Check out the new USENIX Web site. next_inactive up previous


CRAMM: Virtual Memory Support for Garbage-Collected Applications

Ting Yang Emery D. Berger Scott F. Kaplan$^\dag$ J. Eliot B. Moss
tingy@cs.umass.edu emery@cs.umass.edu sfkaplan@cs.amherst.edu moss@cs.umass.edu


Dept. of Computer Science $^\dag$Dept. of Mathematics and Computer Science
University of Massachusetts Amherst Amherst College
Amherst, MA 01003-9264 Amherst, MA 01002-5000

Abstract:

Existing virtual memory systems usually work well with applications written in C and C++, but they do not provide adequate support for garbage-collected applications. The performance of garbage-collected applications is sensitive to heap size. Larger heaps reduce the frequency of garbage collections, making them run several times faster. However, if the heap is too large to fit in the available RAM, garbage collection can trigger thrashing. Existing Java virtual machines attempt to adapt their application heap sizes to fit in RAM, but suffer performance degradations of up to 94% when subjected to bursts of memory pressure.

We present CRAMM (Cooperative Robust Automatic Memory Management), a system that solves these problems. CRAMM consists of two parts: (1) a new virtual memory system that collects detailed reference information for (2) an analytical model tailored to the underlying garbage collection algorithm. The CRAMM virtual memory system tracks recent reference behavior with low overhead. The CRAMM heap sizing model uses this information to compute a heap size that maximizes throughput while minimizing paging. We present extensive empirical results demonstrating CRAMM's ability to maintain high performance in the face of changing application and system load.


1 Introduction

The virtual memory (VM1) systems in today's operating systems provide relatively good support for applications written in the widely-used programming languages of the 80's and 90's, such as C and C++. To avoid the high overhead of heavy page swapping, it is sufficient for these applications to fit their working sets in physical memory [16]. VM systems typically manage physical memory memory with an approximation of a global LRU policy [12,13,15,16,22], which works reasonably well for legacy applications.

However, garbage-collected languages are now increasingly prevalent, ranging from general-purpose languages like Java and C# to scripting languages like Python and Ruby. Garbage collection's popularity derives from its many software engineering advantages over manual memory management, including the elimination of dangling pointer errors and a drastic reduction of memory leaks.

The performance of garbage-collected applications is highly sensitive to heap size. A smaller heap reduces the amount of memory referenced, but requires frequent garbage collections that hurt performance. A larger heap reduces the frequency of collections, thus improving performance by up to 10x. However, if the heap cannot fit in available RAM, performance drops off suddenly and sharply. This is because garbage collection has a large working set (it touches the entire heap) and thus can trigger catastrophic page swapping that degrades performance and increases collection pauses by orders of magnitude [18]. Hence, heap size and main memory allocation need to be coordinated to achieve good performance. Unfortunately, current VM systems do not provide sufficient support for this coordination, and thus do not support garbage-collected applications well.

Choosing the appropriate heap size for a garbage-collected application--one that is large enough to maximize throughput but small enough to avoid paging--is a key performance challenge. The ideal heap size is one that makes the working set of garbage collection just fit within the process's main memory allocation. However, an a priori best choice is impossible in multiprogrammed environments, since the amount of main memory allocated to each process constantly changes. Existing garbage-collected languages either ignore this problem, allowing only static heap sizes, or adapt the heap size dynamically using mechanisms that are only moderately effective. For example, Figure 1 shows the effect of dynamic memory pressure on an industrial-strength Java virtual machine, BEA's JRockit [7], running a variant of the SPECjbb2000 benchmark. The solid line depicts program execution when given a fixed amount of RAM, while the dashed line shows execution under extra periodic bursts of memory pressure. This memory pressure dilates overall execution time by a factor of 220%, and degrades performance by up to 94%.

Figure 1: Impact of bursts of memory pressure on the performance on the JRockit Java virtual machine, which degrades throughput by as much as 94%.
\includegraphics[width=2.8in]{graphs/jrockit.normalized.eps}

The problem with these adaptive approaches is not that their adaptivity mechanism is broken, but rather that they are reactive. The only way these systems can detect whether the heap size is too large is to grow the heap until paging occurs, which leads to unacceptable performance degradation.

Contributions: This paper makes the following contributions. It presents CRAMM (Cooperative Robust Automatic Memory Management), a system that enables garbage-collected applications to predict an appropriate heap size, allowing the system to maintain high performance while adjusting dynamically to changing memory pressure.

CRAMM consists of two parts; Figure 2 presents an overview. The first part is the CRAMM VM system that dynamically gathers the working set size (WSS) of each process, where we define the WSS as the main memory allocation that yields a trivial amount of page swapping. To accomplish this, CRAMM VM maintains separate page lists for each process and computes an LRU reference histogram [25,27] that captures detailed reference information while incurring little overhead (around 1%).

The second part of CRAMM is its heap sizing model, which controls application heap size and is independent of any particular garbage collection algorithm. The CRAMM model correlates the WSS measured by the CRAMM VM to the current heap size. It then uses this correlation to select a new heap size that is as large as possible (thus maximizing throughput) while yielding little or no page faulting behavior. We apply the CRAMM model to five different garbage collection algorithms, demonstrating its generality.

