Abstract: A federated learning method, device, and system are provided, to improve robustness of the federated learning system. The method includes: A first client receives a first value of a parameter of a machine learning model from a server, where the first client is one of a plurality of clients; when the first value of the parameter does not meet a first condition, the first client performs a current round of training based on first training data, the machine learning model, and a local value of the parameter, to obtain a training result of the current round of training, where the first training data is data reserved on the first client; and the first client sends the training result and alarm information to the server, where the alarm information indicates that the first value of the parameter does not meet a requirement.

Abstract: The present invention discloses a hierarchical highly heterogeneous distributed system based deep learning application optimization framework and relates to the field of deep learning in the direction of computational science. The hierarchical highly heterogeneous distributed system based deep learning application optimization framework comprises a running preparation stage and a running stage. The running preparation stage is used for performing deep neural network training. The running stage performs task assignment to all kinds of devices in the distributed system and uses a data encryption module to perform privacy protection to user sensitive data.

Abstract: A request-response based paravirtualized I/O system and method relating to the fields of virtualization and cloud computing includes a request-response application, a front-end drive module and a back-end drive module. The front-end drive module and the back-end drive module interact by means of a transmit queue and a receive queue. The request-response application generates an I/O request, and the front-end drive module writes the I/O request into the transmit queue. The system has two operating modes: a notification mode and a polling mode. The system operates by default in the notification mode. When the request-response application issues a connection establishment or service request, the system switches to the polling mode. This system and method introduce an optimistic polling mechanism, combine the advantages of the notification mode and the polling mode, reduce the number of VM exits and decrease wasting of computing resources, thus improving data path performance.

Abstract: The present invention provides an adaptive computing resource allocation approach for virtual network functions, including the following two steps: Step 1: predicting VNFs' real-time computing resource requirements; Step 1.1: offline profiling different types of VNFs, to obtain a parameter relation between the required amount of computing resources and the ingress packet rate; Step 1.2: online monitoring the network traffic information of each VNF, and predicting VNFs' required amount of computing resources with combination of the parameters in Step 1.1; Step 2: reallocating computing resources based on VNFs' resource requirements. The computing resource allocation approach includes a direct allocation approach and an incremental approach. The adaptive computing resource allocation approach for virtual network functions of the present invention allocates computing resources based on VNFs' actual requirements, and remedies performance bottlenecks caused by fair allocation.

Abstract: The method includes the following steps: step 1. obtaining NUMA topology information of a host machine, and monitoring virtual machine performance events by using a kernel PMU; step 2. implementing a greedy algorithm, and a scheduling decision is obtained; step 3. scheduling, according to the scheduling decision, a virtual CPU (VCPU) and a memory of a virtual machine; step 4. after the scheduling of the virtual machine is complete, redirecting to step 1 to continue performing performance monitoring of the virtual machine.

Abstract: The present invention discloses a hierarchical highly heterogeneous distributed system based deep learning application optimization framework and relates to the field of deep learning in the direction of computational science. The hierarchical highly heterogeneous distributed system based deep learning application optimization framework comprises a running preparation stage and a running stage. The running preparation stage is used for performing deep neural network training. The running stage performs task assignment to all kinds of devices in the distributed system and uses a data encryption module to perform privacy protection to user sensitive data.

Abstract: The present invention provides a multi-resource scheduling method responding to uncertain demands in a cloud scheduler, where two computation formulas for fairness and efficiency are used as cost functions in an optimization problem. For some change sets with uncertain resource demands, a robust counterpart of an original non-linear optimization problem is computationally tractable. Therefore, the present invention models features of these sets with uncertain resource demands, i.e., establishes an ellipsoidal uncertainty model. In this model, each coefficient vector is put into a hyper-ellipsoidal space and used as a metric to measure an uncertainty degree. With the ellipsoidal uncertainty model, a non-linear optimization problem is solved and a resource allocation solution that can respond to dynamically changing demands can be obtained.

