GRID RESOUCE MANAGEMENT Pham Thanh Toan Trinh Quoc Huy Pham The Anh 10/18/2015 1

Outline  Introduction to Grid Resource Management  Hierarchical Model  Abstract Owner Model  Economy/Market Model 210/18/2015

Introduction Resource types 310/18/2015  Physical resource: computer, disk, database, networks, scientific instruments.  Logical resource: Executing applications, Complex workflows…

Introduction Definition 410/18/2015  The term resource management refers to the operations used to control how capabilities provided by Grid resources and services are made available to other entities, whether users, applications or services.

Introduction Definition 510/18/2015  At the heart of the Grid is the ability to discover, allocate, and negotiate the use of network accessible capabilities, be they computational services offered by a computer, application services offered by a piece of software, bandwidth delivered on a network, or storage space provided by a storage system.

Introduction Requirements 610/18/2015  Establish a mutual agreement between resource provider and resource consumer.  Encompass a wide range of different scenarios: task submission, workload management, on-demand access, co-scheduling, resource brokering,…  Encompass not only a commitment to perform a task but also commitments to level of performance, or quality of services.

Introduction Issues 710/18/2015  Coordinate resource usage, the diversity of resource types and the variety of different management modes.  The differing policy requirements of resource owners make difficulties in creating a consistent cross-organized policy domain that delivers the necessary capability to the end user while respecting the policy requirements of the resource owner.  Grid Applications require the concurrently allocation of multiple resources.

Introduction Grid Resource Management Models 810/18/2015  Hierarchical model.  Abstract Owner  Computational Market/Economy Model.

Hierarchical Model 10/18/20159

Hierarchical Model Hierarchical Resource Management  The outcome of Grid Forum’s second meeting.  GRAM, Legion, Ninf.  The major components of this model is divided into passive and active components 10/18/201510

Hierarchical Model Passive components  Resource:  things that can be used for a period of time, and may or may not be renewable  They have owners who may charge others for using resources.  They can be shared or exclusive. 10/18/201511

Hierarchical Model Passive components  Task:  Are consumers of resources  Include both traditional tasks and non-computational tasks (such as file and communication).  Jobs:  Are hierarchical entities: can be composed of sub-jobs or tasks, and sub-jobs maybe themselves contain sub-jobs. The leave of this structure are tasks.  The simplest form of a job is one containing a single task 10/18/201512

Hierarchical Model Passive components  Schedules:  Are mappings of tasks to resource overtime. 10/18/201513

Hierarchical Model Active components  Schedulers:  Compute one or more schedules for input jobs  The unit of scheduling is the job.  The schedulers outside control domains cannot commit resources; these are known as metaschedulers or super schedulers.  Information Services:  Act as databases for describing items of interest to the resource management systems: resource, jobs, schedulers, agents, …  Could be a LDAP, a database,… 10/18/201514

Hierarchical Model Active components  Domain Control Agents (Local resource manager):  Commit resource for use  The set of resources controlled by an agent is a control domain.  Are distinct from Schedulers, but control domains may contain internal Schedulers.  Can provide state information, either through publishing in an Information Service or via direct querying.  The actions outside are called requests. Action inside maybe called command. 10/18/201515

Hierarchical Model Active components  Deployment agents: implement schedules by negotiating with domain control agents to obtains resources and start tasks running.  Users: submit job to the resource management system for execution.  Admission Control Agents: determine whether the system can accommodate additional jobs, and reject or postpone job when the system are saturated. 10/18/201516

Hierarchical Model Active components  Monitors:  Track the progress of jobs.  Obtain job status from the task comprising the job and from the Domain Control Agents where those tasks are running.  Perform outcalls to Job Control Agents and Schedulers to effect remapping, terminating extending lifetime, … 10/18/201517

Hierarchical Model Active components  Job Control Agents:  Are responsible for shepherding a job through the system, and can act both as a proxy for the user and as a persistent point for a job.  Are responsible for coordinating between different components within the resource management system. 10/18/201518

