduced Instruction Set Computer) architecture; a server 63; a blade server 64; a storage device 65; and network and networking components 66. In some embodiments, software components include network application server software 67 and database software 68.

[0074] The virtualization layer 70 provides an abstraction layer from which the following examples of virtual entities can be provided: virtual server 71; virtual storage 72; virtual network 73, including virtual private network; virtual application and operating system 74; and virtual client 75.

[0075] In one example, management layer 80 can provide the functions described below. Resource provisioning 81 provides dynamic procurement of computing resources and other resources that are utilized to perform tasks within the cloud computing environment. Metering and Pricing 82 provide cost tracking as resources are utilized within the cloud computing environment, and billing or invoicing for consumption of these resources. In one example, these resources can include application software licenses. Security provides identity verification for cloud consumers and tasks, as well as protection for data and other resources. User portal 83 allows for access to and task computing environment for consumers and system administrators. Service level management 84 provides cloud computing resource allocation and management such that required service levels are met. Service Level Agreement (SLA) planning and fulfillment 85 provide pre-arrangement for, and procurement of, cloud computing resources according to an SLA.

[0076] Workloads layer 90 provides examples of functionality for which the cloud computing environment can be utilized. Examples of workloads and functions which can be provided from this layer include: mapping and navigation 91; software development and lifecycle management 92; virtual classroom education delivery 93; data analysis processing 94; transaction processing 95; and, pipeline training processing 96.

[0077] Example embodiments can be a system, a method, and / or a computer program product at any possible technical detail level of integration. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present application.

[0078] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0079] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network can comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions to storage media within the respective computing / processing device for execution by a processor.

[0080] Computer readable program instructions for carrying out operations of the present application can be assembly instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, configuration data for an integrated circuit, or source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like, and procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate array (FPGA), or programmable logic array (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present application.

[0081] Aspects of the present application are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.

[0082] These computer-readable program instructions may be provided to a processor of a computer or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / actions specified in one or more blocks of a flowchart and / or block diagram. These computer-readable program instructions may also be stored in a computer-readable storage medium that can direct a computer, programmable data processing apparatus, and / or other devices to operate in a particular manner, such that the computer-readable storage medium in which the instructions are stored includes an article of writing comprising instructions for implementing aspects of the functions / actions specified in one or more blocks of a flowchart and / or block diagram.

[0083] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer-implemented process, such that the instructions, which execute on the computer, other programmable apparatus or other device, perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0084] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of instructions comprising one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions indicated in the blocks may not occur in the order indicated in the figures. For example, two blocks shown consecutively may actually be implemented as a single step, executed simultaneously, substantially simultaneously, with partial or complete time overlap, or these blocks may sometimes be executed in reverse order, depending on the functions involved. It will also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented by a dedicated hardware-based system that performs the specified function or action or executes a combination of dedicated hardware and computer instructions.

Claims

1. A computer-implemented method for predicting pipeline training requirements, the method comprising: Receive one or more worker node characteristics from one or more worker nodes; Extract one or more pipeline features from one or more pipelines to be trained; Extract one or more dataset features from one or more datasets used to train the one or more pipelines; Based on one or more models that relate the features of the one or more worker nodes, one or more pipeline features, and one or more dataset features to the one or more resources, predict the amount of one or more resources required for each of the one or more worker nodes to train the one or more pipelines using the one or more datasets; as well as Identify the one or more worker nodes that require the minimum amount of one or more resources to train the one or more pipeline worker nodes. The one or more pipeline features include the type of estimator, the type of preprocessor, the type of feature engineering and its parameter settings, and the one or more dataset features include the number of data points, the number of features, the number of cross-validation folds, the number of features with categorical values, the number of features with real values, the number of missing values ​​and the number of sparse values.

2. The method according to claim 1, further comprising: Determine the actual amount of resources required for the worker nodes to train the one or more pipelines; as well as The one or more models are adjusted by comparing the predicted resource quantity with the actual resource quantity.