Figure 2: The CRAMM system. The CRAMM VM system efficiently gathers detailed per-process reference information, allowing the CRAMM heap size model to choose an optimal heap size dynamically.
\includegraphics[width=2.6in]{approach.eps}

We have implemented the CRAMM VM system in the Linux kernel and the CRAMM heap sizing model in the Jikes RVM research Java virtual machine [3]. We present the results of an extensive empirical evaluation of CRAMM, including experimental measurements across 20 benchmarks and 5 garbage collectors, as well as comparison with two industrial Java implementations. These results demonstrate CRAMM's effectiveness in maintaining high performance in the face of changes in application behavior and system load.

This work builds on our previous study that introduced an early version of the CRAMM heap sizing model [28]. That study presented a model that was evaluated only in the context of trace-driven simulations. This paper builds on the previous study significantly. It refines the heap sizing model to take into account copying and non-copying regions (required to handle generational collectors), provides more accurate startup adjustment, and more effectively adapts to dynamic memory allocation by polling the underlying VM between collections. Furthermore, it is implemented in a fully functional kernel and JVM, introduces implementation strategies that make its overhead practical, has more efficient overhead control mechanisms, and presents extensive empirical results.

In addition to serving the needs of garbage-collected applications, the CRAMM VM system is the first system to our knowledge to provide per-process and per-file page management while efficiently gathering detailed reference histograms. This information can be used to implement a wide range of recently proposed memory management systems including compressed caching [27], adaptive LRU policies like EELRU [25], and informed prefetchers [20,24].

The remainder of this paper is organized as follows. Section 2 presents an overview of garbage collection algorithms and terminology used in this paper. Section 3 derives the CRAMM heap sizing model, which relates application working set size to heap size. Section 4 describes the CRAMM VM system, which gathers detailed statistics allowing it to compute the precise current process working set size. Section 5 presents empirical results, comparing static and previous adaptive approaches to CRAMM. Section 6 presents work most closely related to ours, and Section 7 concludes.


2 GC Behavior and Terminology

A garbage collector (GC) periodically and automatically finds and reclaims heap-allocated objects that a program can no longer possibly use. We now sketch how, and when, a GC may do this work, and along the way introduce GC terminology and concepts critical to understanding CRAMM.

Garbage collectors operate on the principle that if an object is unreachable via any chain of pointers starting from roots--pointers found in global/static variables and on thread stacks--then the program cannot possibly use the object in the future, and the collector can reclaim and reuse the object's space. Through a slight abuse of terminology, reachable objects are often called live and unreachable ones dead. Reference counting collectors determine (conservatively) that an object is unreachable when there are no longer any pointers to it. Here, we focus primarily on tracing collectors, which actually trace through pointer chains from roots, visiting reachable objects.

The frequency of collection is indirectly determined by the heap size: the maximum virtual memory space that may be consumed by heap-allocated objects. When allocations have consumed more than some portion of the heap size (determined by the collection algorithm), collection is invoked. Thus, the smaller the heap size, the more frequently GC occurs, and the more CPU time is spent on collection.

GC algorithms divide the heap into one or more regions. A non-generational GC collects all regions during every collection, triggering collection when some percentage of the entire heap space is filled with allocated objects. A non-generational GC may have only one region. In contrast, generational GCs partition the regions into groups, where each group of regions, called a generation, contains objects of a similar age. Most commonly, each group consists of a single region. When some percentage of the space set aside for a generation has been filled, that generation, and all younger ones, are collected. Additionally, live objects that survive the collection are generally promoted to the next older generation. New objects are typically allocated into a nursery region. This region is usually small, and thus is collected frequently, but quickly (because it is small). The generational configurations that we consider here have two generations, a nursery and a mature space. Because nursery collection generally filters out a large volume of objects that die young, mature space grows more slowly--but when it fills, that triggers a full heap collection.

Orthogonal to whether a collector is generational is how it reclaims space. Mark-sweep (MS) collection marks the reachable objects, and then sweeps across the allocation region to reclaim the unmarked ones. MS collection is non-copying in that it does not move allocated objects. In contrast, copying collectors proceed by copying reachable objects to an empty copy space, updating pointers to refer to the new copies. When done, it reclaims the previous copy space. We do not consider here collectors that compact in place rather than copying to a new region, but our techniques would work just as well for them. Notice that collectors that have a number of regions may handle each region differently. For example, a given GC may collect one region by copying, another by MS, and others it may never collect (so-called immortal spaces).

Finally, allocation and collection are intertwined. When allocating into an MS-managed region, the allocator may use free lists to find available chunks of space. When allocating into a copying region, it typically increments a free space pointer through the initially empty space. For generational collection, the nursery is usually a copy-collected space, thus allowing fast allocation. The mature space, however, may be a copying- or a non-copying-collected region, depending on the particular collector.


3 CRAMM Heap Sizing Model

The goal of the CRAMM heap sizing model is to relate heap size and working set size, so that, given a current real memory allocation, we can determine a heap size whose working set size just fits in the allocation. The working set size (WSS) for a GCed application is determined almost entirely by what happens during full collections, because full collections touch every reachable heap object. Since live and dead objects are generally mixed together, the working set includes all heap pages used for allocated objects. It also includes the space needed for copied survivors of copying regions. Thus, each non-copying region contributes its size to the working set, while each copying region adds its size plus the volume of copied survivors, which can be as much as the size of the copying region in the worst case.

