Scheduling method and device
Through sensitivity analysis and affinity strategy based on memory fetch indicators, binding business modules and NUMA or NUCA resources is solved, and the shortcomings of existing scheduling solutions in comprehensively considering the advantages and disadvantages of different architectures are improved, and system performance and resource utilization efficiency are improved.
Patent Information
- Application Number
- CN202411186956.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-11-08
AI Technical Summary
The existing NUMA and NUCA architecture scheduling solutions have advantages and disadvantages in specific application scenarios, and it is difficult to comprehensively consider the advantages and disadvantages of different solutions, resulting in low system performance and resource utilization efficiency.
By determining the sensitivity of NUMA and NUCA based on the memory fetch indicator, dividing sensitive categories, and implementing affinity strategy based on the sensitive categories, binding the business module with the NUMA node or NUCA core, and scheduling to the bound node or core for execution , realize efficient utilization of resources.
It improves system performance and resource utilization efficiency, has a wide range of applications, and can transparently utilize hardware resources without changes in business modules. Dynamic optimization strategies ensure continuous positive returns and reduce memory access latency and cache access latency.
Smart Images

Figure CN118916145A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of computer technology, and in particular to the field of data processing technology. Background Art
[0002] In modern high-performance computing systems, NUMA (Non-Uniform Memory Access) and NUCA (Non-Uniform Cache Access) architectures have been widely used to improve the overall system performance in a multi-processor environment.
[0003] The existing NUMA architecture scheduling solutions mainly include the following three:
[0004] First, task affinity scheduling: Based on the affinity between tasks and memory nodes, tasks are scheduled to processors or memory nodes that are closer to their data to reduce the number of remote memory accesses.
[0005] Second, load balancing scheduling: by monitoring the load of each node, tasks are evenly distributed to each node to avoid overloading of some nodes.
[0006] Third, topology-aware scheduling: Use system topology information to make scheduling decisions, considering the topological relationship between processors, memory nodes, and I / O (Input / Output) devices.
[0007] The existing NUCA architecture scheduling solutions mainly include the following three:
[0008] First, static data distribution: at compile time or system startup, data is statically allocated to specific cache modules to reduce cache access latency.
[0009] Second, dynamic data migration: dynamically migrate data to a more appropriate cache location based on runtime access patterns and cache access latency.
[0010] Third, access pattern prediction: By monitoring and predicting data access patterns, data can be migrated to the expected cache location in advance to reduce cache access latency. Summary of the invention
[0011] The embodiments of the present disclosure provide a scheduling method, apparatus, device, storage medium, and program product.
[0012] In a first aspect, an embodiment of the present disclosure proposes a scheduling method, including: determining the NUMA sensitivity of non-uniform memory access and the NUCA sensitivity of non-uniform cache access based on memory access indicators; determining the sensitivity category based on the NUMA sensitivity and the NUCA sensitivity; binding the business module to the NUMA node or the core of the NUCA based on the affinity policy corresponding to the sensitive category; and scheduling the business module to the bound NUMA node or the core of the NUCA for execution.
[0013] In the second aspect, an embodiment of the present disclosure proposes a scheduling device, including: a first determination module, configured to determine the NUMA sensitivity of non-uniform memory access and the NUCA sensitivity of non-uniform cache access based on memory access indicators; a second determination module, configured to determine the sensitivity category based on the NUMA sensitivity and the NUCA sensitivity; a binding module, configured to bind the business module to the NUMA node or the core of the NUCA based on the affinity policy corresponding to the sensitivity category; and a scheduling module, configured to schedule the business module to the bound NUMA node or the core of the NUCA for execution.
[0014] In a third aspect, an embodiment of the present disclosure proposes an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method described in the first aspect.
[0015] In a fourth aspect, an embodiment of the present disclosure proposes a non-transitory computer-readable storage medium storing computer instructions, where the computer instructions are used to cause a computer to execute the method described in the first aspect.
[0016] In a fifth aspect, an embodiment of the present disclosure proposes a computer program product, including a computer program, which implements the method described in the first aspect when executed by a processor.
[0017] The key or important features of the embodiments of the present disclosure are not intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Other features, purposes and advantages of the present disclosure will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings. The drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure. Among them:
[0019] Figure 1 It is a schematic diagram of the structure of a server with NUMA characteristics;
[0020] Figure 2This is a schematic diagram of the structure of a shared LLC;
[0021] Figure 3 is an overall architecture diagram of the scheduling method according to the present disclosure;
[0022] Figure 4 is a flow chart of an embodiment of a scheduling method according to the present disclosure;
[0023] Figure 5 is a flow chart of another embodiment of the scheduling method according to the present disclosure;
[0024] Figure 6 is a flow chart of an embodiment of a scheduling optimization method according to the present disclosure;
[0025] Figure 7 is a flow chart of another embodiment of the scheduling optimization method according to the present disclosure;
[0026] Figure 8 is a structural schematic diagram of an embodiment of a scheduling device according to the present disclosure;
[0027] Fig. 9 It is a block diagram of an electronic device used to implement the scheduling method of the embodiment of the present disclosure. DETAILED DESCRIPTION
[0028] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.
[0029] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present disclosure may be combined with each other. The present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0030] Figure 1 A schematic diagram of the structure of a server with NUMA characteristics is shown.
[0031] like Figure 1 As shown in the figure, the CPU (Central Processing Unit) can access local memory through the memory controller, and can also access the memory of other CPUs through the interconnect bus, that is, remote memory. In addition, local memory access has low latency and high bandwidth, while remote memory access has high latency and low bandwidth, showing the characteristics of non-uniform memory access.
[0032] The NUMA architecture reduces memory access latency by dividing the memory into multiple nodes, each of which consists of one or more processors and local memory. However, due to the non-uniformity of memory access, task scheduling becomes more complicated. How to effectively schedule tasks to minimize remote memory access and optimize overall performance has become an important research topic.
