WO2018141435A1

WO2018141435A1 - Method and apparatus for allocating device resources

Info

Publication number: WO2018141435A1
Application number: PCT/EP2017/080724
Authority: WO
Inventors: Josep Soler Garrido
Original assignee: Siemens Aktiengesellschaft
Priority date: 2017-01-31
Filing date: 2017-11-28
Publication date: 2018-08-09

Abstract

The invention relates to a method for allocating resources of devices (2) of a distributed system (1), wherein each device (2) of the distributed system (1) allocates the locally available resources (5) of said device to tasks in accordance with external execution quality metrics, which the device (2) receives from other devices (2) of the distributed system (1).

Description

Method and device for assigning device resources

The invention relates to a method and apparatus for allocating resources from distributed system devices to tasks executed by distributed system devices.

Distributed systems can include a variety of devices that have different resources and communicate with one another. The resources of the distributed system include, for example, processor resources, memory resources or data transfer resources. The various devices of the distributed system are thus limited in their local resources and are exposed to a dynamic workload. The various devices or nodes of the system can locally execute tasks of a work process or workflow to use their local resources. In this case, for example, local sensor data can be processed which receive the various devices of the distributed system from associated data sources or sensors from local data streams. Using the received sensor data, the distributed system devices can perform various tasks, such as diagnostics, maintenance, monitoring, or other data processing tasks. Workflows or working processes of the distributed Sys tems ^¬ comprise a plurality of tasks that are performed on different devices using multiple data streams. The results of the various individual tasks can be combined or chained to extract results to identify different patterns or certain situations within the distributed system know ^¬ to it. Dependent on this, for example, a control unit of the distributed system can initiate or carry out measures. When executing a work process or workflow advertising the assigned various tasks of the workflow data processing ^¬ resources of the distributed system. The allocation of resources takes place dynamically in many applications. It is therefore an object of the present invention to provide a method and apparatus for allocating resources of wireless devices within a distributed system to tasks wel ^¬ che of devices of the distributed system are performed to establish, in which the allocation of resources is optimized.

This object is fulfilled according to the invention by a method for allocating resources of devices, which has the features of patent claim 1.

The inventive method allows the allocation of the system resources of the distributed system in a robust and simple manner.

Possible embodiments of the method according to the invention are specified in the subclaims.

In one possible embodiment of the method according to the invention, the other tasks that are related to the tasks that can be executed by the respective device are contained in at least one common workflow, which comprises a set of z associated tasks.

In one possible embodiment of the method according to the invention each unit of the distributed system performs the Loka ^¬ le optimization, when a resource allocation to each device is changed, or change the available locally on the device resources, or when the device an update at least one embodiment quality metric from a receives another device of the distributed system.

In a further possible embodiment of the method according to the invention, the predetermined optimization function has a global function of the entire distributed system. In a further possible embodiment of the method according to the invention, the resources of a distributed system device include processor resources, memory resources and / or data transmission resources of the device.

In another possible embodiment of the method according to the invention, the execution quality metrics indicate a quality with which a task is executed by the device.

In another possible embodiment of the method according to the invention, the execution quality metrics have a sensitivity, an accuracy and / or an error rate.

In another possible embodiment of the method according to the invention, a task that can be executed by a device can be executed in different operating modes of the device, which each claim a different set of resources of the device.

In a further possible embodiment of the method according to the invention, a set of permissible operating modes of a device is provided for each task.

In a further possible embodiment of the method according to the invention, the selection of resources of a device for executing a task by switching to another permissible operating mode of the respective device takes place.

The invention further provides, according to another aspect, a resource allocation device having the features specified in claim 13. Possible embodiments of the resource allocation according to the invention are specified in the subclaims. In one possible embodiment of the resource allocation device according to the invention, the resource allocation device has a storage unit for storing the local execution quality metrics and the external resource allocation

Execution quality metrics.

In a further possible embodiment of the resource allocation apparatus according to the invention, the resource assigning apparatus further includes a storage unit for storage of all of brass by the device within a workflow from ^¬ executable tasks, ie which set of tasks T can perform the device locally.

Further, the device stores which sets of tasks within the system are related to any task of the set of tasks and which other devices are capable of executing them.

The invention further provides, according to a further aspect, a distributed system having the features specified in claim 16.

Accordingly, the invention provides a distributed system with a plurality of devices, each of which has resources for processing tasks and each containing a Ressourcenzuwei ^¬ sungsvorrichtung according to one aspect of the invention.

In one possible embodiment of the distributed system according to the invention, the resource allocation devices of the distributed system devices exchange

Execution quality metrics to increase the efficiency of resource utilization within the distributed system.

