 |  |  | Jikes RVM |
Jikes RVM [19] is an open source research virtual machine that provides a flexible testbed for prototyping virtual machine technology. It executes Java bytecodes and runs on the Linux/IA32, AIX/PowerPC, Linux/PowerPC platforms, and OS X/PowerPC platforms. This section briefly provides background on a few relevant aspects of Jikes RVM. More details are available at the project web site [19] and in survey papers [2][12][6].
Jikes RVM is implemented in the Java programming language [3] and uses Java threads to implement several subsystems, such as the garbage collector [10] and adaptive optimization system [6]. Thus, our HPM infrastructure provides insight both into the performance characteristics of Java programs that execute on top of Jikes RVM and into the inner workings of the virtual machine itself. In particular, by gathering per-Java-thread HPM information the behavior of these VM threads is separated from application threads. However, VM services, such as memory allocation, that execute as part of the application thread are not separately distinguished.
As shown in Figure *, Jikes RVM's thread scheduler maps its M Java threads (application and VM) onto N Pthreads (user level POSIX threads). There is a 1-to-1 mapping from Pthreads to OS kernel threads. A command line argument to Jikes RVM specifies the number of Pthreads, and corresponding kernel threads, that Jikes RVM creates. The operating system schedules the kernel threads on available processors. Typically Jikes RVM creates a small number of Pthreads (on the order of one per physical processor). Each Pthread is called a virtual processors because it represents an execution resource that the virtual machine can use to execute Java threads.
To implement M-to-N threading, Jikes RVM uses compiler-supported quasi-preemptive scheduling by having the two compilers (baseline and optimizing) insert yieldpoints into method prologues, epilogues, and loop heads. The yieldpoint code sequence checks a flag on the virtual processor object; if the flag is set, then the yieldpoint invokes Jikes RVM's thread scheduler. Thus, to force one of the currently executing Java threads to stop running, the system sets this flag and waits for the thread to execute a yieldpoint sequence. The flag can be set by a timer interrupt handler (signifying that the 10ms scheduling quantum has expired) or by some other system service (for example, the need to initiate a garbage collection) that needs to preempt the Java thread to schedule one of its own daemon threads. Because yieldpoints are a subset of the GC safe points, i.e., program points where the stack references (or GC roots) are known, only the running Java threads need to reach a GC safe point to begin a stop-the-world garbage collection; threads in the scheduler's ready queue are already prepared for a garbage collection.
| | | Jikes RVM |