Real-time performance adjustment of network equipment
By using packet processing circuits and performance optimization circuits in network connected devices and using genetic algorithms to adjust operating parameters in real time, the problem of unstable performance of network equipment in the prior art is solved, and the bandwidth, delay and packet loss rate are optimized. Adapt to changes in communication traffic.
Patent Information
- Application Number
- CN202111550076.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2020-12-22
- Filing Date
- 2021-12-17
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2041-12-17
AI Technical Summary
In the face of changes in communication traffic, existing network connection devices have difficulty adjusting operating parameters in real time to optimize performance, resulting in unstable performance metrics such as bandwidth, latency and packet loss rate.
Using packet processing circuits and performance optimization circuits, iteratively calculate performance metrics and adjust the value of configuration registers, and modify candidate vectors using genetic algorithms to improve the performance of network devices, including cross-section and mutation steps to select the best parameter set, and adjust the network devices in real time performance.
Real-time adjustment of network equipment performance is realized, transmission bandwidth, reception bandwidth and packet loss rate are improved, communication traffic changes are adapted to the overall performance of network equipment.
Smart Images

Figure CN114726723B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates generally to communication networks and, more particularly, to real-time performance adjustment of network connected devices. Background Art
[0002] A network connection device (NCD), such as a network interface controller, a host channel adapter, a network-enabled graphics processing unit (GPU), a network switch, or a router, typically communicates with a communication network such as Ethernet or InfiniBand™. Summary of the invention
[0003] Embodiments of the invention described herein provide a network device (ND) comprising a packet processing circuit and a performance optimization circuit. The packet processing circuit is connected to a network and is configured to process communication packets for communicating over the network. The packet processing circuit includes a plurality of configuration registers for setting one or more operating parameters of the ND. The performance optimization circuit is configured to improve a performance metric of the ND by iteratively calculating a performance metric and adjusting the value of one or more configuration registers based on the performance metric.
[0004] In some embodiments, the optimization circuit is configured to adjust the value of the one or more configuration registers in response to a parameter rule set, the parameter rule set setting one or more constraints on the one or more operating parameters. In some embodiments, the optimization circuit is configured to adjust the value of the one or more configuration registers in response to a cross-parameter rule set, the cross-parameter rule set setting one or more constraints on the interdependence between at least two of the operating parameters.
[0005] In some embodiments, the optimization circuit is configured to improve the performance metric of the ND by: (i) selecting a set of initial candidate vectors, each vector specifying a value of a corresponding subset of the configuration registers, and (ii) modifying the candidate vectors in an iterative process by adjusting the values of one or more configuration registers in response to the performance metric corresponding to the candidate vectors. In an example embodiment, the optimization circuit is configured to (i) store the value of the performance metric in a first stage of the iterative process, (ii) evaluate the degree of change of the performance metric relative to the stored value in a second later stage of the iterative process, and (iii) refresh one or more of the candidate vectors in response to the value exceeding a threshold.
[0006] In a disclosed embodiment, the optimization circuit is configured to improve the performance metric of the ND by iteratively modifying a set of candidate vectors using a genetic algorithm, each vector specifying a corresponding subset of the configuration registers, the genetic algorithm comprising: i) a parent selection step, in which a candidate vector pair is selected, ii) a crossover step, in which a sub-vector is generated by mixing the values of the candidate vector pair, iii) a mutation step, in which a mutant sub-vector is generated by modifying the sub-vector, and iv) a fittest selection step, in which a candidate vector of an iteration is selected in response to the performance metric of the mutant sub-vector of a previous iteration.
[0007] In one embodiment, the optimization circuit is configured to improve the performance metric of the ND by iteratively modifying a set of candidate vectors using a genetic algorithm, each vector specifying a corresponding subset of the configuration registers, the genetic algorithm comprising: i) a mutation step in which the candidate vectors are modified, and ii) a crossover step in which sub-vectors are generated by mixing the values of pairs of candidate vectors.
[0008] In an example embodiment, the performance metric indicates a receive bandwidth. In another embodiment, the performance metric indicates a transmit bandwidth. In yet another embodiment, the performance metric indicates a packet loss rate. In yet another embodiment, the performance metric indicates a packet pause rate. In a disclosed embodiment, the performance metric indicates a combination of latency and memory utilization. In another embodiment, the performance metric indicates a bit error rate (BER).