In a possible embodiment of the distributed system according to the invention, the devices of the distributed system field devices have each including a processor as ^¬ calculation unit, the data derived in particular from sensors, sensor data, in a workflow, the one or more Tasks includes, processes and transmits the processed data to a control unit of the distributed system.

In the following, possible embodiments of the inventive method and the device according to the invention for assigning device resources of the distributed system to tasks will be explained in more detail with reference to the attached figures.

Show it:

1 is a block diagram illustrating a possible exemplary embodiment of a device having a resource allocation device according to the invention;

FIG. 2 is a flow chart illustrating an embodiment of a method of allocating device resources of a distributed system according to the invention; FIG.

Fig. 3 is another block diagram illustrating a

Embodiment of a distributed system with devices whose resources are assigned by means of a device according to the invention;

4 shows an example of a work process or

Workflow, which contains several tasks, which Ge ^¬ räteressourcen can be assigned using the inventive method and the apparatus according to the invention;

5 is a table of an exemplary list of event detection task operation modes that may be performed by a device;

6 is a graph for logically illustrating an exemplary implementation of various workflows on different distributed system devices; a functional diagram for representing an average workflow overhead as a function of a resource variance according to a simulation; a further functional diagram for representing an average workflow overhead as a function of a resource variance according to another simulation; a distribution of values of an expense or

Cost function for different methods used in a simulation. As can be seen from Fig. 1, a distributed system 1 according to one aspect of the present invention comprises a plurality of devices 2-0, 2-1, 2-2 capable of communicating with each other. For example, the various devices 2-i may communicate with each other via a wireless or wired interface. For example, the devices 2-i are connected to a common bus of the system 1 in order to be able to exchange data or messages with one another. The bus may be a serial or parallel data bus. For example, the devices are 2-i field devices that have data via a field device bus, in particular

Execution quality metrics, can exchange with each other. In the embodiment of the distributed system 1 shown in FIG. 1, the various devices 2-i can also transmit processed data to an optionally existing common unit 3 of the distributed system 1. The unit 3 may include one or more workflows and working process abar w ^¬ BEITEN. In a preferred embodiment of the distributed system no unit 3 is present and the system consists of devices 2, in particular of field devices. The number of devices 2 and the number of layers may vary. A first layer of devices 2 may be connected to sensors. For example, a second layer of devices 2 may serve as a data concentrator and unprocessed data from field devices. Rates 2 received. Further units, for example in a cloud, can collect and process data in a third layer. Each of these work processes or workflows w comprises one or more tasks t. In the exemplary embodiment illustrated in FIG. 1, each device 2-i can be connected to one or more data sources 4-i which supply data for data processing. For example, the device 2-0 of the distributed system 1 is connected to N different data sources 4-1, 4-2 ... 4-m in the illustrated example. The data sources 4-i may include sensors that provide sensor data in different data streams DS to the device 2-0. Other data sources, such as memory units or the like, are also possible. The data sources can also be formed by other devices 2. For example, a device 2 may send raw data or processed data from its sensors to other devices 2 for further data processing. These data streams can be transmitted via a network interface. Each device 2-i has a ^¬ be restricted amount of locally available resources for data processing. In the example shown in FIG. 1, the device 2-0 has N different device resources 5-i. This device resources are 5-i, for example, processor resources, memory resources and data transmission resources of the equipment 2. The various devices 2-i of the, allocated system 1 each have a resource allocation ^¬ device 6 for allocating resources 5-i of the device to task t on, which are executed by devices of the distributed system 1 ^¬ . For each unit 2-i of the distributed system 1 is a predetermined optimization function is provided, the external local execution quality metrics q _d, the device of the self-executable tasks 2- i t, and

Execution quality metrics q _d 'depend on other tasks t', which may be related to the tasks t executable by the respective device 2-i and executable by other devices of the distributed system 1. The other tasks, which together ^¬ menhängen with the locally executable by the respective unit 2-i tasks are preferably contained in a common workflow w, comprising a set of associated tasks. each Distributed system device 1 also has relation data for each task executable by the device, which is a relationship between a local execution quality metric q _d with which a task can be executed by the respective device and resources 5-i of the device. which are assigned to the respective task. The

Execution quality metrics may depend on the resources of the device itself and, where appropriate, resources of the sensors or other device extensions that the device controls.