[0033] Figure 2 A schematic diagram of the structure of a shared LLC (Last Level Cache) is shown.
[0034] like Figure 2 As shown, the shared LLC structure mainly includes UCA (Uniform Cache Access) and NUCA.
[0035] UCA is a centralized shared structure. Multiple processor cores are connected to LLC through a bus or a cross switch, and all processor cores have the same access delay to LLC. This centralized shared LLC can easily become a bottleneck as the number of processor cores increases. In addition, the UCA structure has limited scalability due to the use of buses or cross switches for interconnection. Therefore, the UCA structure is usually used in multi-core processors with a small number of processor cores.
[0036] NUCA is a distributed shared structure. Each processor core has a local LLC and accesses the LLC of other processor cores through on-chip interconnection. In the NUCA structure, the processor core can access all LLCs, but LLCs in different locations have different access delays. When the working set is small, the local cache of the processor core is sufficient to accommodate the working set, and the processor core only uses the local cache; when the working set is large, data that cannot be placed in the local cache can be placed in the remote cache. The NUCA structure requires an efficient cache search and replacement algorithm so that performance is not affected when using remote cache. The NUCA structure usually uses scalable on-chip interconnection (such as wireless mesh networks, etc.) and a directory-based cache consistency protocol. It has good scalability and can effectively support a large number of processor cores. Therefore, the NUCA structure is usually used in multi-core or many-core processors with a large number of cores.
[0037] The NUCA architecture introduces multi-level caches and allows for non-uniform cache access latency to better adapt to the locality requirements of the processor. However, the complexity of the NUCA architecture also brings new challenges, especially in the allocation and scheduling of cache resources.
[0038] By analyzing the existing NUMA and NUCA architecture scheduling solutions, it can be seen that each solution has advantages and disadvantages in specific application scenarios. Designing a scheduling strategy that comprehensively considers the advantages and disadvantages of different solutions can help further improve system performance and resource utilization efficiency.
[0039] Figure 3 The overall architecture diagram of the scheduling method according to the present disclosure is shown.
[0040] The overall architecture of the scheduling method disclosed in the present invention can be divided into an offline part and an online part. The offline part mainly performs NUMA and NUCA sensitivity profiling. The online part mainly performs resource scheduling.
[0041] The online part may include a single machine layer and a cluster layer. The single machine layer may include agents, monitors, distributors, and optimizers, etc. The cluster layer may include a scheduler.
[0042] Monitors can collect resource usage information of containers, nodes, machines, etc. and send this information to agents, offline parts, and optimizers.
[0043] The agent is the main interface of the single-machine layer and the optimization control end of NUMA and NUCA. The agent can be responsible for communicating with the scheduler of the cluster layer and controlling the allocation of resources of the single-machine layer according to the deployment requirements of the scheduler and the control parameters of NUMA and NUCA obtained by the offline part.
[0044] The allocator can be the execution end that ultimately implements NUMA or NUCA binding, and can bind processes to NUMA nodes or NUCA cores as needed.
[0045] The optimizer can be responsible for the runtime optimization of NUMA or NUCA. Based on the information collected by the monitor and the control parameters of NUMA or NUCA obtained in the offline part, it determines whether the module optimized by NUMA or NUCA has achieved the expected goal. If not, the corresponding optimization operation will be performed.
[0046] The cluster layer can be controlled by a scheduler. The scheduler can receive the parameters output by the offline part after NUMA and NUCA sensitivity modeling, make NUMA or NUCA optimization decisions for the module based on the parameters, and send them to the agent at the stand-alone layer.
[0047] The offline part can receive data collected by the monitors of all machines and perform NUMA and NUCA sensitivity modeling. The model is used to predict modules that will benefit from NUMA or NUCA optimization, and the results are output to the cluster-level scheduler. The offline part can perform dynamic binding or unbinding condition judgment, and output parameters to the single-machine-level agent for dynamic NUMA or NUCA binding control.
[0048] Figure 4 A process 400 of an embodiment of a scheduling method according to the present disclosure is shown. The scheduling method comprises the following steps:
[0049] Step 401, based on the memory access index, determine the non-uniform memory access NUMA sensitivity and the non-uniform cache access NUCA sensitivity.
[0050] In this embodiment, NUMA sensitivity and NUCA sensitivity may be determined based on the memory access index.
[0051] The offline part can profile NUMA and NUCA sensitivity.
[0052] In some embodiments, NUMA latency sensitivity can be calculated based on the memory access metric. Based on the NUMA latency sensitivity, NUMA sensitivity can be determined. The greater the NUMA latency sensitivity, the greater the NUMA sensitivity. For example, NUMA latency sensitivity can be directly used as NUMA sensitivity.
[0053] Memory access indicators may include the time to wait for memory access, the time to wait for cache access, the amount of data in the RSS (Resident Set Size) memory page, the amount of data in the cache memory page, and the memory access bandwidth or the number of memory accesses. Based on these memory access indicators, the NUMA latency sensitivity can be calculated. The calculation formula may be as follows:
[0054] NUMA delay sensitivity value = memory access bandwidth × [waiting time for memory access / (waiting time for memory access + waiting time for cache access)] × [data volume of RSS memory page / (data volume of RSS memory page + data volume of cache memory page)].
[0055] Alternatively, NUMA delay sensitivity value = number of memory accesses × [waiting time for accessing memory / (waiting time for accessing memory + waiting time for accessing cache)] × [data volume of RSS memory page / (data volume of RSS memory page + data volume of cache memory page)].
[0056] It is understandable that the larger the NUMA delay sensitivity value is, the more sensitive the service module is to the memory access delay. By appropriately reducing the memory access delay of these service modules, the performance of these service modules can be effectively improved.
