 |  |  | Implementation Details |
To enable Jikes RVM to access the C pmapi API we defined a Java class with a set of native methods that mirrors the functionality of the pmapi interface, represented by the dashed box of the JVM in Figure *. In addition to enabling our VM extensions in Jikes RVM to access these functions, the interface class can also be used to manually instrument arbitrary Java applications to gather aggregate HPM data. We are using this facility to compare the performance characteristics of Java applications when run on Jikes RVM and on other JVMs.
The main extension point in Jikes RVM was to add code in the thread scheduler's context switching sequence to read the hardware counters and real time clock on every context switch in the VM. This information is accumulated into summaries for each virtual processor and Java thread, and written into a per virtual processor trace record buffer. Each virtual processor has two dedicated 4K trace record buffers and a dedicated Java thread, called a trace writer thread, whose function is to transfer the contents of a full buffer to a file. Trace records for a virtual processor are written into an active buffer. When the buffer is full, the virtual processor signals the trace write thread and starts writing to the other buffer.
By alternating between two buffers, we continuously gather trace records with low overhead. By having a dedicated Java thread drain a full buffer, and thus, not suspend the current thread to perform this task, we avoid directly perturbing the behavior of the other threads in the system. This implementation also enables easy measurement of the overhead of writing the trace file because HPM data is gathered for all Java threads, including the threads that are writing the trace files. In our experiments 1.7% of all cycles are spent executing the trace writer threads. The overhead of reading the hardware counters and real time clock on every thread switch and storing the trace information into the buffer is in the measurement noise. Thus, the total overhead of the infrastructure is less than 2%.
Minor changes were also made in the boot sequence of the virtual machine, and in the code that creates virtual processors and Java threads to appropriately initialize data structures. The VM extensions have been available in Jikes RVM [19] since the 2.2.2 release (June 2003).
The disk space required to store trace records is a function of the trace record size, the frequency of thread switches, and the number of virtual processors. For example, running one warehouse in the SPECjbb2000 benchmark on one virtual processor, we found that 12 Kbytes per second were written with a 10 millisecond scheduling quantum.
| | | Implementation Details |