[0009] According to an embodiment of the present invention, there is further provided a method in a network device (ND) connected to a network. The method includes processing a communication packet for communicating through the network using a packet processing circuit in the ND. The packet processing circuit includes a plurality of configuration registers for setting one or more operating parameters of the ND. Using a performance optimization circuit in the ND, a performance metric of the ND is improved by iteratively calculating the performance metric and adjusting the value of one or more of the configuration registers based on the performance metric.
[0010] The present invention will be more fully understood through the following detailed description of its embodiments in conjunction with the accompanying drawings, in which: BRIEF DESCRIPTION OF THE DRAWINGS
[0011] Figure 1 is a block diagram schematically illustrating a network connection device (NCD) with real-time performance adjustment according to an embodiment of the present invention;
[0012] Figure 2 is a flow chart schematically describing a method for production programming and field operation of an NCD with real-time performance adjustment according to an embodiment of the present invention;
[0013] Figure 3 is a flow chart schematically describing a first method for performance optimization of NCD according to an embodiment of the present invention;
[0014] Figure 4 is a flow chart schematically describing a second method for performance optimization of NCD according to an embodiment of the present invention;
[0015] Figure 5 is a flowchart schematically describing a method for checking whether an algorithm refresh is required according to an embodiment of the present invention;
[0016] Figure 6 is a performance versus time waveform graph comparing the receive bandwidth performance of an NCD with fixed operating parameters to the performance of an NCD with real-time performance adjustment according to an embodiment of the present invention;
[0017] Figure 7 is a performance versus time waveform graph comparing the transmission bandwidth performance of an NCD with fixed operating parameters to the performance of an NCD with real-time performance adjustment according to an embodiment of the present invention; and
[0018] Figure 8 is a performance versus time waveform graph comparing the packet loss rate performance of an NCD with fixed operating parameters with the performance of an NCD with real-time performance adjustment according to an embodiment of the present invention. DETAILED DESCRIPTION
[0019] Overview
[0020] Network connection devices (NCDs; also referred to as network devices - NDs), such as network adapters, network-enabled GPUs, packet switches, and routers, typically include configuration registers that set values for various operating parameters. The parameters may include, for example, the size of various queues, the size of Tx bursts, the number of credits, etc. Various communication performance metrics of the NCD may be controlled by setting appropriate values for some or all of the operating parameters. For example, larger queues may improve (increase) the bandwidth performance metric, but may worsen (increase) the communication latency.
[0021] The optimized set of parameters can be calculated by various algorithms and trial and error. However, the optimal set of parameters may change over time as the nature of the communication traffic changes (e.g., a large number of short packets replace a smaller number of long packets).
[0022] According to an embodiment of the present invention, a method and apparatus for continuously modifying operating parameters during actual operation of an NCD in a network to adjust the performance of the NCD in real time is provided. In one embodiment, the NCD uses a genetic algorithm for performance optimization. (For a tutorial on genetic algorithms, see Whitley, Darrell (1994), "A genetic algorithm tutorial" - Statistics and Computing, Vol. 4, No. 2, pp. 65-85.)
[0023] We will refer to the combination of parameters from two parents as "crossover", "cross breeding" or simply "hybridization" hereinafter; we will further refer to the random change of certain parameters of a vector as "mutation". In some embodiments, NCD first generates a set of candidate vectors, each vector including a set of parameters, then crosses vector pairs to generate sub-vectors, mutates the sub-vectors, selects the most suitable sub-vector (e.g., the vector with the best performance, hereinafter referred to as the "best performance vector"), and then re-crosses the selected vector.
[0024] In other embodiments, no "fittest selection" step is required; instead, the NCD mutates and / or crosses vectors based on their relative performance (eg, poorer performing vectors are mutated more aggressively).
[0025] In some embodiments, if the NCD finds that performance has degraded beyond a preset threshold, the NCD can refresh the genetic algorithm by aggressively modifying the vector set.
[0026] Embodiments according to the present invention can be used to optimize various performance metrics, including (but not limited to) transmission bandwidth, reception bandwidth, delay, packet loss rate, memory utilization and power consumption. In other embodiments, combined performance metrics such as concurrent bandwidth and delay can be optimized.