[0057] In some embodiments, NUCA delay sensitivity, NUCA capacity sensitivity, and NUCA shared sensitivity may be calculated based on the memory access index. NUCA sensitivity may be determined based on NUCA delay sensitivity, NUCA capacity sensitivity, and NUCA shared sensitivity. For example, NUCA delay sensitivity, NUCA capacity sensitivity, and NUCA shared sensitivity are weighted and summed to obtain NUCA sensitivity.
[0058] Memory access indicators may include LLC delay sensitivity coefficient, LLC usage time ratio, LLC capacity sensitivity coefficient, business working set size, local LLC capacity, LLC shared data sensitivity coefficient, business shared data size, and number of concurrent threads for shared data. Based on the LLC delay sensitivity coefficient and LLC usage time ratio, NUCA delay sensitivity can be calculated. Based on the LLC capacity sensitivity coefficient, business working set size, and local LLC capacity, NUCA capacity sensitivity can be calculated. Based on the LLC shared data sensitivity coefficient, business shared data size, and number of concurrent threads for shared data, NUCA shared sensitivity can be calculated. The calculation formula may be as follows:
[0059] Computation time = CPU time + memory access time (Formula 1);
[0060] Memory access time = cache access time + memory access time (Formula 2);
[0061] Access cache time = access non LLC time + access LLC time (Formula 3);
[0062] NUCA delay sensitivity = (LLC delay sensitivity coefficient × LLC usage time ratio) (Formula 4);
[0063] NUCA capacity sensitivity = LLC capacity sensitivity coefficient × (business working set size - local LLC capacity) (Formula 5);
[0064] NUCA sharing sensitivity = LLC shared data sensitivity coefficient × business shared data size × number of concurrent threads sharing data (Formula 6).
[0065] It can be understood that the execution time of the business module can be the computing time required for the relevant thread to process a specific task. As shown in Formula 1, the computing time required for the thread to process a specific task can be divided into CPU time and memory access time. In order to reduce the execution time of the business module, it is necessary to reduce the memory access delay. Therefore, NUCA sensitivity identification can estimate the potential benefits of the binding operation by analyzing performance indicators related to memory access. Because the memory access time is composed of the access cache time and the access memory time, and the access cache time includes the access non-LLC time and the access LLC time, as shown in Formula 2 and Formula 3. NUCA delay sensitivity is related to the delay of LLC, as shown in Formula 4, and is related to the proportion of LLC usage time and the LLC delay sensitivity coefficient. NUCA capacity sensitivity is related to the capacity of LLC, as shown in Formula 5, the larger the business working set, the more accesses will access the remote LLC. NUCA sharing sensitivity is related to the shared programming mode of LLC, as shown in Formula 6, which is proportional to the shared data size and the number of concurrent threads sharing data.
[0066] It should be noted that the parameters of the above formula can be trained and output by machine learning of multiple rounds of actual measurements or batch data or by means of a large model.
[0067] Step 402, determining a sensitivity category based on NUMA sensitivity and NUCA sensitivity.
[0068] In this embodiment, the sensitivity category may be determined based on NUMA sensitivity and NUCA sensitivity.
[0069] Different NUMA sensitivities and different NUCA sensitivities may correspond to different sensitivity categories. For example, if the sensitivity of NUMA is greater than the second preset sensitivity threshold, NUMA is sensitive; otherwise, NUMA is insensitive. If the sensitivity of NUCA is greater than the first preset sensitivity threshold, NUCA is sensitive; otherwise, NUCA is insensitive. If both NUMA and NUCA are sensitive, it means that both memory and cache are sensitive, and it belongs to the category of both cache and memory sensitivity. If NUMA is insensitive, but NUCA is sensitive, it means that only cache is sensitive, and it belongs to the category of cache sensitivity. If NUMA is sensitive, but NUCA is insensitive, it means that only memory is sensitive, and it belongs to the category of memory sensitivity. If both NUMA and NUCA are not sensitive, it means that both memory and cache are not sensitive, and it belongs to the category of insensitive.
[0070] Step 403: Bind the service module to the NUMA node or the core of the NUCA based on the affinity policy corresponding to the sensitive category.
[0071] In this embodiment, based on the affinity policy corresponding to the sensitive category, the service module can be bound to the NUMA node or the core of the NUCA.
[0072] Different sensitivity categories may correspond to different affinity policies. For example, for both cache and memory sensitive categories and cache sensitive categories, the affinity policy may be to bind the business module to the NUCA core. For the memory sensitive category, the affinity policy may be to bind the business module to the NUMA node. For the insensitive category, the affinity policy may be to not perform any binding.
[0073] Step 404: Schedule the service module to the bound NUMA node or NUCA core for execution.
[0074] In this embodiment, if there is a bound NUMA node or NUCA core, the service module can be scheduled to the bound NUMA node or NUCA core for execution. If there is no binding, the service module can be scheduled to any NUMA node or NUCA core for execution.
[0075] The disclosed embodiment provides a scheduling method, which allows the service to transparently and efficiently utilize the hardware of the NUMA or NUCA architecture by binding the service module with the hardware resources. In addition, through the close integration of the single machine layer and the cluster layer, the service module can transparently utilize the hardware resources without modification, and has a wide range of applications.
[0076] Figure 5 A process 500 of another embodiment of a scheduling method according to the present disclosure is shown. The scheduling method comprises the following steps:
[0077] Step 501, based on the memory access index, determine the non-uniform memory access NUMA sensitivity and the non-uniform cache access NUCA sensitivity.
[0078] Step 502, determining a sensitivity category based on NUMA sensitivity and NUCA sensitivity.
[0079] In this embodiment, the specific operations of steps 501-502 are already described in Figure 4 In the illustrated embodiment, steps 401 - 402 are described in detail and will not be repeated here.
[0080] Step 503: If the sensitivity of the NUCA is greater than the first preset sensitivity threshold, the service module is bound to the core of the NUCA according to the resource usage.
[0081] In this embodiment, if the sensitivity of the NUCA is greater than the first preset sensitivity threshold, the service module may be bound to the core of the NUCA according to resource usage.
