Microsoft Technology Licensing, LLC

Linkedin Corporation

Badrinath K Sridharan

Haricharan K Ramachandra

Cuong Huu Tran

John W Nicol

Zhenyun Zhuang

David Q He

Ruixuan Hou

Toon Sripatanaskul

Michael Chang

Nicholas P Baggott

Ritesh Maheshwari

Sergiy Zhuk

Christopher Coleman

Thomas Goetze

Arman H Boehm

Tao Feng

The disclosed embodiments provide a system for detecting and managing memory inefficiency in a software program. During operation, the system obtains a first snapshot of a heap for a software program, wherein the first snapshot includes a first set of objects stored in the heap at a first time. Next, the system applies a compression technique to the first snapshot to obtain a first set of inefficiency metrics for the first set of objects, wherein each inefficiency metric in the first set of inefficiency metrics represents a memory inefficiency of an object in the heap at the first time. The system then outputs the first set of inefficiency metrics with additional attributes of the first set of objects to improve identification of the memory inefficiency in the software program.

Compression-based detection of memory inefficiency in software programs

A system, apparatus, and methods are provided for managing on-demand profiling of one or more instances of a software application executing on a plurality of machines within one or more data centers. During operation, the system executes the one or more instances of the software application on the plurality of machines. Next, the system publishes, to a command channel, a command message that comprises a profiling request, wherein the profiling request specifies a subset of the machines. The system then receives, via a data channel, one or more data messages from the subset of the machines, wherein the data messages comprise data gathered by the subset of the machines in response to receiving the command message. Next, the system then evaluates the performance of the software application by aggregating and processing the data messages. Responsive to detecting an anomaly in the performance, the system then executes one or more remedies.

On-demand profiling based on event streaming architecture

Computerized techniques are described for detecting performance degradation during features ramp up. A first version of web content is sent to first client devices. The first version of the web content lacks a particular feature. First performance data that indicates performance of each first client device with respect to processing the web content is received. A second version of the web content is sent to second client devices. The second version of the web content has the particular feature. Second performance data that indicates performance of each second client device with respect to processing the web content is received. The first performance data is compared with the second performance data to determine a difference between a processing metric associated with the first version of the web content and a processing metric associated with the second version of the web content.

Detecting and alerting performance degradation during features ramp-up

The disclosed embodiments provide a system for detecting and managing inefficiency in external services. During operation, the system obtains a snapshot of a data stream transmitted over an external service from a computer system at a first time. Next, the system applies a compression technique to the snapshot to obtain a set of inefficiency metrics for a set of data elements in the snapshot. The system then outputs the set of inefficiency metrics with additional attributes of the data stream to improve identification of inefficiency in the data stream.

Compression-based detection of inefficiency in external services

The disclosed embodiments provide a system for detecting and managing inefficiency in local storage. During operation, the system obtains a first snapshot of data in local storage of a computer system, wherein the first snapshot comprises a first set of data elements in the local storage at a first time. Next, the system applies a compression technique to the first snapshot to obtain a first set of inefficiency metrics for the first set of data elements. The system then outputs the first set of inefficiency metrics with additional attributes of the data to improve management of inefficiency in the data.

Compression-based detection of inefficiency in local storage

The disclosed embodiments provide a system for analyzing data from a monitored system. During operation, the system identifies a difference between a performance of an application and a service-level agreement (SLA) of the application. Next, the system determines a correlation between the performance of the application and a disk input/output (I/O) performance of a data storage device used by the application. When the correlation exceeds a threshold, the system outputs a recommendation to migrate the application between the data storage device and a different type of data storage device.

Performance-based migration among data storage devices

Techniques for categorizing exceptions and logs are described. For example, exception data of an exception that occurred on a machine is accessed. The exception data includes a stack trace of the exception. A determination is made that the exception is unique based on the stack track of the exception. Responsive to the determination that the exception is unique, the exception is categorized, by a machine including a memory and at least one processor, into one or more categories based on the stack trace of the exception.

Systems and methods for categorizing exceptions and logs

A system, method, and apparatus are provided for performing a transparent hybrid data storage scheme in which data are stored as blocks distributed among one or more flash-based storage devices (e.g., solid state drives) and one or more magnetic storage devices (e.g., magnetic disk drives). Files larger than a given size (e.g., 1 MB) are segmented into blocks of that size and stored on one or more devices; blocks of one file may be stored on devices of different types. Periodically, a utility function calculates utility values for each of some or all stored blocks based on frequency of access to the block, frequency of access of a particular type (e.g., random, sequential), a preference regarding where to store the block or the corresponding file, and/or other factors. Blocks having the highest utility values are subject to migration between devices of different types and/or the same type (e.g., for load-balancing).

Transparent hybrid data storage

Techniques for generating, storing, and using service call graphs are provided. In one technique, trace data items generated by different services are correlated to generate a service call graph. Trace data indicates when certain services are called and their respective latencies as a result of a client request. A service call graph may reflect a single trace or multiple traces over a particular period of time. A service call graph may be analyzed to inform administrators of a web site how a web application and the services it relies on are performing. A service call graph may be used to determine whether there are sufficient resources to support a projected increase in traffic to a web application. A service call graph may be used to estimate a cost of a web application. Multiple service call graphs may be compared to determine one or more root causes of a performance problem.

Service call graphs for website performance

Techniques for generating and using service call graphs are provided. In one technique, trace data items generated by different services are correlated to generate a service call graph. Trace data indicates when certain services are called and their respective latencies as a result of a client request. A service call graph may reflect a single trace or multiple traces over a particular period of time. A service call graph may be analyzed to inform administrators of a web site how a web application and the services it relies on are performing. A service call graph may be used to determine whether there are sufficient resources to support a projected increase in traffic to a web application. A service call graph may be used to estimate a cost of a web application. Multiple service call graphs may be compared to determine one or more root causes of a performance problem.

Estimating latency of an application

In order to prevent data thrashing and the resulting performance degradation, a computer system may maintain an application-layer cache space to more effectively use physical memory and, thus, significantly improve an application-memory hit ratio and reduce disk input-output operations. In particular, the computer system may maintain a managed memory cache that is separate from a page cache. The managed memory cache may be managed according to predefined caching rules that are separate from the caching rules in the operating system that are used to manage the page cache, and these caching rules may be application-aware. Subsequently, when data for an application is accessed, the computer system may prefetch the data and associated information from disk and store the information in the managed memory cache based on data correlations associated with the application.

Growing community of inventors

Saratoga, CA, United States of America

Badrinath K Sridharan