3. The method according to claim 1, wherein the one or more worker node characteristics respectively include CPU and the number of cores therein, GPU and the number of cores therein, CPU utilization, GPU utilization, CPU memory, GPU memory, CPU and GPU swapping, and the output of the vmstat and iostat commands.

4. The method according to claim 1, wherein, The one or more models are trained using random forest, GBM, logistic regression, deep neural networks, and autoencoders.

5. The method according to claim 1, wherein, The one or more resources include training time, power consumption, peak CPU utilization, and peak memory utilization.

6. A computer program product for predicting pipeline training requirements, the computer program product comprising: One or more non-transitory computer-readable storage media and program instructions stored on the one or more non-transitory computer-readable storage media capable of executing a method, the method comprising: Receive one or more worker node characteristics from one or more worker nodes; Extract one or more pipeline features from one or more pipelines to be trained; Extract one or more dataset features from one or more datasets used to train the one or more pipelines; Based on one or more models that relate the features of the one or more worker nodes, one or more pipeline features, and one or more dataset features to the one or more resources, predict the amount of one or more resources required for each of the one or more worker nodes to train the one or more pipelines using the one or more datasets; and Identify the one or more worker nodes that require the minimum amount of one or more resources to train the one or more pipeline worker nodes. The one or more pipeline features include the type of estimator, the type of preprocessor, the type of feature engineering and its parameter settings, and the one or more dataset features include the number of data points, the number of features, the number of cross-validation folds, the number of features with categorical values, the number of features with real values, the number of missing values ​​and the number of sparse values.

7. The computer program product according to claim 6, further comprising: Determine the actual amount of resources required for the worker nodes to train the one or more pipelines; as well as The one or more models are adjusted by comparing the predicted resource quantity with the actual resource quantity.

8. The computer program product according to claim 6, wherein, The characteristics of the one or more worker nodes include CPUs and the number of cores therein, GPUs and the number of cores therein, CPU utilization, GPU utilization, CPU memory, GPU memory, CPU and GPU swapping, and the output of the vmstat and iostat commands.

9. The computer program product according to claim 6, wherein, The one or more models are trained using random forest, GBM, logistic regression, deep neural networks, and autoencoders.

10. The computer program product according to claim 6, wherein, The one or more resources include training time, power consumption, peak CPU utilization, and peak memory utilization.

11. A computer system for predicting pipeline training requirements, the system comprising: One or more computer processors, one or more computer-readable storage media, and program instructions stored on the one or more computer-readable storage media, the program instructions being executable by at least one of the one or more processors capable of performing a method comprising: Receive one or more worker node characteristics from one or more worker nodes; Extract one or more pipeline features from one or more pipelines to be trained; Extract one or more dataset features from one or more datasets used to train the one or more pipelines; Based on one or more models that relate the features of the one or more worker nodes, one or more pipeline features, and one or more dataset features to the one or more resources, predict the amount of one or more resources required for each of the one or more worker nodes to train the one or more pipelines using the one or more datasets; and Identify the one or more worker nodes that require the minimum amount of one or more resources to train the one or more pipeline worker nodes. The one or more pipeline features include the type of estimator, the type of preprocessor, the type of feature engineering and its parameter settings, and the one or more dataset features include the number of data points, the number of features, the number of cross-validation folds, the number of features with categorical values, the number of features with real values, the number of missing values ​​and the number of sparse values.

12. The computer system according to claim 11, further comprising: Determine the actual amount of resources required for the worker nodes to train the one or more pipelines; as well as The one or more models are adjusted by comparing the predicted resource quantity with the actual resource quantity.

13. The computer system according to claim 11, wherein, The characteristics of the one or more worker nodes include CPUs and the number of cores therein, GPUs and the number of cores therein, CPU utilization, GPU utilization, CPU memory, GPU memory, CPU and GPU swapping, and the output of the vmstat and iostat commands.

14. The computer system according to claim 11, wherein, The one or more models are trained using random forest, GBM, logistic regression, deep neural networks, and autoencoders.

Images