As schematically illustrated in FIG. 1, each resource allocating device 6 of a device 2-i within the distributed system 1 preferably comprises a receiving unit 6A, a calculating unit 6B and a transmitting unit 6C. The Emp ^¬ capturing unit 6A and 6C, the transmitting unit can be integrated in a transmitter 6 de-receiving unit of the resource allocation device. The receiving unit 6A of the resource allocating device 6 is adapted to receive execution quality metrics q _d of tasks t from other devices 2 of the distributed system 1 executed by the other devices 2. In the embodiment illustrated in FIG. 1, the device 2-0 communicates with two further devices 2-1, 2-2 of the distributed system 1 for exchanging execution quality metrics. In the illustrated example, the receiving unit 6A of the device 2-0 receives a first execution quality metric q _d i 'from the device 2-1 with respect to tasks executed by the device 2-1. Furthermore, the receiving unit 6A of the device 2-0 receives execution quality metrics q _d 2 'from tasks of the further device 2-2 of the distributed system 1, which are executed by this other device 2-2. The calculation unit 6B of the resource allocation device 6 is adapted to a local function from an upstream LF give ^¬ nen to be optimized optimization function (F) derive. The local function LF preferably depends on the local Execution quality metrics q _d, 2-i executable tasks of the equipment used. The derivation of the local function LF, which depends on the local execution quality metrics q _d , from the given optimization function F is performed on the basis of the external execution quality metrics q _d ', which the receiving unit 6A of the respective device 2-i has received from the other devices of the system 1 , In one possible embodiment of the distributed system 1 according to the invention, the optimization function F is a global function of the entire distributed system 1. The optimization function F can have a cost function to be minimized.

The computing unit 6B of the resource allocation device 6 is further adapted to perform a local optimization based on the derived local function LF and the relationship data present in the respective device 2-i. The local optimization is carried out on the basis of abgeleite ^¬ th local function LF and stored relationship data of the respective device 2-i in consideration of the available resources 5 of the respective unit 2-i for selection of resources of the particular device which the affected task to the execution be assigned to.

The resource allocating device 6 further includes the sending unit 6C that is suitable for updating

Execution quality metrics q _d each task performed by the respective device 2-i, which is performed by the device 2-i using the resources of the device 2-i assigned to this task, to any other device of the distributed system 1, the is provided for performing another task, for example, related to one of the executed by the respective device 2-i task. In the embodiment shown in Fig. 1, the transmitting unit 6C of the device 2-0 transmits updated execution quality metrics q _d o of each task t performed by the device 2- 0 by the device 2-0 using the resources 5 assigned to this task t 2-0, to the other two devices 2-1, 2-2 of the distributed system 1, wherein each of the two remaining devices 2-1, 2-2 are provided for the execution of another task, which may be related to one of the tasks executed by the device 2-0 or may belong to the same workflow w.

The resource allocating device 6 preferably has a storage unit for storing the local

Execution quality metrics q _d and the received or

received external execution quality metrics q _d '. The

Resource allocation device 6 preferably also has a further storage unit for storing all tasks t that can be executed by device 2-i within a workflow or workflow w. Further, the apparatus 2 stores are provided which sets or groups of tasks within the system with ir ^¬ vicinity of a task of the set of tasks in relationship, and what other devices of the system 1 are in a position they exit out the various resource allocation devices 6 of the various devices 2- i of the distributed system 1 exchange execution quality metrics among each other to increase the efficiency of resource usage within the distributed system 1. In one possible embodiment, the various devices 2-i of the distributed system 1 have field devices, each of which contains a processor as the calculation unit 6B, which processes data in a workflow w, which includes one or more tasks t, and then processes the processed data to the control unit 3 of the distributed system 1 transmits. The processor and the calculation unit 6B of the unit 2-i, for example, processed Sensorda ^¬ th derived from sensors 4-i, which are connected to the device in question 2-i, or derived from other system devices. 1

2 shows a simple flow chart for illustrating an embodiment of the method according to the invention for allocating resources from devices 2-i of a distributed system. tems 1 to tasks t, which are executed by devices 2 of the distributed system 1, according to another aspect of the invention. Each device 2-i of the distributed system 1 can execute the main steps shown schematically in FIG.

In a first step Sl, the relevant device 2-i receives from other devices of the system 1

Execution quality metrics q _d 'of tasks executed by the other devices 2 of the system 1. It can be sufficient if a single execution quality metric of another device has changed. The

Execution Quality Metrics with Other Devices of System 1 Performing tasks generally are not received by device 2-i at the same time.

They can be received asynchronously, i. Each time an individual device alters its allocation of resources, it tells the other devices.