Abstract: The present invention provides a multi-resource scheduling method responding to uncertain demands in a cloud scheduler, where two computation formulas for fairness and efficiency are used as cost functions in an optimization problem. For some change sets with uncertain resource demands, a robust counterpart of an original non-linear optimization problem is computationally tractable. Therefore, the present invention models features of these sets with uncertain resource demands, i.e., establishes an ellipsoidal uncertainty model. In this model, each coefficient vector is put into a hyper-ellipsoidal space and used as a metric to measure an uncertainty degree. With the ellipsoidal uncertainty model, a non-linear optimization problem is solved and a resource allocation solution that can respond to dynamically changing demands can be obtained.

Abstract: The present invention relates to Data Center Network (DCN) flow scheduling scheme. It provides a dynamic scheduling algorithm and a hybrid of centralized and decentralized scheduling system to improve the performance of DCN and data parallel application. The scheduling system uses a central controller to collect the real-time bandwidth of each node, and schedule the priority as well as transmission rate of each network flow set combined by application context (Coflow [1]). The centralized scheduling avoids a sophisticated system design and hardware (switch) modification to comparing with full decentralized solutions. The combination of centralization and decentralization decreases the average completion time of Coflows, and eventually improve the performance of data parallel applications.

Abstract: The present invention discloses an apparatus and a method for virtual machine scheduling in a non-uniform memory access (NUMA) architecture.

Abstract: The present invention provides an adaptive computing resource allocation approach for virtual network functions, including the following two steps: Step 1: predicting VNFs' real-time computing resource requirements; Step 1.1: offline profiling different types of VNFs, to obtain a parameter relation between the required amount of computing resources and the ingress packet rate; Step 1.2: online monitoring the network traffic information of each VNF, and predicting VNFs' required amount of computing resources with combination of the parameters in Step 1.1; Step 2: reallocating computing resources based on VNFs' resource requirements. The computing resource allocation approach includes a direct allocation approach and an incremental approach. The adaptive computing resource allocation approach for virtual network functions of the present invention allocates computing resources based on VNFs' actual requirements, and remedies performance bottlenecks caused by fair allocation.

Abstract: The present invention relates to Data Center Network (DCN) flow scheduling scheme. It provides a dynamic scheduling algorithm and a hybrid of centralized and decentralized scheduling system to improve the performance of DCN and data parallel application. The scheduling system uses a central controller to collect the real-time bandwidth of each node, and schedule the priority as well as transmission rate of each network flow set combined by application context (Coflow [1]). The centralized scheduling avoids a sophisticated system design and hardware (switch) modification to comparing with full decentralized solutions. The combination of centralization and decentralization decreases the average completion time of Coflows, and eventually improve the performance of data parallel applications.

Abstract: A scheduling method for virtual processors based on the affinity of NUMA high-performance network buffer resources, including: in a NUMA architecture, when a network interface card (NIC) of a virtual machine is started, getting distribution of the buffer of the NIC on each NUMA node; getting affinities of each NUMA node for the buffer of the network interface card on the basis of an affinity relationship between each NUMA node; determining a target NUMA node in combination with the distribution of the buffer of the NIC on each NUMA node and NUMA node affinities for the buffer of the NIC; scheduling the virtual processor to the CPU on the target NUMA node. The problem of affinity between the VCPU of the virtual machine and the buffer of the NIC not being optimal in the NUMA architecture is solved to reduce the speed of VCPU processing network packets.

Abstract: The invention discloses a method for dynamic interrupt balanced mapping method based on the current scheduling states of VCPUs. When the virtual I/O APIC of an SMP virtual machine needs to map a virtual interrupt into a VCPU of the virtual machine after receiving the virtual interrupt, a part of VCPUs in the active state are analyzed according to the scheduling states of all VCPUs of the current VM in a VMM scheduler, and the virtual interrupt is mapped into the active VCPUs to obtain lower interrupt processing delay. If a plurality of VCPUs are in the active state simultaneously, the interrupt load of each active VCPU is considered further, and the interrupt is mapped into the active VCPU with the current lowest current load to further ensure balancing of interrupt processing loads of all VCPUs, and therefore, the loads of VCPUs in the SPMP structure are more symmetrical to promote balancing of the overall performance of all VCPUs in the SMP structure.

