WO2024119513A1

WO2024119513A1 - Device and method for agent for dynamically adapting explicit congestion notification configuration in network system

Info

Publication number: WO2024119513A1
Application number: PCT/CN2022/138121
Authority: WO
Inventors: Guillermo Bernardez GIL; Jose SUAREZ-VARELA; Albert CABELLOS-APARICIO; Pere BARLET-ROS; Xiangle CHENG; Xiang Shi; Shihan XIAO
Original assignee: Huawei Technologies Co., Ltd.
Priority date: 2022-12-09
Filing date: 2022-12-09
Publication date: 2024-06-13

Abstract

An agent for dynamically adapting an Explicit Congestion Notification (ECN) configuration in a network system is provided. The agent is configured to send a first message to one or more neighbor agents, the first message comprising a property of a switch port associated with the agent. The agent is further configured to receive a second message from each neighbor agent, the second message comprising a property of a neighbor switch port associated with the neighbor agent. The agent is further configured to determine the ECN configuration for the switch port associated with the agent based on the property of the switch port and the properties of the neighbor switch ports. A property of a switch port may comprise at least one of a transmission rate, a queue length, and a rate of ECN marked packets at the respective switch port.

Description

DEVICE AND METHOD FOR AGENT FOR DYNAMICALLY ADAPTING EXPLICIT CONGESTION NOTIFICATION CONFIGURATION IN NETWORK SYSTEM

TECHNICAL FIELD

The present disclosure relates to an agent for dynamically adapting an Explicit Congestion Notification (ECN) configuration in a network system, the network system comprising a plurality of agents, and each agent being associated with a switch port. The disclosure further provides a corresponding method and a computer program to perform the method.

BACKGROUND

Congestion Control (CC) plays a fundamental role in optimizing traffic in Data Center Networks (DCNs) . Currently, DCNs implement two main CC protocols: Data Center Transmission Control Protocol (DCTCP) and Data Center Quantized Congestion Notification (DCQCN) . Both protocols, as well as their main variants, are based on ECN, wherein intermediate switches mark packets when they detect congestion. The ECN configuration is thus a crucial aspect on the performance of widely deployed CC protocols.

Nowadays, network experts set static ECN parameters carefully selected to optimize the average network performance (e.g., throughput, latency) . However, high-speed DCNs experience quick and abrupt changes that severely change the network state (e.g., dynamic traffic workloads, incast events and failures) . This leads to under-utilization and sub-optimal performance.

CC has been extensively studied. As a result, there exists a number of solutions for DCNs tackling the problem from different angles, such as round-trip time (RTT) -based, credit-based, or telemetry-based mechanisms. Currently, most production DCNs implement the two main well-established CC protocols DCTCP and DCQCN. The former is the main standard in traditional networks based on the TCP/IP stack, while the latter is the standard in modern RDMA-based networks. Both mechanisms, as well as their enhanced schemes, rely on ECN, so that switches mark packets when they experience congestion, and end-hosts dynamically adapt their transmission rate according to this congestion feedback. The most advanced conventional CC mechanisms, such as High Precision Congestion Control (HPCC) or Swift, have been successfully validated in production networks, showing a good performance compared to their contemporaries. However, these solutions require features that are not widely supported by legacy datacenter equipment, such as in-band network telemetry or accurate real-time RTT measurements.

Some of the proposed CC mechanisms show a very good performance in DCNs. However, they rely on novel network architectures and/or protocol stacks that are not supported by most legacy switches deployed in current DCNs. The most widely deployed CC standards are DCTCP in networks running the TCP/IP stack, and DCQCN in RDMA-based networks. Both of them, nevertheless, rely on static ECN-based mechanisms that do not adapt the marking ECN threshold depending on the observed traffic.

SUMMARY

In view of the above, an objective of this disclosure is to improve conventional solutions for dynamically adapting ECN configuration in network systems. Another objective is to provide an agent and a corresponding method that can operate successfully under traffic and topology changes. Once it is deployed, an objective is that the agent is enabled to cooperate with neighbor agents to optimize a global property of the network. Further, an objective is to make the agent compatible with any datacenter network running ECN-based protocols, such as DCTCP or DCQCN.

These and other objectives are achieved by the solutions provided in the independent claims. Advantageous implementations are further defined in the dependent claims.

According to a first aspect, an agent for dynamically adapting an ECN configuration in a network system is provided. The network system includes a plurality of agents, each agent being associated with a switch port. The agent is configured to: send a first message to one or more neighbor agents, where the first message includes a property of the switch port associated with the agent; receive a second message from each neighbor agent of the one or more neighbor agents, the second message including a property of a neighbor switch port associated with the neighbor agent; and determine the ECN configuration for the switch port associated with the agent based on the property of the switch port and the properties of the neighbor switch ports.

This provides the advantage that the agent can independently adapt the ECN configuration of its associated port taking into consideration the internal properties (or attributes) of the switch port and the received communications (second messages) including the properties of neighboring ports.

In an implementation form of the first aspect, the property of the switch port includes at least one of: a transmission rate, a queue length, and a rate of ECN marked packets of the switch port; and/or the property of the neighbor switch port includes at least one of a transmission rate, a queue length, and a rate of ECN marked packets of the neighbor switch port.

In an implementation form of the first aspect, before sending the first message to one or more neighbor agents, the agent is configured to collect the property of the switch port associated with the agent.

The property of the switch port is commonly supported by commercial switches and can be locally obtained with low computational overhead at microsecond timescales.

In an implementation form of the first aspect, the agent is further configured to retrieve, from the second message, the property of the neighbor switch port associated with the neighbor agent.

This may enable the agent to additionally determine relevant information of the one or more switch ports received in the second messages.

In an implementation form of the first aspect, the agent is further configured to update the property of the switch port associated with the agent by combining the property of the switch port and the properties of the one or more neighbor switch ports retrieved from the second message.

This provides the advantage that the agent may acquire valuable context from the rest of agents, which may be taken into account to adapt the ECN configuration of its respective switch port.

In an implementation form of the first aspect, the agent is further configured to generate an updated first message, where the updated first message includes the updated property of the switch port associated with the agent.

Thereby, the agent may be able to further share the property of the switch port associated with it and that already includes information of the properties of the neighbor agents, to the one or more neighbor agents.

In an implementation form of the first aspect, before the determination of the ECN configuration for the switch port associated with the agent based on the property of the switch port and the properties of the one or more neighbor switch ports, the agent is further configured to perform an iteration procedure including performing a predefined number of iterations.