In a further step S2, a local function, which depends on the local execution quality metrics q _d of the tasks t _{that can be} executed by the respective device 2-i, is first determined by a predetermined optimization function F to be optimized on the basis of the external execution quality metrics q _d ' derived the respective device 2-i from the other devices of the system 1 derived. The to be optimized

Function F is in a possible embodiment a global function of the entire distributed system 1, in particular a global cost function of the distributed system 1. In a further step S3, it is a local optimization ^¬ step on the basis of the derived at step S2 local function and on the basis of relationship data of the respective device 2-i, taking into account the available resources of the respective device 2-i for the selection of resources of the respective device 2-i, which are assigned to the relevant task for their execution. The device has two for each of the unit 2 executable task from relationship ^¬ data. The stored relationship data gives a relationship between a local execution quality metric q _d , with which a task can be executed by the relevant device 2, and the resources 5 of the device 2, which are assigned to the respective task for this purpose. The relationship data can be stored in a memory unit of the device 2.

In a further step S4, the updated

Execution quality metrics q _d each task performed by the respective device 2-i executed by the respective device 2-i using the resources 5 of the device 2 assigned to this task to all other devices 2 of the system 1 which are used to execute a other task are provided. The other task may be related to one of the tasks executed by the respective device 2-i or belong to a same workflow w. Each device 2 of the distributed system 1 may, in one possible embodiment, perform the local optimization in step S3 as soon as a resource allocation at the respective device 2-i is changed and / or a resource availability in the device 2-i changes. The exchange of execution quality metrics may be continuous in one possible embodiment. Alternatively, the replacement of the

Execution quality metrics in steps S1, S4 are triggered when a particular event occurs. The un ^¬ ter the devices 2-i exchanged

Execution quality metrics indicate a quality with which a task by the particular device 2-i of the system is guided out ^¬. 1 In one possible embodiment, the exchanged execution quality metrics have a sensitivity, an accuracy and / or an error rate with which a task is executed by the device 2.

An executable by a device 2 of the system 1 task, which is executable in different operating modes m of the device 2, each claiming a different set of resources 5 of the device 2 and / or resources of device extensions, in particular of sensors or data sources. In a possible In this embodiment, a set of permissible operating modes of the relevant device 2-i is provided for each task. The selection of resources 5 of a device 2-i to perform a task in one possible embodiment by switching to another permissible operating mode of the device in question 2. In the event that the local resources of a device 2 are changed, the

Steps Sl, S2 of the method shown in FIG. 2 ent ^¬ fall.

Fig. 3 shows another schematic diagram for Erläute ^¬ tion of the operation of a distributed system 1, which comprises multiple devices 2-i, the tasks of a workflow w can perform ^¬.

In the method according to the invention, an iterative distributed allocation of resources takes place by means of a distributed resource management. The individual distributed resources ^¬ management by the different devices 2 of the distributed system 1 has the advantage that a local optimization can be done by the different devices 2 of the distributed system. 1 This local optimization based on a derived local function LF can be performed particularly quickly as soon as the resource allocation situation within the distributed system 1 changes. Each device 2-i of the distributed system 1 has information indicating the topology of the workflows or workflows w. Depending ^¬ the unit 2-i has information or data which is ^¬ ben which tasks are executable by the device concerned. 2 The device 2 has for each workflow w on Informa ^¬ tion data which group or set of devices which perform tasks 2, which belongs to the relevant workflow w. In addition, each device 2 has information data of a function F associated with these workflows w. The function F is a function of the local

Execution quality metrics and the external

Execution quality metrics derived from tasks that run on or run through other devices 2 of the distributed system 1.

Each node or device 2 of the distributed system 1 stores data of execution quality metrics which the device 2 has received from other peer devices 2 of the system 1.

Once a resource allocation and resource allocation is changed at a certain device 2 or the resource availability changes locally, the unit 2 in question performs a local optimization on the basis of the derived Loka ^¬ len function by LF, wherein values of the

Execution quality metrics are fixed by the other devices 2 of the distributed system 1 or kept constant.

In this way, an operating mode m can be selected for each task of the device 2. Further, by using the vorlie ^¬ constricting mapping a set or group of data source can configurations of data sources 4-i and values of

Execution quality metrics are read from a lookup table. The output of the local optimization step is the optimal resource assigned to each task that can be executed locally as well as related qualities or quality metrics. The relationship or mapping between the two is known and used by the device. However, the relationship can be expressed in different ways. For example, a list of operating modes (quality resources) may be provided. Alternatively, the quality of a task or its execution quality metric may be defined as a function of the resources used for it. The device 2-i can configure associated data sources 4-i (dashed line in FIG. 1) to obtain data corresponding to the individual tasks of the operating mode. The device 2 transmits its updated

Execution quality metrics to the other peer devices 2, so that they can also perform a local optimization. The resource allocation method according to the invention provides a faster iterative update of data source ^{configurations} in response to changes in device resources, enabling or disabling workflows, changing workflow priorities, device errors, or data source failures.