Figure 3: The effect of heap size on performance and working set size (the number of pages needed to run with at most a 5% slowdown from paging).
\includegraphics[width=3.1in]{graphs/MarkSweep._213_javac.char.eps}

Several properties of GCed applications are important here. First, given adequate real memory, performance varies with heap size. For example, Figure 3 depicts the effect of different amounts of memory (the size of the garbage-collected heap) on performance. This graph is for a particular benchmark and garbage collector (the SPECjvm98 benchmark javac with a mark-sweep garbage collector), but it is typical. On the left-hand side, where the heap is barely large enough to fit the application, execution time is high. As the heap size increases, execution time sharply drops, finally running almost 250% faster. This speedup occurs because a larger heap reduces the number of collections, thus reducing GC overhead. The execution time graph has a $1/x$ shape, with vertical and horizontal asymptotes.

However, the working set size--here given as the amount of memory required to run with at most 5% elapsed time added for paging--has a linear shape. The heap size determines the working set size, as previously described. Our earlier work explores this in more detail [28]. The key observation is that working set size is very nearly linear in terms of heap size.


3.1 GC Working Set Size and Heap Sizing Model

We define heap size, $H$, as the maximum amount of space allowed to contain heap objects (and allocation structures such as free lists) at one time. If non-copy-collected regions use $N$ pages and copy-collected regions allocate objects into $C$ pages, then $H = N + 2 \times
C$. We must reserve up to $C$ pages into which to copy survivors from the original $C$ space, and the collector needs both copies until it is done. The total WSS for the heap during full collection is determined by the pages used for copied survivors, CS: $\mbox{\emph{WSS}} = N + C +
\mbox{\emph{CS}}$. Thus, heap WSS varies from $N+C$ to $N+2\times C$.

As a program runs, its usage of non-copying and copying space may vary, but it is reasonable to assume that the balance usually does not change rapidly from one full collection to the next. We call the ratio of allocable space ($N+C$) to heap size ( $N+2\times C$) the heap utilization, $u$. It varies from 50% for $N=0$ to 100% for $C=0$. Given an estimate of $u$, we can determine $N+C$ from $H$, but to determine WSS, we also need to estimate CS. Fortunately, CS is a property of the application (volume of live objects in copy-collected regions), not of the heap size. As with $u$, we can reasonably assume that CS does not change too rapidly from one full collection to the next.

When adjusting the heap size, we use this equation as our model: $ \Delta H = ( \Delta\mbox{\emph{WSS}} - \Delta\mbox{\emph{CS}} ) / u $. Notice that $\Delta\mbox{\emph{WSS}}$ is just our target WSS (i.e., the real memory allocation the OS is willing to provide) minus our current WSS. The CRAMM VM provides both of these values to the heap size manager.

Starting out: Once the JVM reaches the point where it needs to calculate an initial heap size, it has touched an initial working set of code and data. Thus, the space available for the heap is exactly the volume of free pages the VM system is willing to grant us (call that Free). We wish to set our heap size so that our worst case heap WSS during the first full collection will not exceed Free. But the worst heap WSS is exactly the heap size, so we set $H$ to the minimum of Free and the user-requested initial heap size.

Tracking the parameters: To determine the heap utilization $u$, we simply calculate it at the end of each collection, and assume that the near future will be similar. Estimating $\Delta\mbox{\emph{CS}}$ is more involved. We track the maximum value for CS that we have seen so far, maxCS, and we also track the maximum increment we have seen to CS, maxCSInc. If, after a full collection, CS exceeds maxCS, we assume CS is increasing and estimate $\Delta\mbox{\emph{CS}} =
\mbox{\emph{maxCSInc}}/2$, i.e., that it will grow by 1/2 of the largest increment. Otherwise we estimate $\Delta\mbox{\emph{CS}}$ as $\mbox{\emph{maxCS}} - \mbox{\emph{CS}}$, i.e., that CS for the next full collection will equal maxCS. After calculating $\Delta\mbox{\emph{CS}}$, we decay maxCS, multiplying it by 0.98 (a conservative policy to maintain stable maxCS), and maxCSInc, multiplying it by 0.5 (a more rapidly adjusting policy so that maxCSInc decays away quickly once maxCS reaches its stable state).

Handling nursery collections: Because nursery collections do not process the whole heap, their CS value underestimates survival from future full collections. So, if the nursery size is less than 50% of allocable space, we do not update $H$. For larger nurseries, we estimate $\Delta\mbox{\emph{CS}}$ by multiplying the size of uncollected copying space times $1+\sigma$, where $\sigma$ is the survival rate of the nursery collection, i.e., $\mbox{\emph{CS}}/\nu$, where $\nu$ is the size of the nursery.

This model is a straightforward generalization of our previous one [28], taking into account copying and non-copying regions and modeling startup effects, and eliminates the overhead (8% - 23%) caused by inaccurate startup adjustment in our previous model. Tracking of maxCS and maxCSInc also helps avoid paging. We periodically request the current Free value on allocation slow path, when the allocator tries to request a new chunk of memory from the VM (128KB for MS and 1MB for others). Once Free is less than maxCS, we trigger an immediate collection and resize the heap. This new polling mechanism allows us to adapt to bursts of memory pressure more quickly and effectively than our previous model.


4 VM System Design and Implementation

We now present the CRAMM VM system. We first describe why standard VM systems are insufficient for predictively adaptive heap sizing. We then describe the structure of the CRAMM VM system, followed by detailed discussions of how it calculates working set sizes and how it controls histogram collection overhead.

