Method and apparatus for determining frequency of execution for compiled methods within a virtual machine

<u>Field of the Invention</u>

0001The present invention relates to virtual machines within computer systems. More specifically, the present invention relates to a method and an apparatus for determining the frequency of execution of compiled methods within a virtual machine.

<u>Background of the Invention</u>

0002The Java 2 Platform, Micro Edition (J2ME^TM), has become very popular in the wireless device space. Motorola, Nokia, NTT DoCoMo, RIM, Siemens, and many other key players in the wireless device industry have recently started shipping J2ME-enabled devices in very high volume. It has been estimated that over 200 million J2ME-enabled mobile phones were shipped in 2003 alone.

0003One major limitation of these wireless devices is the relatively small amount of memory available for executing programs. Many methods have been developed to optimize memory usage in these memory-constrained devices. Once such method involves using a dynamic compiler that supports both the execution of interpreted code and compiled code. Code that is executed more often is compiled and executed in compiled mode, while code that is executed less often is executed in interpreted mode. Note that the compiled code executes much faster than the interpreted code; however, the compiled code takes up a lot more space than the executed code.

0004Whether this dynamically generated compiled code resides on the object heap or in a separate code cache region of memory, this compiled code is typically evicted from memory when it is no longer frequently accessed, and this eviction typically takes place through a garbage collection (GC) process. Note that the compiled code may contain pointers to objects which are handled by the GC process.

0005Although the actual eviction of the compiled code happens during GC, the victim selection portion of the eviction process is ideally based on information gathered during mutator execution, wherein the system can gather information to rank compiled methods according to how frequently they are used.

0006As memory constraints become tighter, the quality of this ranking process becomes increasingly important for sustaining high overall execution speed. On the other hand, if the ranking mechanism becomes too complicated, obtaining ranking data may create a significant additional burden on the mutator.

0007In order to determine when a method is in use, it is a common practice to insert software hooks into specific locations in a method. These hooks can gather data and perform various computations when they are encountered during program execution.

0008Current victim selection schemes that use hooks fall into two categories, heavyweight hooks and lightweight hooks. With heavyweight hooks, the system performs some weighting calculation concerning a routine directly in the above hooks. For example, the hook can increment a counter in a Java method. In contrast, lightweight hooks produce a minimal amount of mutator slowdown by computing as little as possible in the hooks. For example, there is a self-modifying code scheme in the Connected Limited Device Configuration HotSpot Implementation (CLDC HI) within J2ME^TM that patches the Java method callee prolog only the first time a hook is executed during a given GC cycle. In subsequent uses of the same routine, there is no mutator overhead.

0009Another way to determine method usage is to use statistical sampling. Statistical sampling operates by periodically analyzing the current stack to determine when routines are executing. (The statistical sampling method also counts as lightweight, since it typically executes relatively infrequently.)

0010However, all of the above-described techniques have drawbacks. They either slow down the mutator significantly, or they do not gather enough information for good victim selection. Furthermore, the interval between GCs is generally too large to establish a useful ranking among large numbers of methods. (Reflecting multiple GC intervals in multiple bits as in CLDC HI does not fully solve this problem, since one GC interval is generally already too long.)

0011<patcit id="pcit0001" dnum="US20020144240A"><text>US 2002/0144240</text></patcit> discloses a method and system for controlling the size of compiled native code. The method employs various heuristics for determining those methods for which the corresponding stack space is to be reclaimed. In one embodiment, a JIT compiler creates a counter for each method and causes it to be incremented each time the method is invoked. Methods can then be sorted by their reference counters and a threshold defined that specifies a minimum number of invocations necessary for a method to be retained.

0012<patcit id="pcit0002" dnum="EP905619A"><text>EP-A 905619</text></patcit> discloses a list management system in which when a process is referenced, a corresponding counter is incremented and a corresponding flag is set. After a trigger interval has expired, a list management procedure accesses the flags to identify and then access those counters with non-zero counts. The processes are then ordered in accordance with the count information. The counters and the flags are reset for the next trigger interval.

<u>Summary of the Invention</u>

0013The invention is defined in the appended claims.