To achieve convergence that will be taken in the event that different nodes or devices perform two updates to sel ^¬ ben time in a sequential manner ensures. For example, if two different devices 2A and 2B perform an update, it is ensured that one of the two devices first performs the update and then transmits the determined execution quality metrics to the other device that updated the received one

Execution quality metrics used for its own local optimization. This may for example be effected by by adding time stamps or sequence numbers to the updated execution quality metrics, which are exchanged between different nodes or devices 2 of the distributed system 1 and preferably additionally defines a predefi ^¬ ned priority between the various devices 2 which all devices 2 of the distributed system 1 is known. For example, if a first node 2A broadcasts an execution quality metric update with the N + 1 sequence, but receives another update from another higher priority peer device 2B that specifies the sequence metadata, that the quality metric sequence N of device 2A has been used, then the node or device 2A can, based on its known priorities, immediately broadcast a new update N + 2 based on the last execution quality metrics of the device 2B. The node or device 2B of the distributed system 1 ignores the update N + 1 of device 2A because it is not based on the last execution quality metrics of the device 2B.

In one possible embodiment, the resource allocating device 6 of the device 2 may perform various optimization procedures. For example, the device 2 leads to Timing by the installation of the system 1 or dauerhaf ^¬ th changes an initial or initial optimization. In another situation, for example in the case of very short-term resource changes, in one possible embodiment, an individual resource management or an individual resource allocation can take place without feedback from the other devices 2 of the distributed system 1. For a longer change in the resource change, the iterative distributed resource allocation or resource allocation can then be carried out with the aid of the method according to the invention.

Fig. 4 shows a schematic representation for a game stick at ^¬ workflow w, which can be carried out by the inventive distributed system 1. In the illustrated example of FIG. 4, the workflow w is used to detect or detect a system event S based on the combination of six independent events or events S 1 to S 6 via logic operations. Each event can be signaled by an associated task t1 to t6, which are executed on a device 2. In the 4 Darge ^¬ presented in Fig., A single Ausführungsqualitätsmet- rik is selected for the particular workflow W or selected. The execution quality metric in the illustrated example is the sensitivity s _c for detecting a complex event c. This embodiment quality metric s _c (up S6 Sl) for the various individual events by the individual sensitivities (El to E6) expressed ^¬ the. In the illustrated example, the combined event sensitivity s _c can be expressed as follows: s _c = (1 - (1- _Sl ) (1-s ₂ )) x (1 - (1-s ₃ ) (1-s ₄ )) x (1 - (1 - s ₅ ) (l - s ₆ ))

For example, a cost function for the workflow w can be defined as a false negative rate FNR: C = 1 -s _c . Devices 2, in particular field devices, are limited in terms of their CPU resources. The sensitivity with which event detection can be achieved depends on the analytical model and sensor configuration. Each configuration requires a corresponding set of device ^¬ resources of the unit 2. The table shown in FIG. 5 shows an example of how different five operating modes m for each task defined who can ^¬. The table belongs to a task t that runs on a particular device 2. The same task t, if executable on different devices of system 1, has a different table for each of these devices. For example ^¬ the task requires more resources of the device in question to achieve the given quality. Each operation mode ^¬ m may for example be defined for executing the task by a differing ^¬ Chen algorithm.

FIG. 6 shows a deployment in which three workflows w, as shown in FIG. 4, can be deployed or deployed. In the example shown, the

Workflows w deployed to nine different field devices 2, each performing two tasks t. In this case ^¬ play case, it is the task of the Ressourcenallokators or the resource assigning apparatus 6 to allocate the limited CPU resources of each device 2 about all tasks in such a way that a cost function for each workflow w is minimized. This is preferably performed dynamically, because the availa- ^¬ cash device resources of the devices 2 change or because new additional tasks can be implemented t. Next ^¬ out the resources of the devices 2 to be dynamically allocated, if new data models are available or change the workflow priorities. Each workflow w includes several tasks t that belong to the workflow w. Complex workflows w can be further grouped into even more complex workflows w. Fig. 6 is a logical diagram showing no physical objects.

A resource allocation mechanism serves to allocate resources 5 to tasks on each device 2. In the distributed system 1, a set or a group of different workflows w may be provided, which each comprise a set or a group of associated tasks, as well as a topology which indicates how the tasks are linked to one another. For each task t a set of operating modes zuläs ^¬ sigen m can be provided which contain the necessary configuration for the data source 4 and the associated necessary device resources. ^r d, j, m ⁼ { ^r d, j, l> ^r d, j, 2 - ^r d, j, R} m> ^me {1- d, j]

An operating mode m represents a particular way in which a task t is to be solved, with an associated set of resources and qualities achieved. An operating mode m may be defined by a particular configuration at data sources 4 or by other parameters, such as a particular algorithm for solving the task.

For each task t, an optionally multidimensional execution quality metric may be indicated as a function of the resources used.