Hierarchical Model Interaction of Components Job Control Agent User A user submit a job to a Job Control Agent. 10/18/201519

Hierarchical Model Interaction of Components Job Control Agent Admission Control Agent User Job Control Agent calls an Admission Agent to examines the resource demands of the job (perhaps consulting with a grid Information System). 10/18/201520

Hierarchical Model Interaction of Components Job Control Agent Admission Control Agent User Scheduler The Admission Agent pass the job to the Scheduler, which performs resource discovery (using the grid Information System). 10/18/201521

Hierarchical Model Interaction of Components Job Control Agent Admission Control Agent User Scheduler Domain Control Agent Scheduler consults with Domain Control Agent to determine the current state and availability of resources. 10/18/201522

Hierarchical Model Interaction of Components Job Control Agent Admission Control Agent User Scheduler Domain Control Agent Deployment The Scheduler computes a set of mappings and passes these mappings to a Deployment Agent. 10/18/201523

Hierarchical Model Interaction of Components Job Control Agent Admission Control Agent User Scheduler Domain Control Agent Deployment The Deployment Agent negotiates with the Domain Control Agents for the resource indicated in schedules and obtains resource reservations for the resources 10/18/201524

Hierarchical Model Interaction of Components Job Control Agent Admission Control Agent User Scheduler Domain Control Agent Deployment These reservations are passed to Job Control Agent. At the proper time, the Job Control Agent works with different Deployment Agent, and the Deployment Agent works with the appropriate Domain Control Agents to start the tasks running. 10/18/201525

Hierarchical Model Interaction of Components Job Control Agent Admission Control Agent User Scheduler Domain Control Agent Deployment Monitor A Monitor tracks progress of the job and may later decide to reschedule if performance is lower than expected. 10/18/201526

Abstract Owner Model 10/18/201527

Abstract Owner Model  Introduction  General Structure of AO  Grid Resources  Negotiating with an AO  Job Shops 10/18/201528

Abstract Owner Model Abstract Owner  When one makes a long distance phone call, who "owns" the resource being used? 10/18/201529

Abstract Owner Model Abstract Owner  Each grid resource be represented by one or more “abstract owners” (AO)  For complex resources, an AO will certainly be a broker for the actual owners or other brokers  Pre-existing contract or protocol to acquire resource 10/18/201530

Abstract Owner Model General Structure of AO - Resource  Resource: any combination of hardware and software  A processor  Another Grid system  Resource will be regarded as object  Methods: initiating and controlling tasks  Attributes: customize the resource 10/18/201531

Abstract Owner Model General Structure of AO - Task  Task: problem/part of problem that need resource to do.  A message to be sent  A process to be executed  A complete multi-process program 10/18/201532

Abstract Owner Model General Structure of AO–External view  Order Window: negotiate to acquire resource  how soon the resource may become available?  how much it might cost?  Pickup Window: deliver resource to client 10/18/ Abstract Owner Order WindowPickup Window External View of AO model

Abstract Owner Model General Structure of AO – External View 10/18/ Client send request and negotiate to AO successful negotiation? Client place order to AO through Order Window AO return resource to Client through Pickup Window Client wait for resource? CANCEL N Y Y N

Abstract Owner Model General Structure of AO – Internal View  InternalStructure  An owner of a physical resource 10/18/ Resource Manager Order WindowPickup Window AO is Resource Owner Physical Resource

Abstract Owner Model General Structure of AO – Internal View  Internal Structure  Higher level broker 10/18/ AO is Broker Manager Order WindowPickup Window Sales Rep. Delivery Rep. AO3 AO2 AO1

Abstract Owner Model Grid Resources  Three type of resources:  Instruments Logically exists at some location. Creates, consumes, or transforms data or information.  Channels Exists to facilitate the explicit transfer of data or information between two or more instruments Connects to an Instrument through a Port  Complexes A collection of (connected) Channel and Instrument resources. 10/18/201537