<u>Brief Description of the Figures</u>

0014Various embodiments of the invention will now be described in detail by way of example only with reference to the following drawings: <ul id="ul0001" list-style="none" compact="compact"><li><figref idref="f0001">FIG.1</figref> illustrates a memory-constrained computing device in accordance with one embodiment of the present invention.</li><li><figref idref="f0001">FIG. 2</figref> illustrates a flag array in accordance with one embodiment of the present invention.</li><li><figref idref="f0002">FIG. 3</figref> illustrates a reference array in accordance with one embodiment of the present invention.</li><li><figref idref="f0003">FIG. 4</figref> presents a flowchart illustrating the process of setting flag bits in accordance with one embodiment of the present invention.</li><li><figref idref="f0003">FIG. 5</figref> presents a flowchart illustrating the process of accumulating the set flag bits in accordance with an embodiment of the present invention.</li></ul>

<u>Detailed Description</u>

<u>Memory-Constrained Computing Device</u>

0015<figref idref="f0001">FIG. 1</figref> illustrates a memory-constrained computing device 100 in accordance with one embodiment of the present invention. Memory-constrained computing device 100 may include any type of computing device, including, but not limited to, a cell phone, a computer system based on a microprocessor, a digital signal processor, a portable computing device, a personal organizer, a device controller, and a computational engine within an appliance. Memory-constrained computing device 100 contains Java Virtual Machine (JVM) 102, which in turn supports execution of both compiled and interpreted methods. (The terms JAVA, JVM and JAVA VIRTUAL MACHINE are trademarks of SUN Microsystems, Inc. of Santa Clara, California.) Note that JVM 102 could be any type of platform-independent virtual machine, and is not meant to be limited to a Java Virtual Machine.

0016JVM 102 contains compiled method store 110. In order to improve performance within JVM 102, some methods are compiled and added to compiled method store 110, which enables these methods to execute faster than the interpreted methods. However, the compiled methods generally take up more space than the interpreted methods, depending on the quality of the compiled code. Furthermore, in many systems, interpreted code cannot be discarded after compilation. Hence, JVM 102 dynamically compiles only the more frequently used methods and places them in compiled store 110. Additionally, JVM 102 regularly evicts less frequently used compiled methods from compiled method store 110 to save space so that subsequent executions of the method take place in interpreted mode. Such evictions may take place during a garbage-collection operation, as well as during some other eviction operation that occurs outside of a garbage-collection operation.

<u>Flag Array</u>

0017<figref idref="f0001">FIG. 2</figref> illustrates a flag array 200 in accordance with one embodiment of the present invention. Flag array 200 is used to facilitate victim selection in a manner that achieves nearly the victim selection quality of heavyweight hooks and the overall cost of lightweight hooks as described previously. In one embodiment of the present invention, the dynamic compiler in JVM 102 generates extra code in (1) the prologue of a callee, (2) directly after a call, and (3) directly before a backwards branch. Using the above three locations, one can catch most relevant method access (except for long loop-free code stretches, which are better executed by the interpreter anyway). Other hook placements (or a subset of the above placements) might also work, though probably with less accuracy and thus reduced method ranking quality.

0018The present approach restricts the hook to a very brief inline action; namely, setting a flag in flag array 200 that corresponds to the routine (method, function) being executed. In the example illustrated in <figref idref="f0001">FIG. 2</figref>, a bit is set in flag array 200 representing the execution of a compiled method in compiled methods 202. The actual machine instruction to set the flag is chosen to address the smallest addressible memory unit (word, byte, bit) without compromising performance. Typically, it sets one byte, although in some architectures, it may set an entire word or a single bit. These flags are contained within flag array 200 which is located at an offset specified by a global variable. Each compiled method has its own index (within compiled methods 202) into flag array 200. Note that indices of evicted routines can be reused.

0019Theoretically, the fact that a method in question has been used at a given point in time will decay in usefulness as mutator execution progresses. As explained above, waiting until a GC happens to harvest the bit and use it for ranking purposes typically takes too long. Gathering more than one bit, on the other hand, increases the mutator burden. The present approach resolves this tension by introducing a third phase in which gathered bits are adjusted to represent the fading of their relevance over time.