In each iteration the following steps are performed: sending the updated first message generated in a previous iteration to the one or more neighbor agents, or in the case of the first iteration send the first message to the one or more neighbor agents, where the updated first message includes the updated property of the switch port associated with the agent; receiving an updated second message from each of the one or more neighbor agents generated in the previous iteration, or in the case of the first iteration receiving the second message from each of the one or more neighbor agents, where the updated second message includes an updated property of the neighbor switch port associated with the neighbor agent; retrieving, from the updated second message, the updated property of the neighbor switch port; updating the property of the switch port associated with the agent by combining the updated property of the switch port calculated in the previous iteration and the updated property of each neighbor switch port retrieved from the one or more second messages; and generating another updated first message including the updated property of the switch port associated with the agent.

At each message passing iteration, the property of the switch port associated with the agent is updated so that valuable context to the rest of agents may be provided, facilitating cooperation between the agents and enabling each agent to adapt the ECN configuration of its corresponding switch port accordingly.

Notably, the predefined number of iterations is a low number, for example 2 or 3 iterations, because a diameter of conventional DCNs is limited. Further, the predefined number of iterations may be a constant value and, thus, there is no need to change the number of iterations when changes in traffic or in topology of the network system occur.

In an implementation form of the first aspect, the determination of the ECN configuration for the switch port associated with the agent based on the property of the switch port and the properties of the neighbor switch ports includes determining an optimal ECN configuration for the switch port based on the final property of the switch port, where the final property of the switch port includes the updated property of the switch port after the last iteration; alternatively, in the case of the first iteration, determining an optimal ECN configuration for the switch port based on the property of the switch port and the properties of the neighbor switch ports received in the one or more second messages.

In an implementation form of the first aspect, the determined optimal ECN configuration for the switch port includes an ECN configuration for the switch port that optimizes the performance of the network system.

This provides the advantage that the agent may cooperate with the one or more neighbor agents to efficiently adapt the ECN configuration of the switch port associated with it to the fast traffic dynamics of conventional DCNs, and that enables to optimize a global flow-aware goal of the network system

In an implementation form of the first aspect, the ECN configuration for the switch port that optimizes the performance of the network system includes at least one of an ECN configuration for the switch port that decreases a Flow completion time (FCT) of the network system, an ECN configuration for the switch port that reduces a buffer occupancy of the network system, and an ECN configuration for the switch port that increases a throughput of the network system.

In an implementation form of the first aspect, the network system is configured based on a DCTCP protocol or a DCQCN protocol.

That is, the agent according to this disclosure is compatible with any datacenter network running DCTCP or DCQCN protocols.

In the first aspect and its implementations, the functions described may be implemented in hardware, software, firmware, or any combination thereof.

According to a second aspect, a method for an agent for dynamically adapting an ECN configuration in a network system is provided. The network system includes a plurality of agents, each agent being associated with a switch port, and the method includes: sending a first message to one or more neighbor agents, the first message including a property of the switch port associated with the agent; receiving a second message from each neighbor agent of the one or more neighbor agents, the second message including a property of a neighbor switch port associated with the neighbor agent; and determining the ECN configuration for the switch port associated with the agent based on the property of the switch port and the properties of the neighbor switch ports.

In an implementation form of the second aspect, the property of the switch port includes at least one of: a transmission rate, a queue length, and a rate of ECN marked packets of the switch port; and/or the property of the neighbor switch port includes at least one of a transmission rate, a queue length, and a rate of ECN marked packets of the neighbor switch port.

In an implementation form of the second aspect, before sending the first message to one or more neighbor agents, the method further comprises collecting the property of the switch port associated with the agent.

In an implementation form of the second aspect, the method further comprises retrieving, from the second message, the property of the neighbor switch port associated with the neighbor agent.

In an implementation form of the second aspect, the method further comprises updating the property of the switch port associated with the agent by combining the property of the switch port and the properties of the one or more neighbor switch ports retrieved from the second message.

In an implementation form of the second aspect, the method further comprises generating an updated first message, where the updated first message includes the updated property of the switch port associated with the agent.

In an implementation form of the second aspect, before determining the ECN configuration for the switch port associated with the agent based on the property of the switch port and the properties of the one or more neighbor switch ports, the method further comprises performing an iteration procedure, including performing a predefined number of iterations.

In an implementation form of the second aspect, determining the ECN configuration for the switch port associated with the agent based on the property of the switch port and the properties of the neighbor switch ports includes determining an optimal ECN configuration for the switch port based on the final property of the switch port, where the final property of the switch port includes the updated property of the switch port after the last iteration; alternatively, in the case of the first iteration, determining an optimal ECN configuration for the switch port based on the property of the switch port and the properties of the neighbor switch ports received in the one or more second messages.

In an implementation form of the second aspect, the determined optimal ECN configuration for the switch port includes an ECN configuration for the switch port that optimizes the performance of the network system.

In an implementation form of the second aspect, the ECN configuration for the switch port that optimizes the performance of the network system includes at least one of an ECN configuration for the switch port that decreases a FCT of the network system, an ECN configuration for the switch port that reduces a buffer occupancy of the network system, and an ECN configuration for the switch port that increases a throughput of the network system.

In an implementation form of the second aspect, the network system is configured based on a DCTCP protocol or a DCQCN protocol.

That is, the method according to this disclosure is compatible with any datacenter network running DCTCP or DCQCN protocols.

According to a third aspect, a computer program is provided, instructions, which when the program is executed by a computer, cause the computer to perform the method according to the second aspect and its implementation forms.

The computer program product according to the third aspect includes the features of the corresponding implementation forms of the method of the second aspect.

The method according to the second aspect and the computer program product according to the third aspect and their implementation forms provide the same advantages and effects as described above for the agent of the first aspect and its respective implementation forms.

The solutions according to this disclosure provide a general agent for dynamically adapting the ECN configuration in a network system that achieves the following advantages: the agent is compatible with widely deployed ECN-based CC protocols (e.g., DCTCP, DCQCN) and, thus, may be directly deployed in other network systems with different traffic distributions. Moreover, the agent may determine the ECN configuration of its respective switch port, enabling the switch port to take actions being aware of the property of its neighbor switch ports. This may achieve a global cooperation among the agents through message communications between neighboring switches, as well as a global reward, which in turn may provide with a better optimization potential than conventional solutions based on local rewards. Further, the agent may be implemented by neural network architectures, that may allow to trained the network system in controlled testbeds and then deploy it directly in real DCNs thereby greatly increasing the commercialization viability of the agent. Further, this disclosure may provide a distributed in-network agent for CC optimization in a network system that may be scalable and robust against possible failures and reconfigurations with respect to centralized solutions, thereby reducing complexity.