[0082] In some embodiments, core binding requires calculating the number of bound cores, then selecting a cache set corresponding to the number of bound cores from a shared cache set, and binding the service module to the cores corresponding to the cache set.
[0083] Based on NUCA sharing sensitivity, resource usage, and utilization threshold, the number of bound cores can be calculated. The calculation formula can be as follows:
[0084] Number of bound cores = NUCA sharing sensitivity × (resource usage / utilization threshold).
[0085] For example, if the resource usage of a business module is 800% (i.e., 8 logical cores), the utilization threshold is 70% (the utilization thresholds of different business modules are different and can be obtained based on actual deployment experience), and the NUCA sharing sensitivity is 1.5, then the number of bound cores is 17.14, which is rounded to 17 cores.
[0086] When selecting a cache set corresponding to the number of bound cores, the core resources corresponding to the cache set with the smallest delay can be selected from the shared cache set. Specifically, according to the cache test program (such as multichase's pingpong), the affinity relationship between the cores can be obtained. Scheduling is performed based on the core's delay topology and current load relationship, and a suitable set of cores is selected for binding. For example, 4 cores need to be bound, but there are two alternative sets of affinity cores, namely, {0, 1, 2, 3} and {24, 25, 26, 27}. Two groups of combinations with low current utilization and not exceeding the binding threshold can be selected for load balancing.
[0087] Step 504: If the sensitivity of NUCA is not greater than the first preset sensitivity threshold, and the sensitivity of NUMA is greater than the second preset sensitivity threshold, the service module is bound to the NUMA node according to the memory access bandwidth and / or the amount of occupied memory.
[0088] In this embodiment, if the sensitivity of NUCA is not greater than the first preset sensitivity threshold, and the sensitivity of NUMA is greater than the second preset sensitivity threshold, the business module can be bound to the NUMA node based on the memory access bandwidth; or the business module can be bound to the NUMA node based on the occupied memory amount; or the business module can be bound to the NUMA node based on the memory access bandwidth and the occupied memory amount.
[0089] In order to ensure that the memory access latency of the business module is as small as possible, the amount of memory occupied by the business module in each NUMA node can be determined. Combined with the memory access bandwidth of each NUMA node, the NUMA node that needs to be bound to the business module is determined. For example, the NUMA node with the largest amount of memory occupied by the business module is bound to the business module. Alternatively, the NUMA node with the largest memory access bandwidth is bound to the business module. Alternatively, the memory access bandwidth and the amount of memory occupied are comprehensively considered, and the weighted sum of the two is taken, and the NUMA node that needs to be bound to the business module is selected based on the weighted result.
[0090] Step 505: If the sensitivity of NUMA is not greater than the second preset sensitivity threshold, and the sensitivity of NUCA is not greater than the first preset sensitivity threshold, the service module is not bound to the NUMA node or the core of NUCA.
[0091] In this embodiment, if the sensitivity of NUMA is not greater than the second preset sensitivity threshold, and the sensitivity of NUCA is not greater than the first preset sensitivity threshold, the service module is not bound to the NUMA node or the core of NUCA.
[0092] Step 506: Schedule the service module to the bound NUMA node or NUCA core for execution.
[0093] In this embodiment, when there is a bound NUMA node or a core of a NUCA, the service module can be scheduled to the bound NUMA node or the core of the NUCA for execution.
[0094] Step 507: Schedule the service module to an unbound NUMA node or a core of a NUCA for execution.
[0095] In this embodiment, when there is no binding, the service module can be scheduled to any NUMA node or NUCA core for execution.
[0096] The disclosed embodiment provides a scheduling method, which allows the service to transparently and efficiently utilize the hardware of the NUMA or NUCA architecture by binding the service module with the hardware resources. In addition, through the close integration of the single machine layer and the cluster layer, the service module can transparently utilize the hardware resources without modification, and has a wide range of applications.
[0097] After the NUMA or NUCA affinity strategy is implemented, the business modules may not be able to guarantee a globally consistent local memory access ratio, and performance fluctuations may occur. In addition, if the online business experiences a surge in traffic (such as stress testing, disaster recovery, special days or events, etc.), at this time, due to the increase in required CPU resources, a single node may not be able to provide sufficient computing power, that is, resource saturation. Therefore, after the scheduling is completed, local dynamic tuning can also be performed. Specifically, single-machine layer optimization is the main focus, and cluster layer optimization is supplemented.
[0098] Figure 6 A process 600 of an embodiment of a scheduling optimization method according to the present disclosure is shown. The scheduling optimization method comprises the following steps:
[0099] Step 601, obtaining monitoring data of core indicators of the business module.
[0100] In this embodiment, monitoring data of core indicators of the business module are obtained, wherein the core indicators may include but are not limited to: IPC (InterProcess Communication), local memory access ratio, cache access ratio, module utilization, whole machine load, etc.
[0101] Step 602 , comparing the monitoring data with the expected threshold of the core indicator according to the time window to determine whether resource saturation occurs.
[0102] In this embodiment, monitoring data is compared with expected thresholds of core indicators according to time windows to determine whether resource saturation occurs.
[0103] For core indicators, you can establish expected thresholds. Compare the monitoring data with the expected thresholds of the core indicators according to the time window to determine whether there are resource and performance issues during operation. If there are performance issues, resource saturation may occur. For example, if the monitoring data exceeds the expected threshold, it can be considered that resource saturation occurs. If the monitoring data does not exceed the expected threshold, it can be considered that resource saturation does not occur.
[0104] Step 603: In response to resource saturation, scheduling optimization is performed on the service module according to the resource saturation type.
[0105] In this embodiment, when resource saturation occurs, scheduling optimization can be performed on the service modules according to the resource saturation type.