Given the heap sizing model presented in Section 3.1, the underlying VM system must provide to a GC-based process both its working set size (WSS) and its main memory allocation,2thus allowing the GC to choose a proper heap size. Unfortunately, we cannot easily obtain this information from standard VM systems, including the Linux VM.

Linux uses a global page replacement policy that manages each physical page within a single data structure for all processes and files. Linux thus has only ordinal information about all pages, giving each page a ranking among the total pool of pages. It has no cardinal information about the reference rates, nor any separation of pages according to process or file. Consequently, it cannot track the LRU reference histogram--the distribution of memory references to pages managed by an LRU queue--which is needed to determine the WSS for each process. Furthermore, it cannot predict how much it could reduce the allocations of files and other processes without inducing heavy page faulting. It therefore cannot wisely choose a main memory allocation to offer to a GC-based process. Finally, even if it chose to reduce the allocations for some files or other processes, global page replacement cannot guarantee that it will replace the pages of those processes first.

The CRAMM VM system addresses these limitations. Figure 2 gives an overview of the CRAMM VM structure and interface. For each file and process, it keeps separate page lists and an LRU reference histogram. It also tracks the mean cost of a major page fault (one that requires disk I/O) so that, along with the histogram and a desired maximum fault rate, it can compute the WSS of a process.

Its ability to compute the WSS of each file and process allows the CRAMM VM to calculate new allocations to each without causing thrashing by assigning too small an allocation. When an allocation is reduced, the separate page lists allow the VM to prefer reclaiming pages from those files and processes that are consuming more than their allocation.

A garbage collector communicates with the CRAMM VM through system calls. First, the collector registers itself as a cooperative process with the CRAMM VM at initialization time. The VM responds with the current amount of free memory, allowing the collector to pick a reasonable initial heap size. Second, after each heap collection, the collector requests a WSS estimate and a main memory allocation from the VM. The collector then uses this information to select a new heap size. If it changes its heap size, it calls on the VM to clear its old histogram, since the new heap size will exhibit a different reference pattern.

Last, the collector periodically polls the VM for an estimate of the free memory--the main memory space that could be allocated to the process without causing others to thrash. If this value is unexpectedly low, then memory pressure has suddenly increased. Either some other system activity is aggressively consuming memory (e.g. the startup of a new process), or this process has more live data (increased heap utilization), and thus is using more memory than expected. The collector responds by pre-emptively collecting the heap and selecting a new heap size.


4.1 CRAMM VM Structure

The CRAMM VM allocates a data structure, called mem_info, for each address space (an inode for files or an mm_struct for processes). This structure comprises a list of pages, an LRU reference histogram, and some additional control fields.

Figure 4: Segmented queue page lists for one address space (file or process).
\includegraphics[width=5.00in]{VMM_structure.eps}\includegraphics[width=1.50in]{histogram.eps}

Figure 4 shows the page list structure of a process. The CRAMM VM manages each address space (the space of a file or a process) much like the Linux VM manages its global queue. For the in-memory pages of each address space, it maintains a segmented queue (SEGQ) structure [5], where the active list contains the more recently used pages and the inactive list contains those less recently used. When a new page is faulted into memory, the VM places it at the head of the active list. If the addition of this page causes the active list to be too large, it moves pages from the tail of the active list to the head of the inactive list. When the process exceeds its main memory allocation, the VM removes a page from the tail of the inactive list and evicts it to disk. This page is then inserted at the head of a third segment, the evicted list. When an address space's WSS exceeds its main memory allocation, the evicted list's histogram data allows the VM to project how large the allocation must be to capture the working set.

The active list is managed using a CLOCK algorithm. The inactive list is ordered by each page's time of removal from the active list. The relative sizes of these two lists is controlled by an adaptive mechanism described in Section 4.3. Like a traditional SEGQ, all inactive pages have their access permissions removed, forcing any reference to an inactive page to cause a minor page fault. When such a page fault occurs, the VM restores the page's permissions and promotes it into the active list, and then updates the address space's histogram. The insertion of a new page into the active list may force other pages out of the active list. The VM manages the evicted list similarly; the only difference is that a reference to an evicted page triggers disk activity.

Replacement algorithm: The CRAMM VM places each mem_info structure into one of two lists: the unused list for the address spaces of files for which there are no open file descriptors, and the normal list for all other address spaces. When the VM must replace a page, it preferentially selects a mem_info from the unused list and then reclaims a page from the tail of that inactive list. If the unused list is empty, the VM selects a mem_info in a round robin manner from the normal list, and then selects a page from the tail of its inactive list.

As Section 5.2 shows, this eviction algorithm is less effective than the standard Linux VM replacement algorithm. However, the CRAMM VM structure can support standard replacement policies and algorithms while also presenting the possibility of new policies that control per-address-space main memory allocation explicitly.

Available Memory: A garbage collector will periodically request that the CRAMM VM report the available memory--the total main memory space that could be allocated to the process. Specifically, the CRAMM VM reports the available memory (available) as the sum of the process's resident set size (rss), the free main memory (free), and the total number of pages found in the unused list (unused). There is also space reserved by the VM (reserved) to maintain a minimal pool of free pages that must be subtracted from this sum:



\begin{displaymath}\mbox{\emph{available}} = \mbox{\emph{rss}} + \mbox{\emph{free}} + \mbox{\emph{unused}} - \mbox{\emph{reserved}}\end{displaymath}

This value is useful to the collector because the CRAMM VM's per-address-space structure allows it to allocate this much space to a process without causing any page swapping. Standard VM systems that use global memory management (e.g., Linux) cannot identify the unused file space or preclude the possibility of page swapping as memory is re-allocated to a process.