For each workflow w, a cost function may be specified as a function of the execution quality metrics. For each device 2 of the system 1 comprises a set of verfüg ^¬ cash resources for executing the task t. This set of available resources may change over time, depending on the current workload of the device 2. The workload of the device 2 is measurable in one possible embodiment.

Use of the method according to the invention it is possible to those operation mode m for each individual task that is running on the respective device 2 t, in such a way to determine men that the sum of the effort, which relate to the actual ^¬ workflow w connected minimized, becomes. An example of an optimization function F is:

without exceeding the available resources R:

The method according to the invention can be based on a distributed optimization algorithm. Each device 2 of the distributed system 1 can individually decide which proportion or percentage of its available resources 5 it assigns to the various individual tasks. Each device 2 is capable of exchanging information with other devices of the system 1 to iteratively achieve a global optimum for the entire distributed system 1. The devices 2 inform their peer devices 2 of the quality with which they perform their own tasks t. This information is used by each device 2 in the next local optimization step to allocate resources to locally executed tasks. In this way a high probability is achieved friendliness to achieve a global optimum, while at the same time keeps the complexity and quantity of the information exchanged ^¬ or data minimal.

Each device 2, for example each field device, lists a set of tasks based on its locally available sensor data and other required data based on its limited resources, for example, the device 2 through a network of other distributed system devices 1 receives. Each device 2-i of the system 1 can, for example, execute T _d tasks. Each unit 2-i may play control a set of S _d local data sources or sensors at ^¬, to generate data for data processing. A set of R device resource classes (for example, CPU, memory, or network bandwidth ...) is the second device of all devices distributed system 1, which perform tasks t, together. The amount of each of these different resources for the data processing tasks may vary for each device 2. The amount of the resource type r available on a device d may be indicated as R _d , _r . This can be distributed to the various data processing tasks. Thus, r _{di ir is} the amount of resource r assigned to a task j on a device d. Thereof must meet the following condition for each type of resource r follows that each device of the system 1 d at each time point: r _d, j, _r <Rd.r Vd, r: d E {1 ... D], E {r 1 ... R}

The results of all tasks that are performed on all devices 2 of the Sys tems ^¬ 1 can be combined in w workflows that are able to extract results in terms of the distributed system. 1 Different tasks t can be combined to a workflow w on a higher layer or layer. This combination does not necessarily have to be done at the device level. In one possible embodiment, it is also possible, for example, to communicate between devices 2 for the organization of the workflows w. The data sources or sensors 4 can support different configurations with different attributes, for example with a certain sampling rate or with a certain accuracy or precision. The various devices 2-i of the system 1 are capable of configuring the data sources 4 and controlling their attributes. Typically, an execution quality metric is assigned to an analytic task. The execution quality metric includes, for example, sensitivity, specificity, or accuracy. The quality of execution of a task t depends on the data model or algorithm used and on the configuration of the data sources 4, for example on a sampling rate of the data sources 4. Tasks that process data with a higher level of execution ^¬ quality, generally require more device resources 5. Therefore, it is possible to define the quality of a task or the execution quality as a function of applied comparable to them resources r. For each task j, which is executed on a device, i.e., the quality q may be defined as follows: dj ⁼ <Ιά,] (Τά,], 1> ^{γ ά,]} 2> - ^r d, j, R) - The quality q as a function of resources r can of course be continuous. In most cases the quality is discretized. For a given task j, which runs on a ^device d (t _d ), it is possible to define a finite set of operating modes m _di as follows: ^m d, j ^e i ^m d, j, l- ^m d , j, Mdj} ·

Each operating mode n _dm has an associated set of required resources r. ^r d, j, m ⁼ { ^r d, j, l> ^r d, j, 2- ^r d, j, R m and therefore the associated quality qd, j, m ^~ di ^' d .m) - An execution quality metric for a workflow w as a function of the qualities of the associated associated tasks. This can be formulated as a cost function, for example, with higher values indicating a ge ^¬ ringere quality. A workflow Wi with a set of related tasks ti the effort and depend on the Kos ^¬ ten belonging to the workflow w, from the

Execution quality metrics of the associated tasks C _L C ^), x £ T _L and thus of the operating mode m of each individual task t, which in turn are a function of the device resources r used by the device 2. FIGS. 7, 8 show simulation results in a functional diagram. In this case, a simulation or an Implemen ^¬ tion is shown, as shown in FIG. 6. There, three different workflows or workflows wi, W2, W3 are shown, each workflow w based on the logical combination of six tasks t, as shown in Fig. 4. These tasks are distributed over nine different devices 2, each performing two tasks t. The various tasks have five different operating modes

Execution quality metrics and resource usage as shown in the table of FIG. 5.

