Abstract: Embodiments are directed to a computer implemented method for generating a drift detector. The method includes generating, using a processor system, drift cases based at least in part on known drift set data of a computer system. The method further includes injecting, using the processor system, the drift cases into the computer system to generate a first data set. The method further includes applying, using the processor system, cleaning rules to the first data set to reduce a size of the first data set and generate a cleaned data set. The method further includes extracting one or more features of the cleaned data set. The method further includes normalizing the extracted one or more features of the cleaned data set. The method further includes training a machine learning system using the extracted and normalized one or more features of the cleaned data, wherein an output of the machine learning system comprises the drift detector.

Abstract: A method of identifying a root cause in a distributed computing environment is provided including collecting metric data of a plurality of application components, collecting metric correlation relationship data, collecting topology relationship data, and collecting transaction tracking relationship data that indicates a group of the application components used to execute a requested transaction. A graph is generated including a plurality of nodes corresponding to the application components by merging the metric correlation relationship data, the topology relationship data, and the transaction tracking relationship data. The nodes of the graph are traversed in an order based on a bi-directional weight for each of a plurality of edges connecting neighboring nodes among the plurality of nodes. A recommendation list is generated including at least one abnormal application component.

Abstract: A method of identifying a root cause in a distributed computing environment includes traversing a plurality of nodes in a call graph starting with an end user node. Each node corresponds to an application component. A response time is calculated between connected pairs of neighboring nodes. A weight is calculated for each of a plurality of edges connecting the neighboring nodes. The nodes are traversed starting with the end user node in an order based on the weight of each of the edges. A root cause score is calculated for each node based on traversing all of the nodes in the order based on the weight of each of the edges. A ranked list is generated.

Abstract: A method of operating a computer system including collecting, from the computer system, data indicative of variations in throughput and response time over a period of time, calculating processing power of the computer system over the period of time, recording a maximal power, calculating a standard deviation of the response time (RT-StdDev), recording the standard deviation of the response time corresponding to a time of the maximal power (RT-StdDevMaxPower), and generating a notification that the computer system is in a bottleneck state using a comparison of a current processing power to the maximal power and a comparison of the RT-StdDev to the RT-StdDevMaxPower.

Abstract: A system detection method and apparatus. Throughput and response time of a system are acquired. Throughput reflects the number of requests processed by the system that are received from outside the system. The response time reflects time required to process requests from outside the system. A system performance characteristic parameter based on the throughput and response time is obtained and used to describe congestion of the system to determine bottlenecks in the system by at least obtaining system performance characteristic parameter based on the throughput and the response time and conducting noise filtering on the system performance characteristic parameter. A bottleneck is determined based on the noise filtered system performance characteristic parameter. The system detection method and apparatus allow detecting a bottleneck and flow control are provided in a simple and effective way to ensure that requests are processed in time.

Abstract: A system detection method and apparatus. Throughput and response time of a system are acquired. Throughput reflects the number of requests processed by the system that are received from outside the system. The response time reflects time required to process requests from outside the system. A system performance characteristic parameter based on the throughput and response time is obtained and used to describe congestion of the system to determine bottlenecks in the system by at least obtaining system performance characteristic parameter based on the throughput and the response time and conducting noise filtering on the system performance characteristic parameter. A bottleneck is determined based on the noise filtered system performance characteristic parameter. The system detection method and apparatus allow detecting a bottleneck and flow control are provided in a simple and effective way to ensure that requests are processed in time.

Abstract: A non-transitory computer program product with computer readable program code for system detection and flow control. The throughput and response time of the system are acquired, where the throughput reflects the number of requests from outside the system that has been processed by the system. The response time reflects time required by the system to process requests from outside the system. Acquiring throughput and response time includes directly sampling the throughput and the response time of the system; and conducting noise filtering on the sampled throughput and response time. A system performance characteristic parameter based on the filtered throughput and response time is obtained and is used to describe congestion condition of the system to determine whether there is a bottleneck in the system. Detecting a bottleneck and flow control is provided in a simple and effective way to ensure that requests are processed in time.

Abstract: A method of operating a computer system including collecting, from the computer system, data indicative of variations in throughput and response time over a period of time, calculating processing power of the computer system over the period of time, recording a maximal power, calculating a standard deviation of the response time (RT-StdDev), recording the standard deviation of the response time corresponding to a time of the maximal power (RT-StdDevMaxPower), and generating a notification that the computer system is in a bottleneck state using a comparison of a current processing power to the maximal power and a comparison of the RT-StdDev to the RT-StdDevMaxPower.

Abstract: Embodiments are directed to a computer implemented method for generating a drift detector. The method includes generating, using a processor system, drift cases based at least in part on known drift set data of a computer system. The method further includes injecting, using the processor system, the drift cases into the computer system to generate a first data set. The method further includes applying, using the processor system, cleaning rules to the first data set to reduce a size of the first data set and generate a cleaned data set. The method further includes extracting one or more features of the cleaned data set. The method further includes normalizing the extracted one or more features of the cleaned data set. The method further includes training a machine learning system using the extracted and normalized one or more features of the cleaned data, wherein an output of the machine learning system comprises the drift detector.