It has to be noted that all devices, elements, units and means described in the present application could be implemented in the software or hardware elements or any kind of combination thereof. All steps which are performed by the various entities described in the present application as well as the functionalities described to be performed by the various entities are intended to mean that the respective entity is adapted to or configured to perform the respective steps and functionalities. Even if, in the following description of specific embodiments, a specific functionality or step to be performed by external entities is not reflected in the description of a specific detailed element of that entity which performs that specific step or functionality, it should be clear for a skilled person that these methods and functionalities can be implemented in respective software or hardware elements, or any kind of combination thereof.

BRIEF DESCRIPTION OF DRAWINGS

The above described aspects and implementation forms of the present disclosure will be explained in the following description of specific embodiments in relation to the enclosed drawings, in which:

FIG. 1 shows a schematic diagram of an agent for dynamically adapting an ECN configuration in a network system according to this disclosure;

FIG. 2 schematically shows an exemplary network system including a plurality of agents, each agent being associated with a switch port, according to this disclosure;

FIG. 3 shows an exemplary flowchart for dynamically adapting an ECN configuration where the agent is implemented by a message passing neural network, according to this disclosure;

FIG. 4 an exemplary flowchart for dynamically adapting an ECN configuration where the agent is implemented in an artificial intelligence chip, according to this disclosure;

FIG. 5 shows a method for an agent for dynamically adapting an ECN configuration in a network system according to this disclosure.

DETAILED DESCRIPTION OF EMBODIMENTS

FIG. 1 shows a schematic diagram of an agent 100 for dynamically adapting an ECN configuration in a network system 1. The network system 1 comprises a plurality of

agents

100, 110 and each

agent

100, 110 is associated with a

switch port

102, 112. In the embodiment of FIG. 1, the agent 100 is associated with the switch port 102, and each of the one or more neighbor agents 110 is associated with a switch port 112.

The agent 100 is configured to send a first message 104 to one or more neighbor agents 110. The first message 104 comprises a property 101 of the switch port 102 associated with the agent 100. The property 101 of the switch port 102 comprises any port-level metrics available at the switch port 102 that may help to optimize a global property of the network system 1, as it will be explained later in this description. Thus, the property 101 of the switch port 102 comprises, for example but not as a limitation, at least one of a transmission rate, a queue length, and a rate of ECN marked packets of the switch port 102. The first message 114 may be generated based on the property of the switch port 102, for example, by a processing block of the agent 100 (optional, thus shown in dashed lines) , which obtains the property 101 of the switch port 102. The agent 100 may comprise some internal attributes based on relevant metrics of its associated port 102 (e.g., utilization, queue length, etc. ) .

The agent 100 may be configured to collect the property 101 of its associated switch port 102 available at the switch port 102, for instance, at the optional processing block as indicated.

The property 101 of the switch port is commonly supported by commercial switches and can be locally obtained with low computational overhead at microsecond (μs) timescales.

Then, the agent 100 is configured to receive a second message 114 from each neighbor agent 110 of the one or more neighbor agents 110. The second message 114 comprises a property 103 of a neighbor switch port 112 associated with the neighbor agent 110. The property 103 of the neighbor switch port 112 comprises any port-level metrics available at the neighbour switch port 112 that may help to optimize a global property of the network system 1, for example but not as a limitation, at least one of a transmission rate, a queue length, and a rate of ECN marked packets of the neighbor switch port 112. Thus, each neighbor agent 110 may comprise some internal attributes based on relevant metrics of its associated port 112 (e.g., utilization, queue length, etc. ) . At the neighbor agent 110, the second message 114 may be generated in a similar manner based on the property 103 of the neighbor switch port 112, than the first message 104 is generated at the agent 100 (e.g., by an optional processing block at the neighbor agent 110) .

The agent 100 may be further configured to retrieve, from each of the received second messages 114, the property 103 of the neighbor switch port 112 associated with the neighbor agent 110.

The agent 100 is configured to determine the ECN configuration 106 for the switch port 102 associated with it, based on the property 101 of the switch port 102 and the properties 103 of the one or more neighbor switch ports 112 received in the plurality of second messages 114 (in FIG. 1, as an example, only one neighbor agent 110 is shown) .

Thereby, each

agent

100, 110 independently adapts the ECN configuration of its associated

port

102, 112, taking into consideration its internal properties (or attributes) and the received communications (second messages) from neighboring agents 110.

Before the determination of the ECN configuration 106 for the switch port 102 associated with the agent 100 based on the property 101 of the switch port 102 and the properties 103 of the one or more neighbor switch ports 112, the agent 100 may be configured to update the property 101 of the switch port 102 associated with the agent 100 by combining the property 101 of the switch port 102 and the properties 103 of the one or more neighbor switch ports 112 retrieved from the second message 114.

Combining the property 101 of the switch port 102 associated with the agent 100 and the properties 103 of the one or more neighbor switch ports 112 retrieved from the second message 114 may comprise, for example, aggregating the properties 103 of each of the one or more neighbor switch ports 112, This can be done through an element-wise operation (e.g. element-wise min or max) .

Then, the agent 100 may be configured to generate an updated first message 104. The updated first message 104 may comprise the updated property 101 of the switch port 102 associated with the agent 100.

The agent 100 is further configured to perform an iteration procedure. The iteration procedure comprises performing a predefined number of iterations.

Notably, the predefined number of iterations is a low value, for example 2 or 3 iterations, because a diameter of conventional DCNs is limited. For example, for a DCN up to three layers, a suitable number of iterations may be equal to three. Further, the predefined number of iterations may be a constant value; hence, there is no need to change the number of iterations when changes in traffic or in topology of the network system 1 occur.

Each iteration comprises sending, by the agent 100, the updated first message 104 to the one or more neighbor agents 110, and receiving, by the agent 100, an updated second message 114 from each of the one or more neighbor agents 110. At the first iteration, the updated first message 104 comprises the first message 104 and the updated second message 114 comprises the second message 114, as disclosed above. Otherwise, the updated first message 104 comprises the updated property 101 of the switch port 102 associated with the agent 100 and that is generated in a previous iteration, and the updated second message 114 comprises an updated property 103 of the neighbor switch port 112 associated with the neighbor agent 110 that is generated in a previous iteration.