[0027] System Description
[0028] Figure 1 1 is a block diagram 100 schematically illustrating a network connection device (NCD) 102 with real-time performance adjustment according to an embodiment of the present invention. The NCD 102 is coupled to a communication network 104, such as Ethernet or InfiniBand TM (or any other suitable communication network). NCD 102 may be any device coupled to a communication network, including but not limited to network adapters such as network interface controllers (NICs) and host channel adapters (HCAs), switches, routers, communication hubs, network-enabled graphics processing units (GPUs), network-enabled data processing units (DPUs), and compute nodes.
[0029] according to Figure 1In the illustrated example embodiment, NCD 102 includes a central processing unit (CPU) 106 that runs application software and transmits packets over a network via packet processing circuitry 108 that processes ingress and egress packets. The packet processing circuitry includes packet communication circuitry 110 that is configured to perform various packet processing and routing tasks, including but not limited to packet parsing, packet classification, security inspection, routing, packet queuing, and the like.
[0030] The packet processing circuit 108 also includes a set of configuration registers 112, a performance monitoring circuit 114, and a performance optimization circuit (POC) 116 (in an embodiment, the POC 116 may include a reduced instruction set computer (RISC) running firmware (FW) code). The configuration registers 112 store various operating parameters, such as queue size, packet size, thresholds for congestion control algorithms, etc. The settings of the configuration registers may affect various aspects of NCD network performance (e.g., bandwidth, latency, packet loss rate, and power consumption). Typically, parameter values follow a parameter rule set that defines the conditions that the parameter values should satisfy, such as upper and lower limits on parameter values, and a cross-parameter rule set that defines the relationship between different parameter values.
[0031] The performance monitoring circuit 114 is configured to measure various aspects of performance; for example, the performance monitoring circuit can record power consumption, communication bandwidth, packet rate, latency, cache miss rate, host fairness metric, and packet loss rate / pause rate. In some embodiments, the performance monitoring circuit includes counters and timers. In an embodiment, the performance monitoring circuit can be configured to record a combined performance function, such as bandwidth divided by power consumption, and output a predefined performance metric.
[0032] Coupled to the performance monitoring circuitry and the configuration circuitry, the POC 116 is configured to modify operating parameters stored in the configuration registers to improve various performance aspects of the NCD. In some embodiments, the POC runs a performance optimization algorithm that is configured to improve performance aspects of the NCD by reading a set of performance metrics from the performance monitoring circuitry, finding a better set of operating parameters, and programming the configuration registers with the parameter set.
[0033] In a typical communication network, the nature of the communication, such as the mix of various packet types, packet arrival frequency, transmission rate (bandwidth), burst rate, number of senders, number of receivers, and / or other network communication patterns, changes over time. Accordingly, the optimal set of parameters that the NCD should use may vary. For example, a small buffer size may be more preferred for communication of short and frequent packets. In an embodiment, the performance optimization algorithm always runs continuously, adjusting the operating parameters according to the changes in the nature of the communication.
[0034] Therefore, according to Figure 1In the example implementation shown, the NCD continuously adjusts its performance by continuously adjusting operating parameters to best adapt to the changing nature of communications.
[0035] It is understood that the structure of the above-mentioned NCD 102 is cited as an example. The NCD according to the disclosed technology is not limited to the above description. For example, in an alternative embodiment, the NCD may include multiple CPUs. In some embodiments, the CPU 106 includes a general-purpose programmable processor, which is programmed with software to perform the functions described herein. The software can be downloaded to the processor in electronic form, such as through a network, or it can be provided and / or stored in a non-temporary tangible medium, such as a magnetic, optical or electronic memory, alternatively or additionally.
[0036] Figure 2 200 is a flow chart schematically depicting a method for production programming and field operation of an NCD with real-time performance adjustment according to an embodiment of the present invention. The flow chart begins with a define relevant features step 202, in which an engineer (typically an NCD supplier engineer) defines a set of performance aspects to be optimized. Such performance aspects may include, for example, bandwidth and latency. Next, in a select relevant registers step 204, the engineer defines a set of registers that store operating parameters that may affect the desired performance aspects.