Abstract: A proxy server and a hierarchical network system and a distributed workload management method. According to one embodiment of this disclosure, the proxy server includes: a rate controller configured to, based on measured request-related information and service quality parameters relative to service levels of requests, periodically determine a dispatch rate for requests of each service level, wherein the sum of the dispatch rate for respective service levels is less than or equal to a predetermined rate; and a request dispatcher configured to dispatch the requests of the corresponding service level in accordance with the dispatch rate determined by the rate controller. One aspect of the disclosure realizes a low overhead, highly scalable, simple and efficient workload management system to achieve QoS assurance and overload protection.

Abstract: A system for determining a root cause of a performance issue of a software application executing in a distributed computer system. The system includes constructing a hierarchical graph of metrics of a component of the distributed computer system based on search traces of the metrics of different users. A parent-child edge between a first metric and a second metric in the hierarchical graph is constructed when there is a sequential association between the search traces. A sibling edge between the first metric and the second metric in the hierarchical graph is constructed when there is no sequential association between the search traces. A metric investigation recommendation is provided to address the performance issue based on traversal of the hierarchical graph of the metrics.

Abstract: A method of identifying a root cause in a distributed computing environment includes traversing a plurality of nodes in a call graph starting with an end user node. Each node corresponds to an application component. A response time is calculated between connected pairs of neighboring nodes. A weight is calculated for each of a plurality of edges connecting the neighboring nodes. The nodes are traversed starting with the end user node in an order based on the weight of each of the edges. A root cause score is calculated for each node based on traversing all of the nodes in the order based on the weight of each of the edges. A ranked list is generated.

Abstract: A method of identifying a root cause in a distributed computing environment is provided including collecting metric data of a plurality of application components, collecting metric correlation relationship data, collecting topology relationship data, and collecting transaction tracking relationship data that indicates a group of the application components used to execute a requested transaction. A graph is generated including a plurality of nodes corresponding to the application components by merging the metric correlation relationship data, the topology relationship data, and the transaction tracking relationship data. Target nodes are identified among the plurality of nodes that correspond to application components having a response time that is above a predefined threshold. A plurality of paths along the graph that include the target nodes are traversed to identify at least one node corresponding to an abnormal application component.

Abstract: A system for determining a root cause of a performance issue of a software application executing in a distributed computer system. The system includes constructing a hierarchical graph of metrics of a component of the distributed computer system based on search traces of the metrics of different users. A parent-child edge between a first metric and a second metric in the hierarchical graph is constructed when there is a sequential association between the search traces. A sibling edge between the first metric and the second metric in the hierarchical graph is constructed when there is no sequential association between the search traces. A metric investigation recommendation is provided to address the performance issue based on traversal of the hierarchical graph of the metrics.

Abstract: A non-transitory computer program product with computer readable program code for system detection and flow control. The throughput and response time of the system are acquired, where the throughput reflects the number of requests from outside the system that has been processed by the system. The response time reflects time required by the system to process requests from outside the system. Acquiring throughput and response time includes directly sampling the throughput and the response time of the system; and conducting noise filtering on the sampled throughput and response time. A system performance characteristic parameter based on the filtered throughput and response time is obtained and is used to describe congestion condition of the system to determine whether there is a bottleneck in the system. Detecting a bottleneck and flow control is provided in a simple and effective way to ensure that requests are processed in time.

Abstract: A method and apparatus for power-efficiency management in a virtualized cluster system. The virtualized cluster system includes a front-end physical host and at least one back-end physical host, and each of the at least one back-end physical host comprises at least one virtual machine and a virtual machine manager. Flow characteristics of the virtualized cluster system are detected at a regular time cycle, a power-efficiency management policy is generated for each of at least one back-end physical host based on the detected flow characteristics, and the power-efficiency management policies are performed.

Abstract: A system detection method and apparatus. The throughput and response time of the system are acquired, where the throughput reflects the number of requests from outside the system that has been processed by the system. The response time reflects time required by the system to process requests from outside the system. The system performance characteristic parameter based on the throughput and response time is obtained and is used to describe congestion condition of the system to determine whether there is a bottleneck in the system. Also provided are a method and device for flow control. The system detection method and apparatus allow detecting a bottleneck and flow control in a simple and effective way to ensure that requests are processed in time.

Abstract: A method and apparatus for power-efficiency management in a virtualized cluster system. The virtualized cluster system includes a front-end physical host and at least one back-end physical host, and each of the at least one back-end physical host comprises at least one virtual machine and a virtual machine manager. Flow characteristics of the virtualized cluster system are detected at a regular time cycle, a power-efficiency management policy is generated for each of at least one back-end physical host based on the detected flow characteristics, and the power-efficiency management policies are performed.

Abstract: A method and apparatus for power-efficiency management in a virtualized cluster system. The virtualized cluster system includes a front-end physical host and at least one back-end physical host, and each of the at least one back-end physical host comprises at least one virtual machine and a virtual machine manager. Flow characteristics of the virtualized cluster system are detected at a regular time cycle, a power-efficiency management policy is generated for each of at least one back-end physical host based on the detected flow characteristics, and the power-efficiency management policies are performed.

Abstract: A system detection method and apparatus. The throughput and response time of the system are acquired, where the throughput reflects the number of requests from outside the system that has been processed by the system. The response time reflects time required by the system to process requests from outside the system. The system performance characteristic parameter based on the throughput and response time is obtained and is used to describe congestion condition of the system to determine whether there is a bottleneck in the system. Also provided are a method and device for flow control. The system detection method and apparatus allow detecting a bottleneck and flow control in a simple and effective way to ensure that requests are processed in time.