Further, each iteration comprises retrieving, by the agent 100, from the updated second message 114, the updated property 103 of the neighbor switch port 112. Then, each iteration comprises updating, by the agent 100, the property 101 of the switch port 102 associated with the agent 100 by combining the updated property 101 of the switch port 102 calculated in the previous iteration and the updated property 103 of each neighbor switch port 112 retrieved from the one or more second messages 114.

Next, each iteration comprises generating, by the agent 100, another updated first message 104 comprising the updated property 101 of the switch port 102 associated with the agent 100.

The determination of the ECN configuration 106 for the switch port 102 associated with the agent 100 based on the property of the switch port 102 and the properties 103 of the neighbor switch ports 112 comprises determining, by the agent 100, an optimal ECN configuration for the switch port 102 based on the final property of the switch port 102, where the final property 101 of the switch port 102 comprises the updated property 101 of the switch port 102 after the last iteration.

Alternatively, in the case of the first iteration, the determination of the ECN configuration 106 for the switch port 102 associated with the agent 100 based on the property of the switch port 102 and the properties 103 of the neighbor switch ports 112 comprises determining an optimal ECN configuration for the switch port 102 based on the property 101 of the switch port 102 and the properties 103 of the neighbor switch ports 112 received in the one or more second messages 114.

Additionally or alternatively, the determination of the ECN configuration 106 for the switch port 102 associated with the agent 100 based on the property of the switch port 102 and the properties of the neighbor switch ports 112 may comprise determining, by the agent 100, an optimal ECN configuration for the switch port 102 associated with it based on the property 101 of the switch port 102 updated in each iteration.

Thereby, at each message passing iteration, the property 101 of the switch port 102 is optimized so as to provide with valuable context to the neighbor agents 110, facilitating cooperation between them and enabling each

agent

100, 110 to optimally adapt the ECN settings 106 of its

respective switch port

102, 112 accordingly. That is, each

agent

100, 110 can independently adapt the ECN configuration of its associated

port

102, 112, taking into consideration its internal attributes and the properties 103 of the one or more neighbor switch ports 112 received in the communications exchanged with the neighbor agents 110. Those internal attributes may be used to update the properties of each

switch port

102, 112, and the

agents

100, 110 actually share the updated

properties

101, 103 of the

respective switch port

102, 112 to their neighbors 110 via

messages

104, 114 over the network 1.

The determined optimal ECN configuration 106 for the switch port 102 comprises an ECN configuration 106 for the switch port 102 that optimizes the performance of the network system 1.

The distributed ECN configuration 106 for the switch port 102 makes the solution according to this disclosure highly scalable to large networks.

The ECN configuration 106 for the switch port 102 that optimizes the performance of the network system 1 comprises at least one of an ECN configuration 106 for the switch port 102 that decreases a flow completion time (FCT) of the network system 1, an ECN configuration for the switch port 102 that reduces a buffer occupancy of the network system 1, and an ECN configuration for the switch port 102 that increases a throughput of the network system 1.

On the determination of the optimal ECN configuration 106 for the switch port 102 based on the property 101 of the switch port 102 and the properties 103 of the neighbor switch ports 112 received in the one or more second messages 114, the agent 100 may use any suitable optimization method that enables to optimize a global flow-level metric of the network system 1 (e.g., the FCT, the buffer occupancy and/or the throughput of the network 1) . Thereby, the message communications between the plurality of

agents

100, 110 of the network system 1 enable a global cooperation between the

agents

100, 110, as they can take actions based not only on the local properties of the

respective switch ports

102, 112, but also on the properties from their surrounding (neighbor) switch

ports

102, 112.

Further, it is mentioned that each of the one or more neighbor agents 110 is configured to perform exactly the same process as described above for the agent 100, thereby all the

agents

100, 110 move from a local to a global context awareness of the network 1.

The network system 1 is configured based on a DCTCP protocol or a DCQCN protocol. That is, the plurality of

agents

100, 110 according to this disclosure are compatible with any datacenter network running DCTCP or DCQCN protocols.

By dynamically adapting ECN configuration of the switch port 102, the agent 100 according to this disclosure may further optimize CC in the network system 1

The agent 100 according to this disclosure, thus, can optimize the handling of congestion notifications within

switches

102, 112, which is a crucial component of CC protocols to efficiently optimize traffic in the network system 1. In this context, both DCTCP and DCQCN protocols implement a similar approach: switches mark a Congestion Experienced (CE) bit of packets in case the queue length exceeds some predefined thresholds.

DCTCP protocol adopts a hard cutoff, for example that all packets are marked when the queue length exceeds a certain value k. Instead, the DCQCN protocol implements a softer RED-like probabilistic approach with three configuration parameters {k _min, k _max, p _max} given in equation (1) :

where q _len is the instantaneous queue length in the port, and p _mark is the probability for marking a packet.

Thus, when the network system 1 according to this disclosure is configured based on a DCTCP protocol, the determined ECN configuration 106 for the switch port 102 associated with the agent 100 may comprise, for example and not as a limitation, a value k that optimizes the performance of the network system 1 (i.e., that optimizes at least one of the FCT, the buffer occupancy and the throughput of the network system 1) .

When the network system 1 according to this disclosure is configured based on a DCQCN protocol, the determined ECN configuration 106 for the switch port 102 associated with the agent 100 may comprise, for example and not as a limitation, values of the quantities k _min, k _max, p _max and/or p _mark that optimize the performance of the network system 1 (i.e., that optimize at least one of the FCT, the buffer occupancy and the throughput of the network system 1) .

The agent 100 may comprise processing circuitry (not shown) configured to perform, conduct or initiate the various operations of the agent 100 described herein. The processing circuitry may comprise hardware and software. The hardware may comprise analog circuitry or digital circuitry, or both analog and digital circuitry. The digital circuitry may comprise components such as application-specific integrated circuits (ASICs) , field-programmable arrays (FPGAs) , digital signal processors (DSPs) , or multi-purpose processors. The agent 100 may further comprise memory circuitry, which stores one or more instruction (s) that can be executed by the processor or by the processing circuitry, in particular under control of the software. For instance, the memory circuitry may comprise a non-transitory storage medium storing executable software code which, when executed by the processor or the processing circuitry, causes the various operations of the agent 100 to be performed. In one embodiment, the processing circuitry comprises one or more processors and a non-transitory memory connected to the one or more processors. The non-transitory memory may carry executable program code which, when executed by the one or more processors, causes the agent 100 to perform, conduct or initiate the operations or methods described herein.