[0037] In the define objective function step 206, the engineer defines the objective function, which may be a performance metric or a combination of multiple performance metrics. Next, in the generate firmware code step 208, the engineer adds the corresponding optimization code to the firmware of the NCD.
[0038] At the customer site, once the NCD is coupled to the network, the NCD's performance optimization circuit 116 ( Figure 1 ) then continuously executes the optimization algorithm step 210. Thus, the user-defined objective function is continuously modified for optimal performance under different network conditions.
[0039] It is understood that the method of the flowchart 200 described above is cited as an example. Other flowcharts may be used in alternative embodiments. For example, the NCD supplier may send optimized firmware to the installed NCD, replace the fixed parameter algorithm or improve the real-time adjustment algorithm.
[0040] Genetic Algorithms
[0041] In some embodiments, the optimization algorithm executed by the performance optimization circuit 116 can be a genetic algorithm, in which a population of candidate solutions to an optimization problem evolves toward better solutions by crossbreeding and randomly mutating candidate solutions, and typically by selecting the most suitable mutations. For a tutorial on genetic algorithms, see Whitley (cited above).
[0042] Figure 3 300 is a flowchart schematically illustrating a first method for optimizing the performance of an NCD according to an embodiment of the present invention. The flowchart is provided by POC 116 ( Figure 1 )implement.
[0043] We will hereinafter refer to a set of values for operating parameters (stored in one or more configuration registers) as a "vector" or "candidate vector"; we further refer to a group of candidate vectors (a "population" in genetic algorithm nomenclature) as a vector set.
[0044] The flow chart begins with a generate initial population step 302, where the POC creates an initial vector set comprising P vectors. In some embodiments, the initial vector set is predefined; in other embodiments, the initial vector set can be generated by randomly changing a single predefined vector. Other suitable techniques for generating the initial vector set can be used in alternative embodiments.
[0045] After creating the initial vector set, the POC enters a crossbreeding step 304 (hereinafter also referred to as crossbreeding), where the POC generates C child vectors from the parent vector pairs (C≥P) of the vector set. Each child vector includes mixed parameters from two parent vectors. In some embodiments, the parent selection of each parameter is fixed (for example, the first n parameters are copied from the first parent, and the remaining parameters are copied from the second parent). In other embodiments, the parent selection of each parameter is random, and in other embodiments, the selection is weighted random, where the parameters of the parent with better performance are more likely to be selected. In an embodiment, step 304 includes a rule set check, which checks whether the parameter values of each generated child vector comply with the preset parameter rule set and the cross parameter rule set.
[0046] The POC then proceeds to a mutation step 306 where each sub-vector is mutated by randomly modifying some or all parameter values. In some embodiments, the strength of the parameter modification (eg, the expected ratio of the mutated parameter to the original parameter value) is determined in response to the performance of the vector.
[0047] Step 306 includes a rule set check, where each vector is checked for compliance with a parameter rule set (eg, minimum and maximum values for each parameter), and for compliance with a cross-parameter rule set. If any rule is not followed, the mutation process may be restarted.
[0048] After step 306, the POC enters a survival of the fittest step 308, where the PC vector with the lowest performance is "killed." The POC then re-enters step 304 to hybridize the remaining P vectors.
[0049] It is important to note that according to the example implementations described herein, the POC does not store the best vector in the configuration registers at any stage - instead, the POC continuously modifies the configuration registers, loads new candidate vectors and measures the resulting performance.
[0050] Figure 4 4 is a flow chart 400 schematically illustrating a second method for optimizing the performance of an NCD according to an embodiment of the present invention. The flow chart is provided by POC 116 ( Figure 1 ) is executed. Figure 4 In the method described in the example embodiment of , if the genetic algorithm generates a population with significantly poorer performance based on the performance of the previous vector set, the vector set is intensively modified. This technique is similar to thermal annealing, attempting to offset the vectors from the local maximum to which the algorithm may have converged. Flowchart 300 ( Figure 3 ) and flowchart 400 are: i) mutation of the parent vector (flowchart 400) versus mutation of the child vector (flowchart 300), and ii) there is no survival of the fittest step in flowchart 400 - the population remains unchanged after the crossover step.
[0051] Flowchart 400 begins with generating an initial vector set step 402, where the POC creates an initial vector set including P vectors. Figure 3 ), the initial vector set may be generated by a variety of procedures, including but not limited to variations of known good vectors, vectors calculated by static algorithms, etc.