The total amount of CPU resources of each device 2 can be modeled as an independent random variable with a Gaussian distribution Ν (μ, σ ² ).

Both simulations lead optimization using different methods through which a random value for the supply to Verfü- existing resources will be used and a random ^¬ set of connections between the tasks and workflows.

In the first simulation shown in FIG. 7, the average value μ of available resources in the devices is 2 μ = 2.2. This means that, on average, the devices 2 have resources running both analytical tasks t in the higher operating mode plus a 10% margin. The variance σ ² ranges from [0 to 2] depending on the respective simulation point.

Fig. 7 shows the average cost (average

False negative rate FNR) for different methods. In the case of a low resource variance, the false negative rate FNR approaches zero, as expected, because the devices 2 have more device resources than are required to perform all tasks in the higher mode of operation. For a larger resource variance, it can be observed that the performance or performance of the individual Resource management (curve I) decreases faster because there kei ^¬ nen way or no way to adjust the allocation of resources and resource allocation to the global tasks to the lack of resources in some devices with 2 resources to other equipment 2 of the system 1 are available to compensate. The distributed algorithm which implements the invention it ^¬ proper resource allocation method, however, approaches according to curve II after two iterations ^¬ exceeded the performance of an optimal conven- tional search algorithm (curve III) sharply, with a much lower complexity level. The curve I shows the Ver ^¬ equal to the FNR for an individual resource management.

8 shows a diagram for illustrating a further simulation. FIG. 8 shows an average workload as a function of a resource variance. In this simulation, the average of the available resources of the devices is μ = 1.6. This means that the devices 2 on average have fewer resources than necessary to perform both analytical tasks in the higher operating mode ^¬ . The variance σ ² ranges from [0 to 2] depending on the simulation point.

The average results are, as expected, lower in this case due to the reduced resource values. The differences or differences between the individual resource management (curve I) and an optimum Ressourcenma ^¬ management (curve III) are also greater because the po ^¬ tentielle benefit of optimization increased.

Figures 7, 8 show an average value of the cost function. The distribution of the cost values for covering different methods ^¬ NEN is also relevant. This is illustrated in the diagram of FIG. 9 for the second simulation. Fig. 9 shows the distribution of the values of the cost function for different methods used in the second simulation. As can be seen from FIG. 9, the performance of the data analysis with the method according to the invention (in Fig. 9 II right) better predictable. The results of the resource allocation method according to the invention (FIG. 9 II) approximate the results of an optimal resource allocation (FIG. 9 III). For comparison, FIG. 9 shows the results for an individual resource management.

The main advantages of the method according to the invention compared to conventional allocation methods are as follows.

With the method according to the invention, it is possible to achieve global optimization by performing simpler local optimizations to directly improve the quality of workflows w, especially data processing workflows. The inventive method has a low complexity. Therefore, it is able to respond very quickly to variations or changes in resource conditions and system conditions. The method according to the invention does not require a central coordination unit. Moreover, the method according to the invention scales well with the size of the distributed system 1. In the method according to the invention, it is also not necessary to transmit internal device information, such as the resources available in a device 2, to other devices 2 of the system 1. The inventiveness proper procedure provides a nearly optimal allocation of resources and resource allocation within the distributed system 1, that is an almost optimal utilization of standing for Availability checked ^¬ supply resources of the distributed system 1 for the execution of workflows w. The inventive method is suited in particular for a distributed system 1 with several ^¬ ren distributed devices 2, in particular field devices.

Claims

claims

1. A method for allocating resources of devices (2) of a distributed system (1),

wherein each distributed system device (2) allocates its locally available resources (5) to tasks in response to external execution quality metrics received by the device (2) from other distributed system devices (2).

2. The method according to claim 1,

wherein each device (2) comprises the distributed system (1) a ^pre-¬ passed optimization function (F) that are performed by local execution quality metrics with each task of a group of tasks of the respective device (2) are executable, and from external

Execution quality metrics associated with a set of tasks executable by other devices (2) of the distributed system (1) are dependent.

3. The method according to claim 1 or 2,

wherein each device (2) of the distributed system (1) for each task of the group of tasks that are through this device (2) ^¬ be moved from having relationship data obtained by a relationship between a local execution quality metric with WEL rather the task the respective device (2) is carried out who can ^¬, and the resources (5) of the device (2) which are assigned to the respective task, specify.