Abstract: The present invention provides a dual-machine hot standby disaster tolerance system for network service in virtualized environment. The system comprises a main server and a standby server, and the main server and the standby server are connected via network; a main VM runs on the main server; a standby VM runs on the standby server; the standby VM is in the alternative state of the application layer semantics of the main VM; the alternative state of the application layer semantics means that the standby VM can serve instead of the main server in view of the application layer semantics, and generate the correct output for any client request. The outputs of the main VM and standby VM are compared according to the alternative rule in order to determine whether a backup is needed, therefore efficiently reducing the backup frequency, and improving the system performance on the basis of ensuring rapid recovery; the present invention greatly reduces the system overhead and increases the system throughput.

Abstract: The invention discloses a method for dynamic interrupt balanced mapping method based on the current scheduling states of VCPUs. When the virtual I/O APIC of an SMP virtual machine needs to map a virtual interrupt into a VCPU of the virtual machine after receiving the virtual interrupt, a part of VCPUs in the active state are analyzed according to the scheduling states of all VCPUs of the current VM in a VMM scheduler, and the virtual interrupt is mapped into the active VCPUs to obtain lower interrupt processing delay. If a plurality of VCPUs are in the active state simultaneously, the interrupt load of each active VCPU is considered further, and the interrupt is mapped into the active VCPU with the current lowest current load to further ensure balancing of interrupt processing loads of all VCPUs, and therefore, the loads of VCPUs in the SPMP structure are more symmetrical to promote balancing of the overall performance of all VCPUs in the SMP structure.

Abstract: The present invention discloses a method for allocating processor resources precisely by means of predictive scheduling based on current credits, wherein the run queue of the Credit scheduler comprises virtual central processing units (VCPUs) with UNDER priority located at the head of the queue, VCPUs with OVER priority, VCPUs with IDLE priority located at the end of the queue and a wait queue for saving all VCPUs with overdrawn credits. Based on credit values of VCPUs, the method predicts the time of the credit overdrawing, and sets a timer which is triggered after the time to notify the Credit scheduler to stop scheduling corresponding VCPU. Thus the method effectively controls credit consumption and achieves the object of precise allocation of processor resources. The method is suitable to multi-core environment, and is also capable of reserving the advantages of the existing Credit scheduler, which are quick response for small task loads and load balancing.

Abstract: The present invention discloses a scheduling method for virtual processors based on the affinity of NUMA high-performance network buffer resources, including: in a NUMA architecture, when a network interface card of a virtual machine is started, getting distribution of the buffer of the network interface card on each NUMA node; getting affinities of each NUMA node for the buffer of the network interface card on the basis of an affinity relationship between each NUMA node; determining a target NUMA node in combination with the distribution of the buffer of the network interface card on each NUMA node and affinities of each NUMA node for the buffer of the network interface card; scheduling the virtual processor to the CPU on the target NUMA node. The present invention solves the problem that the affinity between the VCPU of the virtual machine and the buffer of the network interface card is not optimal in the NUMA architecture, so that the speed of VCPU processing network packets is not high.

Abstract: The present invention discloses a method for allocating processor resources precisely by means of predictive scheduling based on current credits, wherein the run queue of the Credit scheduler comprises VCPUs with UNDER priority located at the head of the queue, VCPUs with OVER priority, VCPUs with IDLE priority located at the end of the queue and a wait queue for saving all VCPUs with overdrawn credits. Based on credit values of VCPUs, the method predicts the time of the credit overdrawing, and sets a timer which is triggered after the time to notify the Credit scheduler to stop scheduling corresponding VCPU. Thus the method effectively controls credit consumption and achieves the object of precise allocation of processor resources. The method is suitable to multi-core environment, and is also capable of reserving the advantages of the existing Credit scheduler, which are quick response for small task loads and load balancing.