Abstract Owner Model Grid Resources  Instruments are divided into subclasses  Compute Instrument: a processor or set of processors along with associated memory, temp files, software, etc.  Archival Instrument: persistent storage of information.  Personal Instrument: interface directly to a human being, ranging from a simple terminal to a more complex 10/18/201538

Abstract Owner Model Negotiating with a AO  Client need to negotiate with AO before using resource  Negotiating process:  Client sends negotiation  AO replied with appropriate resource information  Client accept and place order, or reject the negotiation. 10/18/201539

Abstract Owner Model Negotiating with AO  Sample object attributes:  Client first effectively creates a "sample" resource object of the appropriate structure  Assigns each attribute either (1) a constant value, (2) a "don’t care" value, (3) a variable name  Some variables can be “negotiable”  Variable constraint list: constraints for variables 10/18/201540

Abstract Owner Model Negotiating with AO  Negotiation Style:  Immediate: AO is to schedule the resource immediately  Pending: AO is to return a specified number of sets of proposed values for the negotiation variables  Confirmation: AO is to finish scheduling based on an earlier-returned set of negotiation variable values  Cancel: AO is to cancel earlier Pending negotiation  Pickup Approach: Protocol to be used at Pickup Window 10/18/201541

Abstract Owner Model Negotiating with AO  Authorization: a capability or key which allows the AO to determine the authority of the client to access resources  Bid: maximum price that the client is willing to pay for the  Negotiation ID: like cookie 10/18/201542

Abstract Owner Model Job Shop  Instead of only deal with Resource acquiring, “Job Shop” handle job scheduling.  Job Shop  Estimator  Executor 10/18/ Job Shop (Estimator & Executor) Job Shop (Estimator & Executor) JobResult Job Scheduling AO for Grid Resource Des. Resource

Abstract Owner Model Job Shop – The Estimator  Deals with the customer, acting as a liaison with the grid AO (through its order window) if necessary to help determine how soon the job might be done and how much it might cost.  Requests the resources needed from the grid AO (through the order window).  Records (in the Work To Do List) what needs to be done when the resources are ready. 10/18/201544

Abstract Owner Model Job shop – The Executor  Takes ready resources from the AO delivery window  Dequeues the associated work (tasks) from the Work To Do List  Builds any necessary environment for those tasks based on the resource  Initiates the tasks on the resource and collects the answer.  Frees the resource upon completion, if necessary.  Notifies and returns the answers to the user 10/18/201545

Abstract Owner Model Job Shop 10/18/ Estimator Job Shop Executor Work to do list

Abstract Owner Model AO Summary  AO model is still a proposal  There are many remaining gaps in the model  Even a potentially useful and well-defined (successfully prototyped) AO protocol will not be viable unless it can coexist with other contemporary approaches. 10/18/201547

Economy/Market Model 10/18/201548

Economy/Market Model Overview  Grid environment  Resources are geography distributed and belong to different organization  Has different resource management, policies, charges for different users  Need a market-driven resource management system to satisfy both resource owner and user User : minimize expenses Resource owner: maximize return-on-investment 10/18/201549

Economy/Market Model Overview  Necessitating a grid resource management system:  Provides appropriate tools and services to allow both resource users and owners to express their requirements. Users: specify their “QoS requirements” such as minimize the computational cost and yet meet the deadline for the results. Resource owners: specify their charges, that can vary from time to time and users to users—and terms of use.  Likely that more and more systems are going to emerge based on this concept. 10/18/201550

Economy/Market Model Overview  Economy/Market model for grid resource management  Captures the essentials of both Hierarchical model Abstract Owner model  Many of the similar ones appears to be futuristic  But points out the need for economy in computation implicitly. 10/18/201551

Economy/Market Model Overview  The key components of economy-driven resource management system:  User Applications (sequential, parametric, parallel, or collaborative applications)  Grid Resource Broker (a.k.a., Super/Global/Meta Scheduler)  Grid Middleware  Domain Resource Manager (Local Scheduler or Queuing system) 10/18/201552