4. The method according to any one of the preceding claims 1 to 3, wherein each device (2) of the distributed system (1) performs the following ^¬ steps (S):

(a) receiving (Sl) from other devices (2 ') of the distributed system (1) of a group of external ones

Execution quality metrics, carried out with tasks that belong to the group of tasks that can be executed by other devices (2 ') of the distributed system (1) ^¬; (b) deriving (S2) of a local function (LF), which depends on the local execution quality metrics with each task of the group of tasks that are executable by the per ^¬ weiligen device (2) is executed from a predetermined optimization function (F) based on the external execution quality metrics with which tasks belonging to the group of tasks executable by other devices (2 ') of the distributed system (1) are executed;

(c) performing (S3) a local optimization step on the basis of the derived local function (LF), the relationship data of the respective device (2) and taking into account the available resources (5) of the respective device (2) for the selection of resources (5) the respective device (2) assigned to the tasks of the group of tasks executable by the respective device (2); and

(d) transmitting (S4) each resultant execution quality metric by which each task of the group of tasks executed by the respective device (2), after an updated resource allocation, to each other device (2 ') having a predetermined Op ^{¬ has} trimming function (F) including said execution quality metrics.

5. The method according to claim 4,

the other tasks being related to the tasks executable by the respective device (2) and included in at least one shared workflow comprising a set of related tasks.

6. The method according to claim 4 or 5,

wherein each device (2) of the distributed system (1) performs the local optimization step as soon as resource allocation at the respective device (2) is changed or locally in resources available to the device (2) or as soon as the device (2) receives an update of at least one execution quality metric from another device (2 ^λ ) of the distributed system (1).

7. The method according to any one of the preceding claims 1 to 6, wherein the predetermined optimization function (F) has a global function of the entire distributed system (1).

8. The method according to any one of the preceding claims 1 to 7, wherein the resources (5) of a device (2) of the distributed system (1) processor resources, memory resources and / or data transmission resources.

9. The method according to any one of the preceding claims 1 to 8, wherein the execution quality metrics indicate a quality with which a task is performed by the device (2), wherein the execution quality metrics have a sensitivity, egg ^¬ ne accuracy and / or an error rate.

10. The method according to any one of the preceding claims 1 to 9, wherein an executable by a device (2) task in different ^¬ operating modes of the device (2) is executable, where ^¬ in the operating modes each have a different set of resources (5 ) of the device (2).

11. The method according to claim 10,

wherein for each task a set of permissible modes of operation of a device (2) is provided.

12. The method according to claim 11,

wherein the selection of resources (5) of a device (2) for executing a task by switching to another permissible operating mode of the device (2).

13. A resource allocation device (6) for allocating resources (5) of a device (2) within a distributed system (1), wherein the resource allocating device (6) allocates locally available resources (5) of the device (2) to tasks in response to external execution quality metrics that the device (2) receives from other devices (2 ') of the distributed system (1).

14. Resource allocation device according to claim 13,

Execution quality metrics belonging to a group of tasks executable by other devices (2 ') of the distributed system (1),

wherein each device (2) of the distributed system (1) for each task of the group of tasks that are through this device (2) ^¬ be moved from having relationship data obtained by a relationship between a local execution quality metric with WEL rather the task the respective device (2) can be performed, and the resources (5) of the device (2) which of the depending ^¬ weiligen task are assigned to specify

wherein the resource allocation device (6) of the device (2) comprises:

(a) a receiving unit (6A), which is capable of at ^¬ thereof units (2 ') of the distributed system (1) to receive a group of external execution quality metrics with which tasks belong to the group of tasks which by other devices (2 ') of the distributed system (1) are executable;

(b) a calculation unit (6B) capable of having a local function (LF) derived from the local

Depends on performance quality metrics that each task of the group of tasks executable by the respective device (2) is executing derive predetermined optimization function (F) on the basis of the external execution quality metrics with which tasks belonging to the group of tasks executable by other devices (2 ') of the distributed system (1) are executed,

wherein the calculation unit (6B) is further adapted to a local optimization step based on the derived local function (LF) and the relationship data of the respective device (2) taking into account the available resources (5) of the respective resource selection device (2) perform 5) of the respective device (2) which are assigned to the tasks of the group of tasks which are executable by the jeweili ^¬ ge device (2); and

(c) a transmitting unit (6C), which is suitable for each resul ^¬ animal end exemplary quality metric, with each task of the group of tasks by the respective device (2) is carried out for an updated resource allocation to any other device (2 ' ) having a predetermined optimization function (F) including said execution quality metric.

15. Resource allocation device (6) according to one of claims 13 or 14,

wherein the resource allocating device (6)

a storage unit for storing the local

Execution quality metrics and the external

Execution quality metrics and another storage unit for storing all by the device (2) within one

Workflows executable tasks.

16. Distributed system (1) comprising a plurality of devices (2), each having resources (5) for processing tasks and each having a resource allocation device (6) according to any one of the preceding claims 13 to 15.