4.2 Working Set Size Calculation

The CRAMM VM tracks the current working set size of each process. Recall that the WSS is the smallest main memory allocation for which page faulting degrades process throughput by less than $t\%$. If $t = 0$, space may be wasted by caching pages that receive very little use. When $t$ is small but non-zero, the WSS may be substantially smaller than for $t = 0$, yet still yield only trivial page swapping. In our experiments, we chose $t = 5\%$.

In order to calculate the WSS, the VM maintains an LRU reference histogram $h$ [25,27] for each process. For each reference to a page at position $i$ of the process's page lists, the VM increments $h[i]$.3 This histogram allows the VM to calculate the number of page faults that would occur for each possible memory allocation. The VM also monitors the mean cost of a major fault (majfc) and the time $T$ that each process has spent on the CPU. To calculate the WSS, it scans the histogram backward to find the allocation at which the number of page faults is just below $(T \times t)/
\mbox{\emph{majfc}}$.

Page list position: When a page fault occurs, the referenced page is found within the page lists using a hash map. In order to maintain the histograms, the CRAMM VM must determine the position of that page within the page lists. Because a linear traversal of the lists would be inefficient, the VM attaches an AVL tree to each page list. Figure 4 shows the structure that the VM uses to calculate page list positions in logarithmic time. Specifically, every leaf node in the AVL tree points to a linked list of up to $k$ pages, where $k$ depends on the list into which the node points. Every non-leaf node is annotated with the total number of pages in its subtree; additionally, each non-leaf node is assigned a capacity that is the $k$-values of its children. The VM puts newly added pages into a buffer, and inserts this buffer into the AVL tree as a leaf node when that buffer points to $k$ pages. Whenever a non-leaf node drops to half full, the VM merges its children and adjusts the tree shape accordingly.

When a page is referenced, the VM first searches linearly to find the page's position in the containing leaf node. It then walks up the AVL tree, summing the pages in leaf nodes that point to earlier portions of the page list. Thus, given that $k$ is constant and small, determining a page's list position is performed in time proportional to the height of the AVL tree.

Because the CRAMM VM does not track references to pages in the active list, one leaf node contains pointers to all pages in the active list, and for this leaf node, $k = \infty$. For leaf nodes that point to inactive and evicted pages, $k = 64$ -- a value chosen to balance the work of linear search and tree traversal. The AVL trees have low space overhead. Suppose an application has $N$ 4KB pages, and our AVL node structure is 24 bytes long. Here, the worst case space overhead (all nodes half full, total number of nodes double the number of leaf node) is:


\begin{displaymath}\frac{((\frac{N}{64} \times 2 \times 2) \times 24)}{(N \times 2^{12})} <
0.037\%\end{displaymath}

On average, we observe that the active list contains a large portion (more than half) of the pages used by a process, and thus the overhead is even lower.

LRU histogram: Keeping one histogram entry for every page list position would incur a large space overhead. Instead, the CRAMM VM groups positions into bins. In our implementation, every bin corresponds to 64 pages (256 KB given the page size of 4 KB). This granularity is fine enough to provide a sufficiently accurate WSS measurement while reducing the space overhead substantially.

Furthermore, CRAMM dynamically allocates space for the histogram in chunks of 512 bytes. Given that a histogram entry is 8 bytes in size, one chunk corresponds to histogram entries for 16 MB of pages. Figure 4 shows the data structure for a histogram. We see that, when a process or file uses less than 64 pages (256 KB), it uses only $bin_0$. This approach optimizes space for the common case of small processes and files. Any process or file that requires more than 256 KB but less than 16MB memory uses the level_1 histogram. Larger ones use the level_2 histogram. The worst-case histogram space overhead occurs when a process uses exactly 65 pages. Here, the histogram will need about 0.2% of the memory consumed by the process. In common cases, it is about 8 bytes per 64 pages, which is less than 0.004%.

Major fault cost: Calculating WSS requires tracking the mean cost of a major page fault. The CRAMM VM keeps a single, system-wide estimate majfc of this cost. When the VM initiates a swap-in operation, it marks the page with a time-stamp. After the read completes, the VM calculates the time used to load the page. This new time is then used to update majfc.


4.3 Controlling Histogram Collection Overhead

Because the CRAMM VM updates a histogram entry at every reference to an inactive page, the size of the inactive list determines the overhead of histogram collection. If the inactive list is too large, then too much time will be spent handling minor page faults and updating histogram entries. If the inactive list is too small, then the histogram will provide too little information to calculate an accurate WSS. Thus, we want the inactive list to be as large as possible without inducing too much overhead.

The VM sets a target for minor fault overhead, expressed as a percentage increase in running time for processes, and dynamically adjusts the inactive list size according to this target. For each process, the VM tracks its CPU time $T$ and a count of its minor page faults $n$. It also maintains a system-wide minor fault cost minfc using the same approach as with majfc. It uses these values to calculate the minor fault overhead as $(n \times
\mbox{\emph{minfc}}) / T$. It performs this calculation periodically, after which it resets both $T$ and $n$. Given a target of $1\%$ and a constant threshold for deviation from that target of $0.5\%$, one of three cases may apply:

This simple adaptive mechanism, set to a $1\%$ overhead target and a $0.5\%$ deviation threshold, successfully keeps the overhead low while yielding sufficient histogram information for WSS calculations.