Economy/Market Model Overview  A possible architectures for grid resource management based on computational market model 10/18/201553

Economy/Market Model Grid Resource Broker (GRB)  Acts as a mediator between the user and grid resources using middleware services.  Is responsible for  Resource discovery.  Resource selection.  Binding of software (application).  Data, and hardware resources,.  Initiating computations.  Adapting to the changes in grid resources.  Presents the grid to the user as a single, unified resource. 10/18/201554

Economy/Market Model Grid Resource Broker (GRB)  Includes  Job Control Agent (JCA): is a persistent central component responsible for shepherding a job through the system. takes care of schedule generation, the actual creation of jobs, maintenance of job status, interacting with clients/users, schedule advisor, and dispatcher.  Schedule Advisor (Scheduler): to select those resources that meet user requirements such as meet the deadline and minimize the cost of computation; is responsible for resource discovery (using grid explorer), resource selection, and job assignment (schedule generation). Its key function is while assigning jobs to resources. 10/18/201555

Economy/Market Model Grid Resource Broker (GRB)  Includes (cont.)  Grid Explorer: is responsible for resource discovery by interacting with grid-information server identifying the list of authorized machines keeping track of resource status information.  Trade Manager (TM): works under the direction of resource selection algorithm (schedule advisor) to identify resource access costs. interacts with trade servers and negotiates for access to resources at low cost. It can find out access cost through grid information server if owners post it.  Deployment Agent: is responsible for activating task execution on the selected resource as per the scheduler’s instruction. periodically updates the status of task execution to JCA. 10/18/201556

Economy/Market Model Grid Middleware  Offers services that help in coupling a grid user and (remote) resources through a resource broker or grid enabled application. It offers core services  Remote process management  Co-allocation of resources  Storage access  Information (directory)  Security, authentication  Quality of Service (QoS), such as resource reservation for guaranteed availability and trading for minimizing computational cost.  Some of these services have already been discussed in the hierarchical model. 10/18/201557

Economy/Market Model Grid Middleware  Main components  Trade Server (TS): It is a resource owner agent Negotiates with resource users and sells access to resources. Aims to maximize the resource utility and profit for its owner Consults pricing algorithms/models defined by the users during negotiation Directs the accounting system to record resource usage.  Pricing Algorithms/Methods: Define the prices that resource owners would like to charge users. The resource owners may follow various policies to maximize profit and resource utilization and the price they charge may vary from time to time and one user to another user and may also be driven by demand and supply like in the real market environment.  Accounting System: Responsible for recording resource usage Bills the user as per the usage agreement between resource broker and trade server (resource owner agent) 10/18/201558

Economy/Market Model Domain Resource Manager  Local resource managers  Responsible for managing and scheduling computations across local resources such as workstations and clusters.  Even responsible for offering access to storage devices, databases, and special scientific instruments such as a radio telescope. 10/18/201559

Economy/Market Model Comments  There’re numerous methods for determining/knowing access cost.  The services offered by trade server could also be accessed from or offered by grid information  A trade manager or broker can directly access information services to identify resource access cost and then contact resource agents for confirmation of access.  The trade manager can use these advertised/posted prices or ask/invite for competitive quotes or bids and choose resources that meet user requirements.  The proposed figure is one of the possible alternatives for computational market model and it can vary depending on, particularly, trading protocols like in real world economy.  The different grid systems may follow different and it will be beneficial if they are all interoperable.  The interoperability standards can be evolved through grid user/developer community forums or standardization organizations 10/18/201560

Economy/Market Model Q & A 10/18/201561

Economy/Market Model References 10/18/  Rajkumar Buyya†, Steve Chapin*, and David DiNucci, Architectural Models for Resource Management in the Grid, USA, 2000  Ian Foster and Carl Kesselman (editors), The Grid: Blueprint for a Future Computing Infrastructure, Morgan Kaufmann Publishers, USA, 1999