Dynamic partitioning of applications between clients and servers

• US 8,935,317 B2
• Filed: 06/23/2010
• Issued: 01/13/2015
• Est. Priority Date: 06/23/2010
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Jain, Navendu
• Primary Examiner(s): Duong, Oanh
• Application Number:
  US12/821,161
• Publication Number:
  US 20110320520A1
• Time in Patent Office: 1,665 Days
• Field of Search: 709/203, 709/223, 709/224, 709/226, 718/1
• US Class Current: 709/203
• CPC Class Codes: G06F 9/5072 Grid computing
• CPC Class Codes: G06F 9/5094 where the allocation takes into account power or heat criteria power management in computers in general G06F1/3203; thermal management in computers in general G06F1/206
• CPC Class Codes: Y02D 10/00 Energy efficient computing, e.g. low power processors, power management or thermal management

Abstract

Optimization mechanism that dynamically splits the computation in an application (e.g., cloud), that is, which parts run on a client (e.g., mobile) and which parts run on servers in a datacenter. This optimization can be based on application characteristics, network connectivity (e.g., latency, bandwidth, etc.) between the client and the datacenter, power or energy available at the client, size of the application objects, load in the datacenter, security and privacy concerns (e.g., cannot share all data on the client with the datacenter), and other criteria, as desired.

Claims

1. A computer-implemented optimization system having computer readable media that store executable instructions executed by a processor, comprising:

a cloud application that comprises client components of a client device and server components of a server configured to process workload;

a request component configured to receive a request from the cloud application to process portions of the workload via the client components and the server components, the request includes client resource information of available client resources to process the workload;

an optimization component that models dynamic application partitioning between the client device and the server by obtaining current values of client-specific parameters of the client resource information and computing an optimal dynamic partitioning of the client and server components to process the workload between the client device and the server based on the client-specific parameters of the client resource information; and

a microprocessor that executes computer-executable instructions associated with at least one of the request component or the optimization component.

2. The system of claim 1, wherein the optimal partitioning is based on at least one of energy consumption of the client device, resource footprint of the client device, data dependencies, network connectivity, or service level agreement.

3. The system of claim 1, wherein the optimal partitioning is performed on a per-request basis.

4. The system of claim 1, wherein the optimization component sends a response to the cloud application that defines which client components to run locally against the workload.

5. The system of claim 1, wherein the cloud application is a client application modeled as a directed data flow graph of vertices as processing modules and edges as communication links, the graph having associated costs where each vertex has an associated computation cost, storage cost, and memory footprint cost, and each edge has an associated bandwidth capacity cost and object size cost.

6. The system of claim 5, wherein the costs are measured offline using profiling.

7. The system of claim 5, wherein the costs are learned online during execution of the client application and the server application.

8. The system of claim 1, wherein the optimal partitioning is based on at least one of computation and storage costs of components running on the client device, memory footprint to run components on the client device, bandwidth needed based on the partitioning, data dependencies, power usage by the client device, end-to-end latency as a function of compute time and transmission latency, conservation of minimum battery life of client, or datacenter utilization where the server is sited.

9. The system of claim 1, wherein the optimization component pushes one or more components to the client device to run against the workload.

10. A computer-implemented optimization system having computer readable media that store executable instructions executed by a processor, comprising:

a request component that receives a request from a cloud application to process workload, the cloud application comprises client components of a client and server components of a server for processing the workload;

an optimization component that models dynamic partitioning of the cloud application by obtaining current values of client-specific parameters of the client components and the server components and computing an optimal dynamic partitioning of the client components and server components to process the workload between the client and the server based on the client-specific parameters, client resources and server resources, and sends a response to the client that defines which client components to run locally against the workload; and

a microprocessor that executes computer-executable instructions associated with at least one of the request component or the optimization component.

11. The system of claim 10, wherein the optimal partitioning is based on at least one of energy consumption of the client, resource footprint of the client, network connectivity, security and privacy data, computation concerns, or service level agreement.

12. The system of claim 10, wherein the optimal partitioning is based on speculative execution of server-side components and buffering externally-visible output at the client until output of server-side computation is received.

13. The system of claim 10, wherein the optimal partitioning is based on a middle system between the client and a datacenter.

14. The system of claim 10, where the optimal partitioning is based on a software clone of a physical client device may be hosted in the client on a powerful server which may perform computations and storage on behalf of or in collaboration with the physical client device and may communicate with the client device as needed.

15. The system of claim 10, where the optimal partitioning considers trade-offs between communication latency, precision, and battery consumption.

16. The system of claim 10, wherein the optimal partitioning is based on at least one of computation costs of components running on the client device, memory footprint to run components on the client device, bandwidth needed based on the partitioning, power usage by the client device, end-to-end latency as a function of compute time and transmission latency, data dependencies, conservation of minimum battery life of the client, or datacenter utilization where the server is sited.

17. The system of claim 10, wherein the cloud application is modeled to include computation and storage costs associated with processor resources of the client, a computation cost associated with processor resources of the server, a memory footprint cost incurred during execution, a bandwidth capacity cost between the client and the server, and an object size cost of data transferred between the client and the server during computations.

18. The system of claim 17, wherein the costs are measured offline using profiling or learned online during execution of the client and the server.

19. A computer-implemented optimization method performed by a computer system executing machine-readable instructions, the method, comprising acts of:

receiving from a cloud application a request for processing workload via a client device and a server;

receiving client resource availability information of the client device and server resource availability information of the server to process the workload;

modeling dynamic application partitioning of the cloud application between the client device and the server by obtaining current values of parameters of the respective client and server resource availability information and computing an optimal dynamic partitioning of the cloud application;

partitioning the computation of components of the cloud application that include server components of the server and client components of the client device between the client device and the server based on the current values of the parameters of the respective client and server resource availability information; and

processing the workload using the components as partitioned.

20. The method of claim 19, further comprising repeating receiving, partitioning, and processing based on a new request received from the client device for processing a new workload.

21. The method of claim 19, further comprising sending a response to the client device that defines which of the client components to run locally on the workload.

22. The method of claim 19, further comprising partitioning the components to minimize a utility function that considers computational latency at the client device, communication latency between the client device and the server, and energy consumed at the client device in computing and communicating data for the client application.

23. The method of claim 19, further comprising partitioning the components based on memory footprint associated with the client components and the server components.

24. The method of claim 19, further comprising caching a database of at least one of application state, intermediate results, features, or objects of interest on the client device and on a datacenter side.

25. The method of claim 19, further comprising computing characteristics of one or more middle components as part of receiving, partitioning, and processing.

26. The method of claim 19, further comprising partitioning the components based on failures of execution of different components and an operating environment.

27. The method of claim 19, further comprising partitioning the components based on a utility computing model of cloud infrastructures where utilization of server resources in a datacenter incurs monetary costs which are dynamic.

28. The method of claim 19, further comprising partitioning the components based on developer or user preferences on how and where to run the components on a client side and a server side, which change during application execution resulting in components migrating from client to server, or server to client, and between different server nodes.