Size adjustment calculations: CRAMM assigns each process a target inactive size, initially $0$. When CRAMM adjusts the inactive list size, it is really setting this target size. Assume that a process has $P_A$ pages in the active list and $P_I$ in the inactive list. Depending on the overhead's relationship to its threshold, the new target will be:

By choosing the smaller of $P_A$ and $P_I$ in these equations, we make the adjustments small if either list is small, thus not changing the target too drastically. These formulas also ensure that at least some constant change is applied to the target, ensuring a change that will have some effect. We also put an upper bound on the refilling adjustment to prevent flushing too many pages into the inactive list at a time. Finally, we decrease the target inactive list size more aggressively than we increase it because low overhead is a more critical and sensitive goal than accurate histogram information. We also refill more aggressively than we increase because an absence of minor faults is a strong indication of an inadequate inactive list size.

Whenever a page is added to the active list, the VM checks the current inactive list size. If it is less than its target, then the VM moves several pages from the active list to the inactive list (8 pages in our implementation). When an adjustment triggers refilling, the VM immediately forces no more than 256 pages into the inactive list to match its new target. As this adjustment only resets the target size and usually does not move pages immediately, the algorithm is largely insensitive to the values of these parameters.

Adaptivity triggers: In the CRAMM VM, there are two events that can trigger an inactive list size adjustment. The first, adjust_interval, is based on running time, and the second, adjust_count, is based on the number of minor faults.

For every new process, its adjust_interval is initialized to a default value ( $\frac{1}{16} \mbox{\emph{sec}}$). Whenever a process is scheduled, if its running time since the last adjustment exceeds its adjust_interval value, then the VM adjusts the inactive list size.

The adjust_count variable is initialized to be $(\mbox{\emph{adjust\_interval}} \times 2\%) /
\mbox{\emph{minfc}}$. If a process suffers this number of minor faults before adjust_interval CPU time has passed, then its overhead is well beyond the acceptable level. At each minor fault, the VM checks whether the number of minor faults since the last adjustment exceeds adjust_count. If so, it forces an adjustment.


5 Experimental Evaluation

We now evaluate our VM implementation and heap size manager. We first compare the performance of the CRAMM VM with the original Linux VM. We then add the heap size manager to several collectors in Jikes RVM, and evaluate their performance under both static and dynamic real memory allocations. We also compare them with the JRockit [7] and HotSpot [19] JVMs under similar conditions. Finally, we run two concurrent instances of our adaptive collectors under memory pressure to see how they interact with each other.


5.1 Methodology Overview

We performed all measurements on a 1.70GHz Pentium 4 Linux machine with 512MB of RAM and 512MB of local swap space. The processor has 12KB I and 8KB D L1 caches and a 256KB unified L2 cache. We installed both the ``stock'' Linux kernel (version 2.4.20) and our CRAMM kernel. We ran each of our experiments six times in single-user mode, and always report the mean of the last five runs. In order to simulate memory pressure, we used a background process to pin a certain volume of pages in memory using mlock.

Application platform: We used Jikes RVM v2.4.1 [3] built for Linux x86 as our Java platform. We optimized the system images to the highest optimization level to avoid run-time compilation of those components. Jikes RVM uses an adaptive compilation system, which invokes optimization based on time-driven sampling. This makes executions non-deterministic. In order to get comparable deterministic executions, we took compilation logs from 7 runs of each benchmark using the adaptive system, and directed the system to compile methods according to the log from the run with the best performance. This is called the replay system. It is deterministic and highly similar to typical adaptive system runs.

Collectors: We evaluate five collectors from the MMTk memory management toolkit [9] in Jikes RVM: MS (mark-sweep), GenMS (generational mark-sweep), CopyMS (copying mark-sweep), SS (semi-space), and GenCopy (generational copying). All of these collectors have a separate non-copying region for large objects (2KB or more), collected with the Treadmill algorithm [6]. They also use separate non-copying regions for meta-data and immortal objects. We now describe the other regions each collector uses for ordinary small objects. MS is non-generational with a single MS region. GenMS is generational with a copying nursery and MS mature space. CopyMS is non-generational with two regions, both collected at every GC. New objects go into a copy region, while copy survivors go into an MS region. SS is non-generational with a single copying region. GenCopy is generational with copying nursery and mature space. Both generational collectors (GenMS and GenCopy) use Appel-style nursery sizing [4] (starts large and shrinks as mature space grows).

Benchmarks: For evaluating JVM performance, we ran all benchmarks from the SPECjvm98 suite (standard and widely used), plus those benchmarks from the DaCapo suite [10] (an emerging standard for JVM GC evaluation) that run under Jikes RVM, plus ipsixql (a publicly available XML database program) and pseudojbb (a variant of the standard, often-used SPECjbb server benchmark with a fixed workload (140,000 transactions) instead of fixed time limit). For evaluating general VM performance, we used the standard SPEC2000 suite.

Presented: Many results are similar, so to save space we present results only from some representative collectors and benchmarks. For collectors, we chose SS, MS, and GenMS to cover copying, non-copying, and generational variants. For benchmarks, we chose javac, jack, pseudojbb, ipsixql, jython, and pmd.