[0106] When determining resource saturation, it can be divided into three dimensions: the entire machine, NUMA resources, and NUCA resources. Different types of resource saturation can correspond to different scheduling optimization strategies. For example, for the saturation of the entire machine resources, the scheduling optimization strategy can be to migrate at least part of the business modules to other machines. For NUMA resource saturation, the scheduling optimization strategy can be to unbind the bound NUMA nodes and schedule the business modules to the unbound NUMA nodes for execution. For NUCA resource saturation, the NUCA cores can be dynamically unbound or bound, and the business modules can be scheduled to the corresponding NUCA cores for execution.
[0107] The disclosed embodiment provides a scheduling optimization method, which allows the business to transparently and efficiently utilize the hardware of the NUMA or NUCA architecture by binding the business module with the hardware resources. In addition, through the close integration of the single-machine layer and the cluster layer, the business module can transparently utilize the hardware resources without modification, and has a wide range of applications. In addition, the dynamic optimization strategy can be dynamically optimized in combination with the runtime characteristics of the business module to ensure continuous positive returns. Through intelligent scheduling algorithms and data distribution strategies, system resources are maximized, memory access latency and cache access latency are reduced, thereby improving overall system performance. This not only helps to improve the operating efficiency of compute-intensive applications, but also helps to achieve higher throughput and better user experience in resource-constrained environments.
[0108] Figure 7 A process 700 of another embodiment of a scheduling optimization method according to the present disclosure is shown. The scheduling optimization method comprises the following steps:
[0109] Step 701, obtaining monitoring data of core indicators of the business module.
[0110] Step 702 , comparing the monitoring data with the expected threshold of the core indicator according to the time window to determine the resource saturation type.
[0111] In this embodiment, the specific operations of steps 701-702 are already described in Figure 6 In the illustrated embodiment, steps 601-602 are described in detail and will not be repeated here.
[0112] Step 703, determining whether the entire machine resources are saturated.
[0113] In this embodiment, the saturation of the entire machine resources is judged. If the entire machine resources are saturated, step 704 is executed; if the entire machine resources are not saturated, step 705 is executed.
[0114] Step 704: Migrate at least part of the business modules to other machines.
[0115] In this embodiment, if the resources of the entire machine are saturated, the cluster layer may be notified to migrate at least part of the business modules to other machines.
[0116] In response to the problem of machine resource saturation, it will be detected at the single machine layer that effective optimization cannot be performed, that is, the resource demand can no longer be met on the local machine. At this time, you can request the cluster layer resource management platform to migrate the business module.
[0117] In some embodiments, the business modules are sorted according to preset indicators, and the business modules sorted in preset positions are migrated to other machines, thereby quickly alleviating the problem of resource saturation of the entire machine. The preset indicators may include but are not limited to: saturated resource utilization, business importance, etc. For example, the business modules are sorted in descending order according to saturated resource utilization, and the top three business modules are migrated to other machines.
[0118] Step 705, determining whether NUMA resource saturation occurs.
[0119] In this embodiment, if the whole machine resources are not saturated, NUMA resource saturation determination is performed. If the NUMA resources are saturated, step 706 is executed; if the NUMA resources are not saturated, step 707 is executed.
[0120] Step 706: unbind the bound NUMA node and schedule the service module to the unbound NUMA node for execution.
[0121] In this embodiment, if serious NUMA load saturation and load imbalance occur, NUMA dynamic unbinding can be performed, for example, unbinding the bound NUMA nodes and scheduling the service module to the unbound NUMA node for execution.
[0122] Step 707, determine whether NUMA unbinding exists.
[0123] In this embodiment, if there is no serious NUMA load saturation and load imbalance, it can be determined whether there is NUMA unbinding. If there is NUMA unbinding, step 708 is executed; if there is no NUMA unbinding, step 709 is executed. The occurrence of NUMA unbinding may be caused by the unbinding of NUMA nodes due to NUMA resource saturation in the previous round of resource scheduling.
[0124] Step 708: Bind the service module to the NUMA node according to the memory access bandwidth and / or the amount of memory occupied, and schedule the service module to the bound NUMA node for execution.
[0125] In this embodiment, if NUMA unbinding exists, the business module can be bound to the NUMA node according to the memory access bandwidth, and the business module can be scheduled to the bound NUMA node for execution; or the business module can be bound to the NUMA node according to the amount of occupied memory, and the business module can be scheduled to the bound NUMA node for execution; or the business module can be bound to the NUMA node according to the memory access bandwidth and the amount of occupied memory, and the business module can be scheduled to the bound NUMA node for execution.
[0126] In order to ensure that the memory access latency of the business module is as small as possible, the amount of memory occupied by the business module in each NUMA node can be determined. Combined with the memory access bandwidth of each NUMA node, the NUMA node that needs to be bound to the business module is determined. For example, the NUMA node with the largest amount of memory occupied by the business module is bound to the business module. Alternatively, the NUMA node with the largest memory access bandwidth is bound to the business module. Alternatively, the memory access bandwidth and the amount of memory occupied are comprehensively considered, and the weighted sum of the two is taken, and the NUMA node that needs to be bound to the business module is selected based on the weighted result.
[0127] Step 709: Determine whether NUCA resource saturation occurs and performance is degraded.
[0128] In this embodiment, if there is no NUMA unbinding, NUCA resource saturation and performance judgment can be performed. If NUCA resource saturation occurs and performance is reduced, step 710 is executed; if NUCA resource saturation does not occur or performance is reduced, the process ends.
[0129] Step 710: dynamically unbind or bind the core of the NUCA, and schedule the service module to the corresponding core of the NUCA for execution.
[0130] In this embodiment, if NUCA resources are saturated and performance is degraded, the NUCA cores may be dynamically unbound or bound, and the service modules may be scheduled to the corresponding NUCA cores for execution.
[0131] If NUCA resources are saturated and performance degrades, you can recalculate the allocation of core resources corresponding to NUCA, and perform binding and unbinding. When a core set that meets the conditions can be allocated, rebind; if there is no core set that meets the conditions, unbind. In addition, pay attention to the NUMA sensitivity of the business module after unbinding. If NUMA is sensitive, you need to bind NUMA resources, otherwise do not bind.
