Paper Summary

This work introduces PerfScope, an online performance bug inference tool for production cloud computing systems that does not need source code and consumes very little system resources. This effort may reduce the search scope for bug-related functions on prominent open source systems to a tiny proportion.

Paper Details

Algorithm

PerfScope is only capable of diagnosing software issues, such as a program hang or a system slowdown. To keep track of recent system calls, PerfScope uses a 5-minute timeframe to collect data.

PerfScope performs its function as described below.

1. A FSM(Finite State Machine) is used, so that, numerous call episodes are constructed for each function call and performed offline and system libraries are excluded to facilitate monitoring of user-defined functions.

2. A change in the thread or CPU ID, indicative of a context shift, is detected by the online bug inference technique, which extracts the execution units.

3. A hierarchical clustering method is used to locate anomalous units, and the appearance vector of the system is recovered using Euclidean distances. These aberrant units are then mapped to bug-related functions and ranked depending on abnormality level.

4. PerfScope can deduce the system call route, which may be utilized to learn more about the bug's origin and course of spread.

Experiments

Tests on various open source server systems including Hadoop, HDFS, Casandra, Tomcat, Apache, Lighttpd, and MySQL are performanced well which showed PerfScope has appropriately discovered and solved actual performance problems. When a performance anomaly is identified, perfscope makes a 90-second system call trace to identify and rank a subset of likely issues. Moreover the paper gives examples on Cassandra, HDFS and Lighttpd-1 to show how PerfScope infer bug. And PerfScope achieves a very low overhead of 1.8 percent runtime.

Strong Points

1. PerfScope is an online bug inference tool so there is a wide range of applications and it owns a high productivity.

2. The unavailability of source code makes PerfScope simple to implement and does not infringe on user privacy.

3. Overhead is small enough to be used in a production environment.

Weak Points

1. The situation in which numerous functions cause performance concerns at the same time is not addressed.

2. The paper's PerfScope evaluation is confined to a single node server, hence its behavior in a multi-node cluster is unknown.

Brain Storm

Taking the source code into consideration may result in more accurate detection of performance issues. For example, fuzzing methods may be used to uncover potential performance issues during the compilation step.