In this exemplary embodiment, the

agents

100, 110 may enable in-network optimization at

intermediate switch ports

102, 112. Further, the agent 100 may provide with a fully decentralized mechanism. The only observable mechanism in the network system 1, thus, may be the message communications between neighboring

agents

100, 110. The first message 104 and the second message 114 provide contextual information to the agent 100 and, moreover, a pattern at which first message 104 and the second message 114 are sent may be recognizable.

The

agents

100, 110 of this exemplary embodiment may reduce the network system 1 complexity by providing with a solution that is more scalable and robust against possible failures and reconfigurations with respect to conventional centralized solutions. Further, as the agent 100 is compatible with widely deployed ECN-based CC protocols (e.g., DCTCP or DCQCN) , the agent 100 may be directly deployed in other network systems 1 with different traffic distributions, increasing the efficiency and the commercialization viability of the agent 100.

Notably, the agent 100 may provide with topology-aware local context, as it may enable its associated switch port 102 to take actions (for example, to determine whether a packet may be marked) while being aware of the properties 103 of its neighbor switch ports 112. Thereby, global cooperation is achieved through message communications between the agent 100 and the one or more neighbor agents 110. Additionally, a global reward may be achieved, since the agent 100 may determine the ECN configuration for the switch port 102 that optimizes a goal of the whole network system 1, which may achieve better results than conventional solutions based on local rewards.

FIG. 2 schematically depicts an exemplary network system 1 according to this disclosure. Same elements are labelled with the same reference signs.

The exemplary network system 1 comprises a plurality of

agents

100, 110, exemplary agents A ₁ to A ₉, and each

agent

100, 110 is associated with a

switch port

102, 112. Each agent 100, exemplary agent A ₄, may be configured to send a first message 104 to one of more neighbor agents 110, exemplary agents A ₂, A ₃, A ₅, A ₇ and A ₉, and to receive a second message from each of the one or more neighbor agents 110 A ₂, A ₃, A ₅, A ₇ and A ₉, where the first message 104 comprises a property 101 of the switch port 102 associated with the agent 100 and each of the second messages 114 comprises a property 103 of the neighbor switch port 112 associated with the neighbor agent 110.

In the network system 1 according to this disclosure, each

port

102, 112 can be an egress port. Further, there can be as

many agents

100, 110 as ports in each switch, so that each

agent

100, 110 can control the ECN configuration of a particular port.

In the example of Fig. 2, each agent 100, exemplary agent A ₄, may also send the first message 104 to other exemplary agents, for examples to agents A ₁, A ₆ and/or A ₈, and may also receive the second message from them. This is not limited in this disclosure.

The exemplary agents A ₁, A ₄, A ₆ and A ₈ may form a spine of the network system 1, and the exemplary agents A ₂, A ₃, A ₅, A ₇ and A ₉ may form a leaf of the network system 1.

The exemplary network system 1 of FIG. 2 may be configured based on a DCTCP protocol or a DCQCN protocol.

The agent 100 according to this disclosure may be implemented in a type of Graph Neural Networks (GNN) called Message Passing Neural Network (MPNN) . This MPNN architecture may help the

agents

100, 110 to process and model information of the state of the network system 1, and may enable propagating local information over the whole network system 1 by allowing message communications between

agents

100, 110, so that they can update their internal attributes depending on the second message 114 received from each neighbor agent 110. Furthermore, there is a direct mapping between the MPNN execution and the actual network system , namely that switch

ports

102, 112 can be represented as node/link entities in the GNN, and the information exchanged internally by port-

level agents

100, 110 can be transmitted via

messages

104, 114 sent through the network system infrastructure.

Then, by providing proper rewards for the actions that the agents 100 can take, each port-

level agent

100, 110 may eventually learn what information is relevant to exchange with its neighbor agents 110, and may also discover a manner to adapt the ECN configuration 106 of tis corresponding switch port 102 to optimize the global objective (e.g., minimize the FCT, minimize the buffer occupancy, and/or maximize the throughput of the network system 1) . As all the

agents

100, 110 have the same internal parameters (i.e., the

respective switch ports

102, 112 have the same type of properties) , the

agents

100, 110 can optimize the

property

101, 103 of the

respective switch ports

102, 112 jointly. After training, the agent 100 implemented in MPNN can be replicated and deployed across distributed devices of any other DCN.

FIG. 3 depicts an exemplary embodiment of a flowchart performed by the agent 100 according to this disclosure. Same elements are labelled with the same reference signs. In the embodiment of FIG. 3, the agent 100 is implemented in a MPNN

In the embodiment of FIG. 3, each of the plurality of

agents

100, 100, i.e., the agent 100 and the one or more neighbor agents 110, may be implemented in a MPNN and may achieve the same processes as disclosed in the following for the agent 100. The first message 104 may comprise the property 101 of the switch port 102 associated with the agent 100 in the form of a hidden state of the agent 100, and each of the second messages 114 may comprise the property 103 of the neighbor switch port 112 associated with the neighbor agent 110 in the form a hidden state of the neighbor agent 110.

In this exemplary embodiment, the agent 100 may comprise three main modules: an Information Retrieval Module 302, a Hidden State Generation Module 304, and an Action Selection Module 306.

The Information Retrieval Module 302 may receive the second message 114 from each of the one or more neighbor agents 110. The Information Retrieval Module 302 can be implemented by a Neural Network (NN) module that may be configured to process the second messages 114 received from the one or more neighbor agents 110, and to further select relevant information of the hidden states of the neighbor agents 110 that should be considered.

The Information Retrieval Module 302 may be further configured to aggregate the selected relevant information of the hidden states of the neighbor agents 110. The aggregation can be done, for example, through an element-wise operation (for example element-wise min or max) .

Further, the Information Retrieval Module 302 may be configured to output the aggregated information of the hidden states of the neighbor agents 110.

The output of the Information Retrieval Module 302 may be used by the Hidden State Generation Module 304. The Hidden State Generation Module 304 can be based on a NN, and can be configured to combine the aggregated selected information of the hidden states of the neighbor agents 110 with the hidden state of the agent 100, thus generating a new or updated hidden state for the agent 100.