[0132] In some embodiments, the number of bound cores can be calculated based on NUCA sharing sensitivity, resource usage, and utilization thresholds. If there is a cache set corresponding to the number of bound cores, the cores corresponding to the cache set are bound, and the business module is scheduled to the bound NUCA cores for execution; if there is no cache set corresponding to the number of bound cores, the bound NUCA cores are unbound, and the business module is scheduled to the unbound NUCA cores for execution.
[0133] It is understandable that when rebinding, it is necessary to consider the NUMA node or NUCA core where the running business module is located, and when selecting a new NUMA node or NUCA core for binding, follow the smallest possible changes. For example, if it is already bound to {0, 1, 2, 3}, the new binding target can be {2, 3, 4, 5} or {24, 25, 26, 27}, then {2, 3, 4, 5} is preferred.
[0134] The disclosed embodiment provides a scheduling optimization method, which allows the business to transparently and efficiently utilize the hardware of the NUMA or NUCA architecture by binding the business module with the hardware resources. In addition, through the close integration of the single-machine layer and the cluster layer, the business module can transparently utilize the hardware resources without modification, and has a wide range of applications. In addition, the dynamic optimization strategy can be dynamically optimized in combination with the runtime characteristics of the business module to ensure continuous positive returns. Through intelligent scheduling algorithms and data distribution strategies, system resources are maximized, memory access latency and cache access latency are reduced, thereby improving overall system performance. This not only helps to improve the operating efficiency of compute-intensive applications, but also helps to achieve higher throughput and better user experience in resource-constrained environments.
[0135] Further references Figure 8 As an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a scheduling device, which is similar to Figure 4 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.
[0136] like Figure 8As shown, the scheduling device 800 of this embodiment may include: a first determination module 801, a second determination module 802, a binding module 803 and a scheduling module 804. The first determination module 801 is configured to determine the NUMA sensitivity of non-uniform memory access and the NUCA sensitivity of non-uniform cache access based on the memory access index; the second determination module 802 is configured to determine the sensitivity category based on the NUMA sensitivity and the NUCA sensitivity; the binding module 803 is configured to bind the service module to the NUMA node or the core of the NUCA based on the affinity policy corresponding to the sensitivity category; the scheduling module 804 is configured to schedule the service module to the bound NUMA node or the core of the NUCA for execution.
[0137] In this embodiment, in the scheduling device 800, the specific processing of the first determination module 801, the second determination module 802, the binding module 803 and the scheduling module 804 and the technical effects thereof can be referred to respectively. Figure 4 The relevant descriptions of steps 401-404 in the corresponding embodiment are not repeated here.
[0138] In some optional implementations of this embodiment, the first determination module 801 is further configured to: calculate the NUMA delay sensitivity based on the waiting time for accessing the memory, the waiting time for accessing the cache, the amount of data in the resident set size RSS memory page, the amount of data in the cache memory page, and the memory access bandwidth or the number of memory accesses; and determine the NUMA sensitivity based on the NUMA delay sensitivity.
[0139] In some optional implementations of this embodiment, the first determination module 801 is further configured to: calculate the NUCA delay sensitivity based on the last-level cache LLC delay sensitivity coefficient and the LLC usage time ratio; calculate the NUCA capacity sensitivity based on the LLC capacity sensitivity coefficient, the business working set size and the local LLC capacity; calculate the NUCA sharing sensitivity based on the LLC shared data sensitivity coefficient, the business shared data size and the number of concurrent threads of shared data; determine the NUCA sensitivity based on the NUCA delay sensitivity, the NUCA capacity sensitivity and the NUCA sharing sensitivity.
[0140] In some optional implementations of this embodiment, the binding module 803 includes: a first binding submodule, configured to bind the business module to the core of the NUCA according to resource usage if the sensitivity of the NUCA is greater than a first preset sensitivity threshold.
[0141] In some optional implementations of this embodiment, the first binding submodule is further configured to: calculate the number of bound cores based on NUCA sharing sensitivity, resource usage and utilization threshold; select a cache set corresponding to the number of bound cores; and bind the business module to the core corresponding to the cache set.
[0142] In some optional implementations of this embodiment, the binding module 803 includes: a second binding sub-module, configured to bind the business module to the NUMA node according to memory access bandwidth and / or occupied memory if the sensitivity of NUCA is not greater than the first preset sensitivity threshold and the sensitivity of NUMA is greater than the second preset sensitivity threshold.
[0143] In some optional implementations of this embodiment, the binding module 803 includes: a third binding sub-module, configured to not bind the business module to the NUMA node or the NUCA core if the sensitivity of NUMA is not greater than the second preset sensitivity threshold and the sensitivity of NUCA is not greater than the first preset sensitivity threshold.
[0144] In some optional implementations of this embodiment, the scheduling device 800 also includes: an acquisition module, configured to acquire monitoring data of core indicators of the business module; a third determination module, configured to compare the monitoring data with the expected threshold of the core indicators according to the time window to determine whether resource saturation occurs; and an optimization module, configured to perform scheduling optimization on the business module in response to the occurrence of resource saturation according to the type of resource saturation.
[0145] In some optional implementations of this embodiment, the optimization module includes: a first optimization submodule configured to migrate at least part of the business modules to other machines in response to the occurrence of saturation of the resources of the entire machine.
[0146] In some optional implementations of this embodiment, the first optimization submodule is further configured to: sort the business modules according to preset indicators, and migrate the business modules sorted at preset positions to other machines, wherein the preset indicators include at least one of the following: saturated resource utilization, business importance.
[0147] In some optional implementations of this embodiment, the optimization module includes: a second optimization sub-module, configured to unbind the bound NUMA nodes in response to NUMA resource saturation, and schedule the business module to an unbound NUMA node for execution.