<u>Reference Array</u>

0020<figref idref="f0002">FIG. 3</figref> illustrates a reference array 300 in accordance with one embodiment of the present invention. Reference array 300 maps the flag indices back to their respective compiled methods, such as compiled method 112, compiled method 302, and compiled method 304. Furthermore, an extra "count" field is associated with each method. This association can either be provided by a third array, count array 310, or by placing an extra field into each method instance. Using the above data structures, the system can iterate over all compiled routines and update their "count" fields according to which flags have been set. After the count fields are updated, all flags are then reset.

0021In one embodiment of the present invention, the process of updating the count fields involves two phases. In the first phase, all counts are lowered to fade the relevance of method usage over time. Then, in the second phase, the flag (with possibly some weight factor) is added only to those routines which have a flag set in the given interval.

0022In another embodiment of the present invention, there is only a single phase in which all counts for methods with set flags are updated, but the weight factor for a set flag is increased at every interval. The count lowering can thus be delayed until GC time, after which the flag weight is reset. Note that this scheme further reduces impact on the mutator.

0023In a third embodiment of the present invention, count decay is triggered by reaching a certain count threshold for any of the compiled methods. This decay process helps to model typical program execution, which typically involves frequent access to a small compiled working set of methods, wherein the working set evolves over time. During normal program execution, certain methods are executed much more frequently than other methods during a specific time period. Over time, the set of methods that are executed most frequently (the working set) changes as the program executes. In this embodiment, the count decay process is important because it reduces the importance of methods that have executed less recently, which helps to result in stronger count values for methods in the working set.

0024It is also important to consider when more methods are being interpreted rather then executed. In order to increase overall performance, in one embodiment of the present invention, a shift in the workload to the interpreter triggers a decay round. For example, this can be approximated by observing the ratio of interpreted execution to compiled execution during a specific time interval. In effect, this embodiment of the present invention makes the choice of interval length virtually irrelevant for the decay processing, in that decay "time flow" is decoupled from wall clock time.

0025Furthermore, at GC time, all methods can be ranked by their accumulated count. The system can either first sort them by count, or can repeatedly search for victims below a certain count threshold, and keep increasing the threshold until a given eviction target is met. Note that the summarization interval needs to be short enough to maintain the relevance of set flags in flag array 200. It also needs to be long enough to spread out mutator burden. In balancing these demands, the system must consider mutator speed, i.e. the rate at which flag setting hooks are hit.

0026The system also needs to handle index overflow. Each index is assigned to a method before it gets compiled. In cases where there is no available index, the system can either prevent compilation until an index will be freed, or can attempt to usurp the index of a routine with a low current ranking. In the rare case that the original index owner sets a flag, an inaccuracy occurs. However, this is likely to be a rare event, which is best ensured against by choosing a large enough array so that index overflow is unlikely to occur. To accomplish this, the array size can be proportional to the heap size. (For example, for a 1 MB heap, an array with 500 elements could be used).

<u>Setting Flag Bits</u>

0027<figref idref="f0003">FIG. 4</figref> presents a flowchart illustrating the process of setting flag bits in flag array 200 in accordance with one embodiment of the present invention. The system starts by determining when a compiled method, such as compiled method 112, is executed (step 402). This can be accomplished by the use of the execution hooks described previously. Finally, the system sets a flag bit in flag array 200 indicating that the compiled method has executed (step 406).

<u>Accumulating Set Flag Bits</u>

0028<figref idref="f0003">FIG. 5</figref> presents a flowchart illustrating the process of accumulating the set flag bits from flag array 200 into count array 310 in accordance with one embodiment of the present invention. At some periodic interval, the system scans flag array 200 for set flag bits (step 502). For each set flag bit, the system increments a count in count array 310 that corresponds to the flag bit in flag array 200 (step 504). Once all of the appropriate counts in count array 310 have been incremented, the system clears all of the flag bits in flag array 200 for the next periodic interval (step 506). Note that while incrementing count array 310, if the system detects an overflow in one of the counts, the system may shift the bits of count array 310 to introduce a decay as described above.

0029The data structures and code described in this detailed description are typically stored on a computer-readable storage medium, which may be any device or medium that can store code and/or data for use by a computer system. This includes, but is not limited to, magnetic, semiconductor, and optical storage devices such as disk drives, magnetic tape, flash memory, CDs (compact discs) and DVDs (digital versatile discs or digital video discs), and computer instruction signals embodied in a transmission medium (with or without a carrier wave upon which the signals are modulated). For example, the transmission medium may include a communications network, such as the Internet.