5.2 VM Performance

For the CRAMM VM to be practical, its baseline performance (i.e., while collecting useful histogram/working set size information) must be competitive when physical RAM is plentiful. We compare the performance of the CRAMM VM to that of the stock Linux kernel across our entire benchmark suite.4 For each benchmark, we use the input that makes it runs longer than 60 seconds.

Figure 5 summarizes the results, which are geometric means across all benchmarks: SPEC2000int, SPEC2000fp, and all the Java benchmarks (SPECjvm98, DaCapo, pseudojbb, and ipsixql) with five different garbage collectors. While the inactive list size adjustment mechanism effectively keeps the cost of collecting histogram data in the desired range (e.g., 0.59% for SPEC2Kint and 1.02% for SPEC2Kfp), the slowdown is generally about 1-2.5%. We believe this overhead is caused by CRAMM polluting the cache when handling minor faults as it processes page lists and AVL trees. This, in turn, leads to extra cache misses for the application. We verified that at the target minor fault overhead, CRAMM incurs enough minor faults to calculate the working set size accurately with respect to our 5% page fault threshold.

Figure 5: Virtual memory overhead (% increase in execution time) without paging, across all benchmark suites and garbage collectors.
\includegraphics[width=3in]{vm-overhead.eps}

CRAMM's performance is generally somewhat poorer on the Java benchmarks, where it must spend more time handling minor faults caused by the dramatic working set changes between the mutator and collector phases of GCed applications. However, the fault handling overhead remains in our target range. Overall, CRAMM collects the necessary information at very low overhead in most cases, and its performance is competitive to that of the stock kernel.


5.3 Static Memory Allocation

To test our adaptive mechanism, we run the benchmarks over a range of requested heap sizes with a fixed memory allocation. We select memory allocations that reveal the effects of large heaps in small allocations and small heaps in large allocations. In particular, we try to evaluate the ability of our mechanism to grow and shrink the heap. We run the non-adaptive collectors (which simply use the requested heap size) on both the stock and CRAMM kernels, and the adaptive collectors on the CRAMM kernel, and compare performance.

Figure 6: Static Memory Allocation: MarkSweep
\includegraphics[width=2.1in]{graphs/FastAdaptiveMarkSweep._213_javac.CA.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveMarkSweep._228_jack.CA.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveMarkSweep.ipsixql.CA.eps}
\includegraphics[width=2.1in]{graphs/FastAdaptiveMarkSweep.pseudojbb.CA.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveMarkSweep.jython.CA.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveMarkSweep.pmd.CA.eps}

Figure 7: Static Memory Allocation: GenMS
\includegraphics[width=2.1in]{graphs/FastAdaptiveGenMS._213_javac.CA.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveGenMS._228_jack.CA.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveGenMS.ipsixql.CA.eps}
\includegraphics[width=2.1in]{graphs/FastAdaptiveGenMS.pseudojbb.CA.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveGenMS.jython.CA.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveGenMS.pmd.CA.eps}

Figure 6 and Figure 7 show execution time for benchmarks using MS and GenMS collectors respectively with a static memory allocation. For almost every combination of benchmark and requested heap size, our adaptive collector chooses a heap size that is nearly optimal. It reduces total execution time dramatically, or performs at least as well as the non-adaptive collector. At the leftmost side of each curve, the non-adaptive collector runs at a heap size that does not consume the entire allocation, thus under-utilizing available memory, collecting too frequently and inducing high GC overhead. The adaptive collector grows the heap size to reduce the number of collections without incurring paging. At the smallest requested heap sizes, this adjustment reduces execution time by as much as 85%.

At slightly larger requested heap sizes, the non-adaptive collector performs fewer collections, better utilizing available memory. One can see that there is an ideal heap size for the given benchmark and allocation. At that heap size, the non-adaptive collector performs well--but the adaptive collector often matches it, and is never very much worse. The maximum slowdown we observed is 11% across all the benchmarks. (Our working set size calculation uses a page fault threshold of $t = 5\%$, so we are allowing a trivial amount of paging--while reducing the working set size substantially.)

Once the requested heap size goes slightly beyond the ideal, non-adaptive collector performance drops dramatically. The working set size is just slightly too large for the allocation, which induces enough paging to slow execution by as much as a factor of 5 to 10. In contrast, our adaptive collector shrinks the heap so that the allocation completely captures the working set size. By performing slightly more frequent collections, the adaptive collector consumes a modest amount of CPU time to avoid a lot of paging, thus reducing elapsed time by as much as 90%. When the requested heap size becomes even larger, the performance of our adaptive collector remains the same. However, the execution time of the non-adaptive collector decreases gradually. This is because it does fewer collections, and it is collections that cause most of the paging.

Interestingly, when we disable adaptivity, the CRAMM VM exhibits worse paging performance than the stock Linux VM. LRU-based eviction algorithm turns out to be a poor fit for garbage collection's memory reference behavior. Collectors typically exhibit loop-like behavior when tracing live objects, and LRU is notoriously bad in handling large loops. The Linux VM instead uses an eviction algorithm based on a combination of CLOCK and a linear scan over the program's address space, which happens to work better in this case.


5.4 Dynamic Memory Allocation

The results given so far show that our adaptive mechanism selects a good heap size when presented with an unchanging memory allocation. We now examine how CRAMM performs when the memory allocation changes dynamically. To simulate dynamic memory pressure, we use a background process that repeatedly consumes and releases memory. Specifically, it consists of an infinite loop, in which it sleeps for 25 seconds, mmap's 50MB memory, mlock's it for 50 seconds, and then unlocks and unmaps the memory. We also modify how we invoke benchmarks so that they run long enough (we give pseudojbb a large transaction number, and iterate javac 20 times).