[0148] In some optional implementations of this embodiment, the optimization module includes: a third optimization sub-module, which is configured to, in response to the absence of NUMA resource saturation but the presence of NUMA unbinding, bind the business module to the NUMA node according to memory access bandwidth and / or occupied memory, and schedule the business module to the bound NUMA node for execution.
[0149] In some optional implementations of this embodiment, the optimization module includes: a fourth optimization submodule configured to dynamically unbind or bind the NUCA cores in response to NUCA resource saturation and performance degradation.
[0150] In some optional implementations of this embodiment, the fourth optimization submodule is further configured to: calculate the number of bound cores based on NUCA sharing sensitivity, resource usage and utilization threshold; in response to the existence of a cache set corresponding to the number of bound cores, bind the cores corresponding to the cache set, and schedule the business module to the cores of the bound NUCA for execution; in response to the absence of a cache set corresponding to the number of bound cores, unbind the cores of the bound NUCA, and schedule the business module to the cores of the unbound NUCA for execution.
[0151] In the technical solution disclosed herein, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.
[0152] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.
[0153] Fig. 9 A schematic block diagram of an example electronic device 900 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.
[0154] like Fig. 9 As shown, the device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the device 900 can also be stored. The computing unit 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.
[0155] A number of components in the device 900 are connected to the I / O interface 905, including: an input unit 906, such as a keyboard, a mouse, etc.; an output unit 907, such as various types of displays, speakers, etc.; a storage unit 908, such as a disk, an optical disk, etc.; and a communication unit 909, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 909 allows the device 900 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.
[0156] The computing unit 901 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 901 performs the various methods and processes described above, such as the scheduling method. For example, in some embodiments, the scheduling method may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 908. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 900 via the ROM 902 and / or the communication unit 909. When the computer program is loaded into the RAM 903 and executed by the computing unit 901, one or more steps of the scheduling method described above may be performed. Alternatively, in other embodiments, the computing unit 901 may be configured to perform the scheduling method in any other appropriate manner (e.g., by means of firmware).
[0157] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.
[0158] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.
[0159] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
[0160] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).
[0161] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.
[0162] A computer system may include a client and a server. The client and the server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises through computer programs running on respective computers and having a client-server relationship to each other. The server may be a cloud server, a server in a distributed system, or a server combined with a blockchain.
[0163] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions provided by this disclosure can be achieved, and this document does not limit this.
[0164] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.
Claims
1. A scheduling method, comprising: Based on the memory access index, determine the NUMA sensitivity of non-uniform memory access and the NUCA sensitivity of non-uniform cache access; Determining a sensitivity category based on the NUMA sensitivity and the NUCA sensitivity; Binding the service module to the NUMA node or the core of the NUCA based on the affinity policy corresponding to the sensitive category; The service module is scheduled to be executed on the bound NUMA node or NUCA core.
2. The method according to claim 1, wherein: The determining the NUMA sensitivity of non-uniform memory access based on the memory access indicator includes: Calculate NUMA latency sensitivity based on the waiting time for memory access, the waiting time for cache access, the amount of data in the resident set size (RSS) memory page, the amount of data in the cache memory page, and the memory access bandwidth or the number of memory accesses. Based on the NUMA latency sensitivity, the NUMA sensitivity is determined.
3. The method according to claim 1, wherein: The determining the NUCA sensitivity of the non-uniform cache access based on the memory access indicator includes: Calculate the NUCA delay sensitivity based on the last-level cache LLC delay sensitivity coefficient and LLC usage time ratio; Calculate NUCA capacity sensitivity based on LLC capacity sensitivity coefficient, business working set size, and local LLC capacity; Calculate NUCA sharing sensitivity based on LLC shared data sensitivity coefficient, business shared data size, and number of concurrent threads sharing data; The NUCA sensitivity is determined based on the NUCA delay sensitivity, the NUCA capacity sensitivity, and the NUCA sharing sensitivity.
4. The method according to claim 1, wherein: The binding of the service module to the NUMA node or the core of the NUCA based on the affinity policy corresponding to the sensitive category includes: If the sensitivity of the NUCA is greater than a first preset sensitivity threshold, the service module is bound to the core of the NUCA according to resource usage.
5. The method according to claim 4, wherein: Binding the service module to the core of the NUCA according to the resource usage includes: Calculate the number of bound cores based on NUCA sharing sensitivity, resource usage, and utilization thresholds; Selecting a cache set corresponding to the number of bound cores; Bind the service module to the core corresponding to the cache set.
6. The method according to claim 1, wherein: in, The binding of the service module to the NUMA node or the core of the NUCA based on the affinity policy corresponding to the sensitive category includes: If the sensitivity of NUCA is not greater than a first preset sensitivity threshold, and the sensitivity of NUMA is greater than a second preset sensitivity threshold, the service module is bound to the NUMA node according to memory access bandwidth and / or occupied memory amount.
7. The method according to claim 1, wherein: The binding of the service module to the NUMA node or the core of the NUCA based on the affinity policy corresponding to the sensitive category includes: If the sensitivity of NUMA is not greater than the second preset sensitivity threshold, and the sensitivity of NUCA is not greater than the first preset sensitivity threshold, the service module is not bound to the NUMA node or the core of NUCA.
8. The method according to any one of claims 1 to 7, wherein: The method further comprises: Obtain monitoring data of core indicators of the business module; Comparing the monitoring data with the expected threshold of the core indicator according to the time window to determine whether resource saturation occurs; In response to occurrence of resource saturation, scheduling optimization is performed on the service module according to the type of resource saturation.
9. The method according to claim 8, wherein: The scheduling optimization of the service module according to the resource saturation type includes: In response to the occurrence of resource saturation of the entire machine, at least part of the business modules are migrated to other machines.
10. The method according to claim 9, wherein: Migrating at least part of the business module to another machine includes: The business modules are sorted according to preset indicators, and the business modules sorted at preset positions are migrated to the other machines, wherein the preset indicators include at least one of the following: saturated resource utilization rate and business importance.