[0052] The POC then enters the first check refresh step 404 and checks whether the performance of the current vector set is significantly worse than the performance of the previous vector set (e.g., less than a preset amount), and therefore, an algorithm refresh is required (see below). Figure 5 406. In step 406, the POC mutates the vectors of the vector set (e.g., 20%) by modifying the parameters specified by the vectors. Figure 4 In the example implementation shown, the strength of variation is determined in response to the performance of the vector; for example, in one implementation, the distance (e.g., the difference) between the performance of the vector and the performance of the best performing vector in a set of vectors is calculated, and each parameter of the vector is multiplied or divided (randomly selected) by a factor determined in response to the distance.
[0053] The mutated vectors generated in step 406 comprise a new vector set. In a second check refresh step 408, the POC checks the performance of the new set to determine if a refresh is required. If a refresh is not required in step 408, the POC proceeds to a hybridization step 410 and generates a new P vector set by hybridizing vector pairs from the current vector set (hybridization reference Figure 3 After step 410, the POC re-enters step 404 and starts a new iteration.
[0054] If, in step 404 or step 408, the POC determines that a refresh is needed, the POC enters a refresh step 412, in which the POC introduces intensive changes to the parameter values of the vector set. The POC then re-enters the mutation step 406. In each of steps 406, 410, and 412, as the POC generates new vectors, the POC checks the new vectors against the parameter rule set and the cross-parameter rule set, and regenerates vectors that do not comply with any or both rule sets.
[0055] In short, according to Figure 4 In the example method described in , POC continuously modifies the operating parameters of NCD by mutating candidate vectors and hybridizing mutated vectors to form new candidate vectors. Each generated vector is activated (i.e., the parameters specified by the vector are stored in the configuration registers of NCD) and the respective performance is measured, affecting further mutation and / or hybridization. If this process causes the performance to degrade by more than a preset threshold (relative to past performance), POC refreshes the algorithm by intensively modifying the set of vectors.
[0056] Figure 5 500 is a flowchart schematically illustrating a method for checking whether an algorithm refresh is required according to an embodiment of the present invention. The flowchart is provided by POC 116 ( Figure 1 ) is executed and can be called during the execution of the genetic algorithm (e.g., Figure 4 Steps 404, 408).
[0057] The flowchart begins at a load register step 502, where the POC loads the parameter values specified by the first vector into the corresponding configuration registers. Next, the POC enters a measure and record step 504, where the POC reads the relevant performance metrics of the vector (e.g., from the performance monitoring circuit 112, Figure 1 ), and record the performance metric. The POC then reenters step 502 and repeats the loop including steps 502 and 504 for all vectors in the vector set.
[0058] Next, in the find best performance step 506, the POC examines the recorded performance metrics and finds the best performance achieved by the current vector set. Finally, in the check performance degradation step 508, the POC checks whether the performance degradation relative to the previous performance meets a certain condition, for example, greater than a preset threshold. If so, the POC determines that the algorithm should be refreshed; otherwise, the algorithm can continue (e.g., enter the hybridization step 410 or the mutation step 404, Figure 4 ).
[0059] It should be understood that the above Figure 3 , Figure 4 and Figure 5 The methods described in are example methods cited as examples. Methods according to the disclosed technology are not limited to the description above. For example, in an alternative embodiment, when the parameters specified by the mutation or hybridization vector are loaded into the configuration register, if a severe performance degradation is observed, the vector can be discarded from the vector set. In some embodiments, the strength of the mutation is determined by a combination of absolute and relative performance metrics. In one embodiment, multiple copies of the algorithm can be run simultaneously, for example, for orthogonal performance metrics.
[0060] Examples of relevant performance aspects
[0061] Embodiments according to the present invention can be used to optimize various NCD performance aspects. The following is a partial list of examples.
[0062] 1. Receive bandwidth. Receive bandwidth can be varied by manipulating parameters such as buffer size, coalescing width, and "in-flight byte limit" (the number of bytes the NCD keeps in memory when sending a read request, to store expected completion and return data, i.e., the number of bytes allowed to be "in-flight" on the way to the NCD). In an embodiment, the POC can adjust operating parameters in real time to optimize receive bandwidth in response to a changing traffic mix.
