International Business Machines Corporation

Robert J Urguhart

Frank Eliot Levine

Robert Francis Berry

Christopher Michael Richardson

Chester Charles John, Jr

Riaz Yousuf Hussain

William Preston Alexander, Iii

Jesse Mannes Gordon

Donald Lawrence Mulvey

Rangarajan Thiruvenkata Chari

An apparatus and method for cataloging symbolic data for use in performance analysis of computer programs is provided. The apparatus and method stores symbolic data for loaded modules during or shortly after a performance trace and utilizes the stored symbolic data when performing a performance analysis at a later time. A merged symbol file is generated for a computer program, or application, under trace. The merged symbol file contains information useful in performing symbolic resolution of address information in trace files for each instance of a module. During post processing of the trace information generated by a performance trace of a computer program, symbolic information stored in the merged symbol file is compared to the trace information stored in the trace file. The correct symbolic information in the merged symbol file for loaded modules is identified based a number of validating criteria. The correct symbolic information for the loaded modules may then be stored as an indexed database that is used to resolve address information into corresponding symbolic information when providing the trace information to a display for use by a user.

Apparatus and method for performing symbolic resolution of modules using static representations of a trace

A method and system for profiling a program using periodic trace sampling is provided. During the execution of the program, sample-based profiling of the executing program is performed&mdash;for a predetermined period, a profiler performs trace processing for the program, after which the profiler pauses and does not perform trace processing for a predetermined period or only performs lightweight processing for a predetermined period. The periods controlling the profiler may be selected by a user, and the periods may be measured by temporal or non-temporal metrics. The user may also specify parameters that are used to filter events so that profiling is performed only for specified threads or methods. The profiler cycles through these periods, during which selected events are processed to generate a profile of the execution flows within the program. For each sample period, a tree data structure is generated in which nodes of the tree data structure represent the routines of the program that execute during the sample period, as may be indicated by entry and exit events caused by the execution of the routines. At the start of each sample period, execution flow information may be used to create an initial tree data structure. When the execution of the program is complete, the tree data structures from each sample period are merged into a resulting tree data structure.

Method and system for periodic trace sampling for real-time generation of segments of call stack trees augmented with call stack position determination

An apparatus and method for cataloging symbolic data for use in performance analysis of computer programs is provided. The apparatus and method stores symbolic data for loaded modules during or shortly after a performance trace and utilizes the stored symbolic data when performing a performance analysis at a later time. A merged symbol file is generated for a computer program, or application, under trace. The merged symbol file contains information useful in performing symbolic resolution of address information in trace files for each instance of a module. During post processing of the trace information generated by a performance trace of a computer program, symbolic information stored in the merged symbol file is compared to the trace information stored in the trace file. The correct symbolic information in the merged symbol file for loaded modules is identified based on a number of validating criteria. The correct symbolic information for the loaded modules may then be stored as an indexed database that is used to resolve address information into corresponding symbolic information when providing the trace information to a display for use by a user.

Apparatus and method for creating a trace file for a trace of a computer program based on loaded module information

Apparatus and method for creating an indexed database of symbolic data for use with trace data of a computer program

A method and system for profiling a program using periodic trace sampling is provided. During the execution of the program, sample-based profiling of the executing program is performed&mdash;for a predetermined period, a profiler performs trace processing for the program, after which the profiler pauses and does not perform trace processing for a predetermined period. The periods controlling the profiler may be selected by a user, and the periods may be measured by temporal or non-temporal metrics. The profiler cycles through these periods, during which selected events are processed to generate a profile of the execution flows within the program. For each sample period, a tree data structure is generated in which nodes of the tree data structure represent the routines of the program that execute during the sample period, as may be indicated by entry and exit events caused by the execution of the routines. When the execution of the program is complete, the tree data structures from each sample period are merged into a resulting tree data structure.

Method and system for periodic trace sampling for real-time generation of segments of call stack trees

A method and system for compensating for instrumentation overhead in trace data is provided. To profile a program, the program is executed to generate trace records that are written to a trace file. A set of event trace records in a trace file is processed to determine one or more trace overhead compensation values. The trace overhead compensation values are determined by computing, for each successive pair of event trace records in the trace file, a difference between a first timestamp in a preceding event trace record and a second timestamp in a succeeding event trace record. The minimum value of all of these differences is then stored as a trace overhead compensation value representing an amount of time for instrumentation processing of an event. The one or more trace overhead compensation values are applied to the event trace records during subsequent processing of the set of event trace records.

Method and system for compensating for instrumentation overhead in trace data by detecting minimum event times

Understanding a third-party vendor's benchmark can be difficult when its source code is proprietary. In particular, the factors on which the benchmark's score depend can not easily be discovered. Furthermore, the sensitivity of the score to the execution speed of various components of the benchmark, which is important for optimizing the score in a competitive environment, is not readily accessible. The present invention shows how insertion of known delays into selected components of a benchmark can be used to elucidate its computational structure, profile it, find the sensitivity of its score to the time spent in a functional component. In all these uses of the present invention, execution of a predetermined amount of code is used to introduce a known delay into a target function or Java method. The benchmark is then re-executed to determine the delay's impact on the benchmark's score. Varying the number of iterations of the delay code, and varying the method or function in which it is inserted, are used to achieve different purposes.

Alternative profiling methodology and tool for analyzing competitive benchmarks

A method and system for compensating for instrumentation overhead in trace data by computing average minimum event times is provided. In order to profile a program, the program is executed to generate trace records that are written to a trace file. A set of trace event records are processed, and the trace events are represented as one or more nodes in a tree data structure. One or more performance statistics are stored at each node in the tree data structure, and a performance statistic at each node is processed to determine an overhead compensation value. The overhead compensation value is determined by computing a local overhead value for each node in the tree data structure. The total execution time of a routine corresponding to the event represented by the node is retrieved, and the local overhead value is computed as the average of the execution time over the number of calls to the routine and the number of calls from the routine to other routines. The minimum of all of the local overhead values is the maximum possible global value used as the overhead compensation value. The overhead compensation value is then applied to the performance statistic at each node.

Method and system for compensating for instrumentation overhead in trace data by computing average minimum event times

The present invention is a system, method, and computer readable medium for representing program event trace information in a way which is very compact and efficient, and yet supports a wide variety of queries regarding system performance. The tracing and reduction of the present invention may be dynamic, in which case information is obtained and added to the trace representation in real-time. Alternately, the tracing and reduction may be static, in which case a trace text file or binary file is obtained from a trace buffer, and the reduction takes place using the trace file as input. The trace information, whether obtained statically or dynamically, is represented as a tree of events. The present invention may be used to present many types of trace information in a compact manner which supports performance queries. For example, the tree structure of the present invention may reflect the call stacks observed during a program's execution, and statistics regarding the time spent in the various routines and call stacks may be stored at each node of the tree. The tree structure may be used to store performance information regarding Java bytecodes executed, memory allocated, or other types of performance information. The trace information is presented in a manner which is compact and efficient, and may be used to quickly and easily answer a variety of performance queries. Further, a system may be traced for long periods of time without losing useful performance information due to a limited buffer space.

Growing community of inventors

Austin, TX, United States of America

Robert J Urguhart