The Information Retrieval Module 302 and the Hidden State Generation Module 304 may conform a Communication Module 308 that can be identified to the operation of the MPNN. Variants of the MPNN may consider that the functions executed by the Information Retrieval Module 302 and the Hidden State Generation Module 304 may not necessarily be modeled by a NN.

Then, the Communication Module 308, additionally or alternatively the Hidden State Generation Module 304, may be configured to generate an updated first message 104 comprising the updated hidden state for the agent 100, and to subsequently send the updated first message 104 to the one or more neighbor agents 110. The Communication Module 308 may be executed for a predefined number of iterations.

Each of the one or more neighbor agents 110 may perform exactly the same process, so that all the

agents

100, 110 move from a local to a global context awareness.

After the iterative execution of the Communication Module 308, the Action Selection Module 306 of the agent 100 may be configured to determine an optimal ECN configuration for the switch port 102 associated with the agent 100 based on the final hidden state of the agent 100, which takes into account the updated hidden state of each of the one or more neighbor agents 110. The final hidden state of the agent 100 may comprise the updated hidden state of the agent 100 after the last iteration.

The Action Selection Module 306 can be implemented by a NN that produces, as an output, a final action policy as depicted in FIG. 3, where the final action policy may comprise the ECN configuration 106 of the switch port 102 associated with the agent 100.

The determined optimal ECN configuration for the switch port 102 associated with the agent 100 may comprise an ECN configuration 106 for the switch port 102 that optimizes the performance of the network system 1. The ECN configuration 106 for the switch port 102 that optimizes the performance of the network system 1 comprises at least one of the ECN configuration 106 for the switch port 102 that decreases the FCT of the network system 1, the ECN configuration for the switch port 102 that reduces the buffer occupancy of the network system 1, and the ECN configuration for the switch port 102 that increases the throughput of the network system 1.

In this exemplary embodiment, the hidden state representations of the

agents

100, 100 are central elements, since they may be the actual content of message communications. The hidden states of the

agents

100, 110 need to encode not only the

agent

100, 110 discriminant features required to define its individual behavior, but also relevant information about the context to facilitate the action decision making of neighboring agents 110.

The general flowchart performed by this exemplary embodiment of the agent 100 is now described. In a first step, the agent 100 may collect some port-level metrics available at its corresponding switch port 102, for example and not as a limitation, at least one of the transmission rate or utilization of the switch port 102, the instantaneous queue length and the rate of ECN-marked packets in the switch port 102. These metrics represent the input features depicted in FIG. 3.

In a second step, the agent 100 may initialize its hidden state based on tis corresponding input features. That is, the input features are encoded into a hidden state of the agent 100, and the hidden state of the agent 100 may be a fixed-size vector. The initialization may be performed, for example, by adding the input features and applying zero-padding to fit a dimension (i.e., the fixed size) of the hidden state vector for the agent 100.

In a third step, the agent 100 may sends the first message 104 to the one or more neighbor agents 110. The first message may comprise the hidden state of the agent 100. Given that each

agent

100, 110 represents

egress ports

102, 112, the agent 100 may actually send the first message 104 to all ports 112 that can potentially receive traffic from it.

In a fourth step, the agent 100 may receive the hidden states of all of its neighbor agents 110, and may processes them as follows:

● First, the hidden state of each neighbor agent 110 may be combined with the hidden state of the agent 100 through a Message Function 310. The Message Function 310 can be implemented by a NN module, for example by the Information Retrieval Module 302.

● Then, the combined hidden state of the agent 100 and hidden states of the one of more neighbor agents 110 may be aggregated using a predefined Aggregation Function 312, e.g., an element-wise max or min function.

● Next, the agent 100 may update its own hidden state with the aggregated hidden states. This task can be performed by an Update Function 314, which may be implemented by a NN module, for example by Hidden State Generation Module 304.

Thereby, the agent 100 may generate an updated hidden state of the agent 100 that potentially incorporates information of the one or more neighbor agents 110.

In a fifth step, the message exchange and processing described in the third and fourth steps may be repeated in a predefined number of iterations. In each iteration, the updated hidden states of the one or more neighbor agents 110 received in the one or more second messages 114 may be considered. During these iterations:

● There may be a recognizable pattern of periodic message communications between neighboring agents 110.

● Assuming that the updated hidden state of the agent 100 may add some local topological awareness through the updated hidden state of the neighbor agents 110 received in the second messages 114, a range of the network 1 is expanded at successive message passing iterations. At each iteration, the agent 100 may have access to information of more distant points (agents 110) in the network 1.

● Therefore, it may be expected that the hidden state of the agent 100 and the hidden states of the neighbor agents 110 may evolve from sparse data to much more dense representations in hidden state vectors as the iterations of the message passing procedure are executed.

The number of iterations may be predefined, and typically is a low value, e.g. 2 or 3 iterations, because the diameter of DCNs is limited.

In a sixth step, at the end of the message passing, the agent 100 may individually evaluate its final hidden state through a Readout Function 316. The Readout Function 316 can be implemented by another NN module, for example the Action Selection Module 306.

The Readout Function 316 may provide the agent 100 with its action policy, i.e., with the optimal ECN configuration for the switch port 102 associated with the agent 100.

The exemplary implementation of the agent 100 by MPNN may be trained using, for example and not as a limitation, a Reinforcement Learning algorithm.

In this exemplary embodiment, the agent 100 may be trained offline, (e.g., in a controlled testbed) and then can be directly deployed in production networks, without the need for re-training on premises. Hence, it may avoid uncertainty of online training in production networks. In addition, offline training may enable to perform extensive testing before deployment, with the possibility of issuing certifications determining the operational ranges that the agent 100 can safely support (e.g., link capacities or maximum network size) . This certification process is aligned with the conventional way in which networking products are currently commercialized.

Further, during training, the agent 100 may learn how to communicate with the one or more neighbor agents 110 to dynamically adapt the ECN configuration 106 of its respective switch port 102 that optimizes a global flow-level metric (e.g., FCT) of the network system 1. The message communications may enable a global cooperation between the

agents

100, 110, as they can take actions based not only on their local state information, but also on data from their neighbor agents 110.

In this exemplary embodiment, the agent 100 is compatible with any data center network running DCTCP or DCQCN protocols. That is, when the agent 100 and the neighbor agents 110 are implemented by MPNN, the network system 1 may be configured based on a DCTCP protocol or a DCQCN protocol.