[0063] Figure 6 is a performance versus time waveform graph comparing the receive bandwidth performance of an NCD with fixed operating parameters to the performance of an NCD with real-time performance adjustment according to an embodiment of the present invention. The vertical axis is the receive bandwidth in units of gigabits per second (Gbps). The horizontal axis is time in units of seconds. The dashed line shows the performance using real-time parameter adjustment according to an embodiment of the present invention, while the solid line depicts the performance when an optimized static parameter set is used. Seven different communication combinations were exercised. It can be seen that the performance when using real-time parameter adjustment is almost always better than the fixed parameter performance.
[0064] 2. Transmission bandwidth. Transmission bandwidth can be controlled by manipulating operating parameters such as packet length, buffer size, coalescing width, and "occupied bytes in the air" limit. In an embodiment, POC adjusts transmission bandwidth by adjusting relevant operating parameters in real time, adopting a changing communication mix.
[0065] Figure 7 1 is a performance versus time waveform diagram comparing the transmission bandwidth performance of an NCD with fixed operating parameters and the performance of an NCD with real-time performance adjustment according to an embodiment of the present invention. It can be observed that the real-time adjustment performance is significantly better than the fixed parameter performance.
[0066] 3. Packet loss rate (in lossy networks). Packet loss rate in lossy networks can be reduced by manipulating operational parameters such as buffer size, coalescing width, and "bytes in the air" limit. In an embodiment, the parameters are adjusted in real time to minimize packet loss as the nature of the communication changes. Figure 8 FIG. 1 is a performance versus time waveform diagram comparing the packet loss rate performance of an NCD with fixed operating parameters and the performance of an NCD with real-time performance adjustment according to an embodiment of the present invention. The vertical axis is the packet loss rate in packets per second. It can be seen that when the real-time parameter adjustment method according to an embodiment of the present invention is used, the packet loss rate is significantly reduced.
[0067] 4. Pause Rate (Lossless Network) - In an embodiment, the POC adjusts operating parameters such as buffer size, coalescing width, and air occupancy byte limit in real time to minimize operational pauses.
[0068] 5. Combined Latency / Memory Utilization - In one embodiment, the POC adjusts the rebalancing algorithm parameters, such as the algorithm comparator type and the number of idle lines, in real time to achieve minimum latency and maximum physical memory utilization.
[0069] 6. Bit Error Rate (BER) Performance - Real-time adjustment of BER reduction algorithm parameters, such as feed-forward equalization weights and continuous-time linear equalization parameters in optical communication devices, to achieve minimum BER.
[0070] 7. Real-time adjustment of parameters of the Data Center Quantized Congestion Notification (DCQCN) congestion control algorithm, such as the minimum time between congested packets and the reaction point threshold, to maximize bandwidth and minimize latency, packet loss rate, and packet pause rate (e.g., the DCQCN algorithm is described in Zhu et al., "Congestion Control for Large-Scale RDMA Deployments," SIGCOMM'15, August 17-21, 2015).
[0071] 8. Combined Performance Metrics - To avoid improving certain performance metrics at the expense of other performance metrics, combined performance metrics can be defined and optimized. For example, power consumption should be minimized, but bandwidth cost should not be high. In one embodiment, the performance metric can be bandwidth / power consumption (e.g., in Gbps / watt). In other embodiments, more complex performance metrics can be used that combine multiple performance aspects.
[0072] The configurations and methods disclosed above, including NCD 102, POC 116; flowchart 200, flowchart 300, flowchart 400, and flowchart 500 are example configurations and methods shown purely for conceptual clarity. Any other suitable configurations and flowcharts may be used in alternative embodiments.
[0073] Elements of the POC 116 and / or NCD 102 may be implemented using suitable hardware, such as software in one or more application specific integrated circuits (ASICs) or field programmable gate arrays (FPGAs), hardware, or a combination of hardware and software elements. In some embodiments, the POC 116 includes a general purpose programmable processor that is programmed with software to perform the functions described herein. The software may be downloaded to the processor in electronic form, such as over a network, or it may alternatively or additionally be provided and / or stored on a non-transitory tangible medium, such as a magnetic, optical, or electronic memory.
