 |  |  | Profiling Java Workloads |
The Java Virtual Machine Profiler Interface (JVMPI) defines a general purpose mechanism to obtain profile data from a Java VM [31]. JVMPI supports CPU time profiling (using statistical sampling or code instrumentation) for threads and methods, heap profiling, and monitor contention profiling. Our work differs in that we are interested in infrastructure that measures the architectural level performance of Java applications.
Java middleware and server applications are an important class of emerging workloads. Existing research uses simulation and/or hardware performance counters to characterize these workloads. Cain et al. [13] evaluate the performance of a Java implementation of the TPC-W benchmark and compare the results to SPECweb99 and SPECjbb2000. The TPC-W benchmark models an online bookstore, and the Java implementation is a combination of Java Servlets that communicate with a database system using JDBC. Cain et al. use hardware counters to measure the performance of the entire benchmark on an IBM multiprocessor with eight RS64-III processors. They also use simulation to experiment with new architectural features. Our infrastructure enables us to distinguish the performance between the various threads of the VM and application, on different processors, at regular time intervals.
Luo and John [21] evaluate SPECjbb2000 and VolanoMark on a Pentium III processor using the Intel hardware performance counters. Seshadri, John, and Mericas [28] use hardware performance counters to characterize the performance of SPECjbb2000 and VolanoMark running on two PowerPC architectures. The focus of both studies was to compare Java server applications to the SPECint2000 benchmarks, which are written in C. They obtain aggregate counter information over a significant portion of the benchmarks running a single processor, whereas we gather hardware performance data on multiple processors on a per thread basis at regular intervals.
Karlsson et al. [20] characterize the memory performance of Java server applications using real hardware and a simulator. They measure the performance of SPECjbb2000 and ECPerf on a 16-processor Sun Enterprise 6000 server. Karlsson et al. use the hardware counters to measure coarse-grained events. The results are for the execution of the entire benchmark, and they do not distinguish between the VM and the application. Our infrastructure enables us to obtain fine-grained performance measurements information in real time.
Dufour et al. [16] introduce a set of architecture- and virtual machine-independent Java bytecode-level metrics for describing the dynamic characteristics of Java applications. Their metrics give an objective measure of aspects like array or pointer intensiveness, degree of polymorphism, allocation density, degree of concurrency, and synchronization. Our work analyzes workload characteristics on the architectural level, covers both application, library and the virtual machine behavior, and investigates behavioral patterns over time.
| | | Profiling Java Workloads |