The determined ECN configuration 106 for the switch port 102 associated with the agent 100 may comprise, for example and not as a limitation, a value k that optimizes the performance of the network system 1 configured based on a DCTCP protocol. As a further example, the determined ECN configuration 106 for the switch port 102 associated with the agent 100 may comprise values of the quantities k _min, k _max, p _max and/or p _mark that optimize the performance of the network system 1 configured based on a DCQCN protocol, as depicted in FIG. 3.

Recent works have focused on the use of machine learning (ML) techniques to produce data-driven solutions that can efficiently adapt to the network dynamics. Some of the conventional ML-based CC proposals, including for example AuTO, Aurora or Orca. However, these solutions are not compatible with widely deployed equipment in datacenters, as they propose to re-implement the network stack. A more recent solution, an automatic run-time optimization scheme (ACC) proposes to perform in-network optimization, by dynamically adapting the ECN configuration on switches. This solution has shown a remarkable performance in production environments and it is compatible with current datacenter equipment running widely deployed ECN-based CC protocols (e.g., DCTCP, DCQCN) . Nevertheless, ACC is designed for online training and, thus, it needs to be re-trained to gradually learn how to adapt to the current network conditions. In general, online training is not always appropriate in production environments, as it carries an implicit uncertainty on what would be the resulting performance of agents after re-training.

Thus, this exemplary embodiment for the agent 100 may alleviate the problems and disadvantages of conventional solutions based on ML and NN DCNs.

FIG. 4 shows an exemplary embodiment of a flowchart performed by the agent 100 according to this disclosure. Same elements are labelled with the same reference signs.

In the exemplary embodiment of FIG. 4, the agent 100 and the neighbor agents 110 may be implemented, for example and not as a limitation, in

chips

400, 410 optimized for Artificial Intelligence (AI) applications. The

chips

400, 410 may be embedded in

switch ports

102, 112 associated with the

agents

100, 110. Each

chip

400, 410 may comprise a network interface card (NIC) .

In a first step, the agent 100 may retrieve the property of the switch port 102 associated with it from its local NIC. The property 101 of the switch port 102 may comprise any port-level metrics available at the switch port 102 that may help to optimize a global property of the network system 1, for example and not as a limitation, at least one of a transmission rate of the switch port 102 in the form of bytes transmitted by the NIC, depicted as tx _rate in FIG. 4, a queue length, q _len, of the switch port 102 and a rate of ECN marked packets, ECN _marks of the switch port 102. These properties are commonly supported by commercial switches, and can be locally obtained by the agent 100 with low computational overhead at timescales of the order of microseconds.

In a second step, the agent 100 may start a communication with the one or more neighbor agents 110 deployed in neighbor switches 112 to gain a local context. This communication is done by sending the first message 104 to the one or more neighbor agents 110.

In this embodiment, the communications may be performed, for example and not as a limitation, through a NN-driven message passing, as disclosed above for the exemplary implementation of the agent 100 by MPNN depicted in FIG. 3, where the

agents

100, 110 exchange the first message 104 and the second messages 114 directly encoded by NN modules in order to find the optimal ECN configuration 106 for the switch port 102 in its associated NIC.

This communication only requires exchanging a few bytes with neighboring switches 112 and can take a few microseconds (μs) . The base link propagation delay in production DCNs is typically of the order of 1 μs.

To set the optimal ECN configuration 106, for example the quantities k _min, k _max, p _max and/or p _mark in the case that the network system 1 is configured based on a DCQCN protocol, the agent 100 may directly interface with forwarding chips using a respective Application Programming Interface (API) , which is typically vendor-specific.

In this exemplary embodiment, the agent 100 may be further configured to mark packets according to the determined optimal ECN configuration 106 for the switch port 102.

Further, in this exemplary embodiment, the agent 100 may be configured to act as an end-host.

In a third step, a CC protocol can be executed at the end-host. The end-host may adjust the flow transmission rate of the

respective switch port

102, 112 based on ECN feedback. This process may be as follows: if a host receives an ECN-marked packet, the host may notify it to a sender by using an acknowledgement ACK.

When a sender receives the ACK, it may re-compute the flow rate of its respective switch port 102 according to a protocol-specific algorithm. The protocol-specific algorithm may comprise, for example, an Additive-Increase/Multiplicative-Decrease protocol.

This CC mechanism may enable to gradually react at one-RTT timescales that are approximately 10 μs in high-speed DCNs. The in-network optimization achieve in this disclosure is orthogonal and complementary to the selection of the flow rate control algorithm (e.g., DCTCP or DCQCN) .

Further, the agent 100 according to this disclosure is compatible with any ECN-based CC protocol and can be deployed along with any other well-established traffic optimization techniques, such as flow scheduling.

FIG. 5 shows an exemplary method 500 for an agent 100 for dynamically adapting an ECN configuration in a network system 1 according to this disclosure. The network system 1 comprises a plurality of

agents

100, 110, each

agent

100, 110 being associated with a

switch port

102, 112, as disclosed before in this disclosure.

The method 500 comprises a step 502 of sending a first message 104 to one or more neighbor agents 110. The first message 104 comprises a property 101 of the switch port 102 associated with the agent 100, and the property of the switch port 102 comprises any port-level metrics available at the switch port 102 that may help to optimize a global property of the network system 1, for example and not as a limitation, at least one of a transmission rate, a queue length, and a rate of ECN marked packets of the switch port 102.

Then, the method 500 comprises a step 304 of receiving a second message 114 from each neighbor agent 110 of the one or more neighbor agents 110. The second message 114 comprises a property 103 of a neighbor switch port 112 associated with the neighbor agent 110. The property 103 of the neighbor switch port 112 comprises any port-level metrics available at the neighbour switch port 112 that may help to optimize a global property of the network system 1, for example and not as a limitation, at least one of a transmission rate, a queue length, and a rate of ECN marked packets of the neighbor switch port 112.

The method 500 further comprises a step 306 of determining the ECN configuration 106 for the switch port 102 associated with the agent 100 based on the property 101 of the switch port 102 and the properties 103 of the neighbor switch ports 112.

The method 500 may further comprise actions according to the described aforementioned embodiments of the agent 100. Hence, the method 500 achieves the same advantages as the agent 100.

The present disclosure further provides a computer program comprising instructions that, when the program is executed by a computer, cause the computer to perform the method 500 shown in FIG. 5.