11. The method according to claim 9, wherein: The scheduling optimization of the service module according to the resource saturation type includes: In response to the occurrence of NUMA resource saturation, the bound NUMA nodes are unbound, and the service module is scheduled to an unbound NUMA node for execution.
12. The method according to claim 11, wherein: The scheduling optimization of the service module according to the resource saturation type includes: In response to the absence of NUMA resource saturation but the presence of NUMA unbinding, the service module is bound to the NUMA node according to memory access bandwidth and / or occupied memory amount, and the service module is scheduled to the bound NUMA node for execution.
13. The method according to claim 9, wherein: The scheduling optimization of the service module according to the resource saturation type includes: In response to NUCA resource saturation and performance degradation, the cores of the NUCA are dynamically unbound or bound.
14. The method according to claim 13, wherein: The dynamic unbinding or binding of the NUCA core includes: Calculate the number of bound cores based on NUCA sharing sensitivity, resource usage, and utilization thresholds; In response to the existence of a cache set corresponding to the number of bound cores, binding the cores corresponding to the cache set, and scheduling the service module to the cores of the bound NUCA for execution; In response to the absence of a cache set corresponding to the number of bound cores, the bound cores of the NUCA are unbound, and the service module is scheduled to be executed on the cores of the unbound NUCA.
15. A scheduling device, comprising: A first determination module is configured to determine the non-uniform memory access NUMA sensitivity and the non-uniform cache access NUCA sensitivity based on the memory access indicator; A second determination module is configured to determine a sensitivity category based on the NUMA sensitivity and the NUCA sensitivity; A binding module configured to bind the service module to the NUMA node or the core of the NUCA based on the affinity policy corresponding to the sensitive category; The scheduling module is configured to schedule the service module to the bound NUMA node or NUCA core for execution.
16. The device according to claim 15, wherein: The first determining module is further configured to: Calculate NUMA latency sensitivity based on the waiting time for memory access, the waiting time for cache access, the amount of data in the resident set size (RSS) memory page, the amount of data in the cache memory page, and the memory access bandwidth or the number of memory accesses. Based on the NUMA latency sensitivity, the NUMA sensitivity is determined.
17. The device according to claim 15, wherein: The first determining module is further configured to: Calculate the NUCA delay sensitivity based on the last-level cache LLC delay sensitivity coefficient and LLC usage time ratio; Calculate NUCA capacity sensitivity based on LLC capacity sensitivity coefficient, business working set size, and local LLC capacity; Calculate NUCA sharing sensitivity based on LLC shared data sensitivity coefficient, business shared data size, and number of concurrent threads sharing data; The NUCA sensitivity is determined based on the NUCA delay sensitivity, the NUCA capacity sensitivity, and the NUCA sharing sensitivity.
18. The device according to claim 15, wherein: The binding module includes: The first binding submodule is configured to bind the service module to the core of the NUCA according to resource usage if the sensitivity of the NUCA is greater than a first preset sensitivity threshold.
19. The device according to claim 18, wherein: The first binding submodule is further configured to: Calculate the number of bound cores based on NUCA sharing sensitivity, resource usage, and utilization thresholds; Selecting a cache set corresponding to the number of bound cores; Bind the service module to the core corresponding to the cache set.
20. The device according to claim 15, wherein: The binding module includes: The second binding submodule is configured to bind the business module to the NUMA node according to the memory access bandwidth and / or the amount of occupied memory if the sensitivity of NUCA is not greater than the first preset sensitivity threshold and the sensitivity of NUMA is greater than the second preset sensitivity threshold.
21. The device according to claim 15, wherein: The binding module includes: The third binding submodule is configured to not bind the service module to the NUMA node or the core of the NUCA if the sensitivity of the NUMA is not greater than the second preset sensitivity threshold and the sensitivity of the NUCA is not greater than the first preset sensitivity threshold.
22. The device according to any one of claims 15 to 21, wherein: The device also includes: An acquisition module, configured to acquire monitoring data of core indicators of the business module; A third determination module is configured to compare the monitoring data with the expected threshold of the core indicator according to the time window to determine whether resource saturation occurs; The optimization module is configured to perform scheduling optimization on the service module in response to occurrence of resource saturation according to the type of resource saturation.
23. The device according to claim 22, wherein: The optimization module includes: The first optimization submodule is configured to migrate at least part of the business modules to other machines in response to the occurrence of resource saturation of the entire machine.
24. The device according to claim 23, wherein: The first optimization submodule is further configured to: The business modules are sorted according to preset indicators, and the business modules sorted at preset positions are migrated to the other machines, wherein the preset indicators include at least one of the following: saturated resource utilization rate and business importance.
25. The device according to claim 22, wherein: The optimization module includes: The second optimization submodule is configured to unbind the bound NUMA nodes in response to NUMA resource saturation, and schedule the service module to an unbound NUMA node for execution.
26. The device according to claim 25, wherein The optimization module includes: The third optimization submodule is configured to, in response to the absence of NUMA resource saturation but the presence of NUMA unbinding, bind the business module to the NUMA node according to memory access bandwidth and / or occupied memory, and schedule the business module to the bound NUMA node for execution.
27. The device according to claim 22, wherein: The optimization module includes: The fourth optimization submodule is configured to dynamically unbind or bind the NUCA cores in response to NUCA resource saturation and performance degradation.
28. The device according to claim 27, wherein The fourth optimization submodule is further configured to: Calculate the number of bound cores based on NUCA sharing sensitivity, resource usage, and utilization thresholds; In response to the existence of a cache set corresponding to the number of bound cores, binding the cores corresponding to the cache set, and scheduling the service module to the cores of the bound NUCA for execution; In response to the absence of a cache set corresponding to the number of bound cores, the bound cores of the NUCA are unbound, and the service module is scheduled to be executed on the cores of the unbound NUCA.
29. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 14.
30. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the method of any one of claims 1-14.
31. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 14.