[0074] It should be understood that the above-mentioned embodiments are cited by way of example, and the present invention is not limited to what has been specifically shown and described above. On the contrary, the scope of the present invention includes the combination and sub-combination of the various features described above, as well as changes and modifications that those skilled in the art would think of after reading the foregoing description and are not disclosed in the prior art. The documents incorporated by reference in this patent application should be considered as an integral part of this application, unless any term defined in these incorporated documents in a manner that conflicts with the explicit or implicit definition in this specification, only the definition in this specification should be considered.
Claims
1. A network device (ND) include: a packet processing circuit connected to a network and configured to process communication packets for communicating over the network, the packet processing circuit including a plurality of configuration registers for setting one or more operating parameters of the ND; and A performance optimization circuit is configured to iteratively perform the following steps: calculating a performance metric of current values of the plurality of configuration registers; as well as Values of one or more configuration registers are adjusted to values different than current values, wherein the adjustment is performed regardless of the calculated performance metric of the current values of the plurality of configuration registers.
2. The ND of claim 1, wherein the optimization circuit is configured to adjust the values of the one or more configuration registers in response to a parameter rule set, the parameter rule set placing one or more constraints on the one or more operating parameters.
3. The ND of claim 1 , wherein the optimization circuit is configured to adjust the values of the one or more configuration registers in response to a cross-parameter rule set, the cross-parameter rule set setting one or more constraints on interdependencies between at least two of the operating parameters.
4. The ND of claim 1 , wherein the optimization circuit is configured to manage a set of a plurality of candidate vectors, each vector specifying a different value of the configuration register; and in each iteration to: calculating a performance metric for each candidate vector; and The set of multiple candidate vectors is modified.
5. The ND according to claim 4, wherein the optimization circuit is configured to: storing the value of the performance metric during a first phase of an iterative process; evaluating, at a second later stage of the iterative process, the extent to which the performance indicator has changed relative to the stored value; and One or more of the candidate vectors are refreshed in response to the value exceeding a threshold. 6 . The ND according to claim 5 , wherein the optimization circuit is configured to generate a new candidate vector by mixing values of candidate vector pairs of non-discarded candidate vectors to replace the discarded candidate vector. The ND of claim 1 , wherein the performance metric indicates a reception bandwidth. The ND of claim 1 , wherein the performance metric indicates a transmission bandwidth.
9. The ND of claim 1, wherein the performance metric indicates a packet loss rate.
10. The ND of claim 1, wherein the performance metric indicates a packet pause rate.
11. The ND of claim 1, wherein the performance metric indicates a combination of latency and memory utilization.
12. The ND of claim 1, wherein the performance metric indicates a bit error rate (BER).
13. A method in a network device (ND) connected to a network, the method include: processing, using packet processing circuitry in the ND, communication packets for communicating over the network, wherein the packet processing circuitry includes a plurality of configuration registers for setting one or more operating parameters of the ND; as well as Using the performance optimization circuit in the ND, the performance metric of the ND is improved by iteratively performing the following steps: calculating a performance metric of current values of the plurality of configuration registers; as well as Values of one or more configuration registers are adjusted to values different than current values, wherein the adjustment is performed regardless of the calculated performance metric of the current values of the plurality of configuration registers.
14. The method of claim 13, wherein adjusting the value of the one or more configuration registers is performed in response to a parameter rule set, the parameter rule set placing one or more constraints on the one or more operating parameters.
15. The method of claim 13, wherein adjusting the value of the one or more configuration registers is performed in response to a cross-parameter rule set, the cross-parameter rule set placing one or more constraints on interdependencies between at least two of the operating parameters.
16. The method of claim 13, further comprising managing a set of a plurality of candidate vectors, each vector specifying a different value for the configuration register; and In each iteration the following steps are performed: calculating a performance metric for each candidate vector; and The set of multiple candidate vectors is modified.
17. The method of claim 16, wherein improving the performance metric of the ND include: storing the value of the performance metric during a first phase of an iterative process; evaluating, in a second later stage of the iterative process, the extent to which the performance indicator has changed relative to the stored value; as well as One or more of the candidate vectors are refreshed in response to the value exceeding a threshold.
18. The method of claim 13, wherein the candidate vectors are generated by mixing values of non-discarded candidate vector pairs.