The computer program may be included in a computer readable medium. The computer readable medium may comprise essentially any memory, such as a ROM (Read-Only Memory) , a PROM (Programmable Read-Only Memory) , a 15 EPROM (Erasable PROM) , a Flash memory, an EEPROM (Electrically Erasable PROM) , or a hard disk drive.

The computer program achieves the same advantages as the method 500 and as the agent 100.

The present disclosure has been described in conjunction with various embodiments as examples as well as implementations. However, other variations can be understood and effected by those persons skilled in the art and practicing the claimed matter, from the studies of the drawings, this disclosure and the independent claims. In the claims as well as in the description the word "comprising" does not exclude other elements or steps and the indefinite article "a" or "an" does not exclude a plurality. A single element or other unit may fulfill the functions of several entities or items recited in the claims. The mere fact that certain measures are recited in the mutual different dependent claims does not indicate that a combination of these measures cannot be used in an advantageous implementation.

Claims

An agent (100) for dynamically adapting an Explicit Congestion Notification, ECN, configuration in a network system (1) , the network system (1) comprising a plurality of agents (100, 110) , each agent (100, 110) being associated with a switch port (102, 112) , and the agent (100) being configured to:

send a first message (104) to one or more neighbor agents (110) , the first message (104) comprising a property (101) of the switch port (102) associated with the agent (100) ;

receive a second message (114) from each neighbor agent (110) of the one or more neighbor agents (110) , the second message (114) comprising a property (103) of a neighbor switch port (112) associated with the neighbor agent (110) ; and

determine the ECN configuration (106) for the switch port (102) associated with the agent (100) based on the property (101) of the switch port (102) and the properties (103) of the neighbor switch ports (112) .
The agent (100) according to claim 1, wherein

the property (101) of the switch port (102) comprises at least one of: a transmission rate, a queue length, and a rate of ECN marked packets of the switch port (102) ; and/or

the property (103) of the neighbor switch port (112) comprises at least one of: a transmission rate, a queue length, and a rate of ECN marked packets of the neighbor switch port (112) .
The agent (100) according to claim 1 or 2, wherein before sending the first message (104) to one or more neighbor agents, the agent (100) is configured to:

collect the property (101) of the switch port (102) associated with the agent (100) .
The agent (100) according to one of claims 1 to 3, further configured to:

retrieve, from the second message (114) , the property (103) of the neighbor switch port (112) associated with the neighbor agent (110) .
The agent (100) according to one of the claims 1 to 4, wherein the agent (100) is further configured to:

update the property (101) of the switch port (102) associated with the agent (100) by combining the property (101) of the switch port (102) and the properties (103) of the one or more neighbor switch ports (112) retrieved from the second message (114) .
The agent (100) according to one of claims 1 to 5, further configured to:

generate an updated first message (104) , wherein the updated first message (104) comprises the updated property (101) of the switch port (102) associated with the agent (100) .
The agent (100) according to one of claims 1 to 6, wherein before the determination of the ECN configuration for the switch port (102) associated with the agent (100) based on the property (101) of the switch port (102) and the properties (103) of the one or more neighbor switch ports (112) , the agent (100) is further configured to:

perform an iteration procedure comprising performing a predefined number of iterations, wherein in each iteration the following steps are performed:

sending the updated first message (104) generated in a previous iteration to the one or more neighbor agents (110) , or in the case of the first iteration sending the first message (104) to the one or more neighbor agents (110) , wherein the updated first message (104) comprises the updated property (101) of the switch port (102) associated with the agent (100) ;

receiving an updated second message (114) from each of the one or more neighbor agents (110) generated in the previous iteration, or in the case of the first iteration receiving the second message (114) from each of the one or more neighbor agents (110) , wherein the updated second message (114) comprises an updated property (103) of the neighbor switch port (112) associated with the neighbor agent (110) ;

retrieving from the updated second message (114) the updated property (103) of the neighbor switch port (112) ;

updating the property (101) of the switch port (102) associated with the agent (100) by combining the updated property (101) of the switch port (102) calculated in the previous iteration and the updated property (103) of each neighbor switch port (112) retrieved from the one or more second messages (114) ; and

generating another updated first message (104) comprising the updated property (101) of the switch port (102) associated with the agent (100) .
The agent (100) according to one of the claims 1 to 7, wherein the determination of the ECN configuration (106) for the switch port (102) associated with the agent (100) based on the property (101) of the switch port (102) and the properties (103) of the neighbor switch ports (112) comprises:

determining an optimal ECN configuration for the switch port (102) based on the final property (101) of the switch port (102) , wherein the final property (101) of the switch port (102) comprises the updated property (101) of the switch port (102) after the last iteration; or

in the case of the first iteration, determining an optimal ECN configuration for the switch port (102) based on the property (101) of the switch port (102) and the properties (103) of the neighbor switch ports (112) received in the one or more second messages (114) .
The agent (100) according to claim 8, wherein the determined optimal ECN configuration for the switch port (102) comprises an ECN configuration (106) for the switch port (102) that optimizes the performance of the network system (1) .
The agent (100) according to claim 9, wherein the ECN configuration (106) for the switch port (102) that optimizes the performance of the network system (1) comprises at least one of: an ECN configuration (106) for the switch port (102) that decreases a flow completion time, FCT, of the network system (1) , an ECN configuration (106) for the switch port (102) that reduces a buffer occupancy of the network system (1) , and an ECN configuration (106) for the switch port (102) that increases a throughput of the network system (1) .
The agent (100) according to one of the claims 1 to 10, wherein the network system (1) is configured based on a Data Center Transmission Control Protocol, DCTCP, or a Data Center Quantized Congestion Notification, DCQCN, protocol.
A method (500) for an agent for dynamically adapting an Explicit Congestion Notification, ECN, configuration in a network system (1) , the network system (1) comprising a plurality of agents, each agent being associated with a switch port, and the method (500) comprising:

sending (502) a first message (104) to one or more neighbor agents (110) , the first message (104) comprising a property (101) of the switch port (102) associated with the agent (100) ;

receiving (504) a second message (114) from each neighbor agent (110) of the one or more neighbor agents (110) , the second message (114) comprising a property (103) of a neighbor switch port (112) associated with the neighbor agent (110) ; and

determining (506) the ECN configuration (106) for the switch port (102) associated with the agent (100) based on the property (101) of the switch port (102) and the properties (103) of the neighbor switch ports (112) .
A computer program comprising instructions which, when the program is executed by a computer, cause the computer to perform the method (500) according to claim 12.