Table 1: Dynamic Memory Allocation: Performance of Adaptive vs. Non-Adaptive Collectors
Benchmark Collector Enough Memory Adaptive Collector Non-Adaptive Collector Adaptive
(Memory) (Heap Size) T(sec) MF T(sec) cpu MF T(sec) cpu MF Yes/No
pseudojbb SS (160M) 297.35 1136 339.91 99% 1451 501.62 65% 24382 0.678
(160M) MS (120M) 336.17 1136 386.88 98% 1179 928.49 36% 47941 0.417
GenMS (120M) 296.67 1136 302.53 98% 1613 720.11 48% 39944 0.420
javac SS (150M) 237.51 1129 259.35 94% 1596 455.38 68% 24047 0.569
(140M) MS (90M) 261.63 1129 288.09 95% 1789 555.92 47% 25954 0.518
GenMS (90M) 249.02 1129 263.69 95% 2073 541.87 50% 33712 0.487


Figure 8: Dynamic Memory Allocation (pseudojbb): Heap Adjustment and Throughput
\includegraphics[width=2.1in]{graphs/FastAdaptiveSemiSpace.pseudojbb.heap.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveMarkSweep.pseudojbb.heap.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveGenMS.pseudojbb.heap.eps}
\includegraphics[width=2.1in]{graphs/FastAdaptiveSemiSpace.pseudojbb.throughput.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveMarkSweep.pseudojbb.throughput.eps} \includegraphics[width=2.1in]{graphs/FastAdaptiveGenMS.pseudojbb.throughput.eps}

Table 1 summarizes the performance of both non-adaptive and adaptive collectors under this dynamic memory pressure. The first column gives the benchmarks and their initial memory allocation. The second column gives the collectors and their requested heap sizes respectively. We set the requested heap size so that the benchmark will run gracefully in the initial memory allocation. We present the total elapsed time (T), CPU utilization (cpu), and number of major faults (MF) for each collector. We compare them against the base case, i.e., running the benchmark at the requested heap size with sufficient memory. The last column shows adaptive execution time relative to non-adaptive. We see that for each collector the adaptive mechanism adjusts the heap size in response to memory pressure, nearly eliminating paging. The adaptive collectors show very high CPU utilization and dramatically reduced execution time.

Figure 8 illustrates how our adaptive collectors change the heap size while running pseudojbb under dynamic memory pressure. The graphs in the first row demonstrate how available memory changes over time, and the corresponding heap size chosen by each adaptive collector. We see that as available memory drops, the adaptive collectors quickly shrink the heap to avoid paging. However, if our adaptive collectors do not periodically polling for current available memory and collect before the heap is filled up, the CPU utilization falls below 80% and can be as low as 53%. Likewise, they grow the heap responsively when there is more available memory. One can also see that the difference between the maximum and minimum heap size is approximately the amount of memory change divided by heap utilization $u$, consistent with our working set size model presented in Section 3.1.

Figure 9: Running Two Instances of Adaptive Collectors.
\includegraphics[width=2.1in]{graphs/twoJikesRVM-1.eps} \includegraphics[width=2.1in]{graphs/twoJikesRVM-2.eps} \includegraphics[width=2.1in]{graphs/twoJikesRVM-3.eps}

We also compare the throughput of the adaptive and non-adaptive collectors (the second row in Figure 8), by printing out the number of transactions finished as time elapses for pseudojbb. These curves show that memory pressure has much less impact on throughput when running under our adaptive collectors. It causes only a small disturbance and only for a short period of time. The total execution time of our adaptive collectors is a little longer than that of the base case, simply because they run at a much smaller heap size (and thus collect more often) when there is less memory. The non-adaptive collectors experience significant paging slowdown when under memory pressure.

As previously mentioned, JRockit and HotSpot do not adjust heap size well in response to changing memory allocation. Figure 10 compares the throughput of our adaptive collectors with that of JRockit and HotSpot. We carefully choose the initial memory allocation so that the background process imposes the same amount of relative memory pressure as for our adaptive collectors. However, being an experimental platform, Jikes RVM's compiler does not produce as efficient code as these commercial JVMs. We thus normalize the time for each of them to the total execution time that each JVM takes to run when given ample physical memory. The results show that both JRockit and HotSpot experience a large relative performance loss. The flat regions on their throughput curves indicate that they make barely any progress when available memory suddenly shrinks to less than their working set. Meanwhile, our adaptive collector changes the heap size to fit in available memory, maintaining high performance.

Figure 10: Throughput under dynamic memory pressure, versus JRockit and HotSpot.
\includegraphics[width=3in]{graphs/throughput.normalized.eps}

Finally, we examine how our adaptive collectors interact with each other. We started two instances using adaptive collectors with a certain memory allocation (220MB), and let them adjust their heap sizes independently. We explored several combinations of collector and benchmark: the same collector and benchmark, the same collector and different benchmarks, and different collectors with different benchmarks. The experiments show that, for all these combinations, our adaptive collectors keep CPU utilization at least 91%. Figure 9 shows the amount of available memory observed by each collector and their adapted heap size over time. We see that, after bouncing around a little, our adaptive collectors tend to converge to heap sizes that give each job a fair share of available memory, even though each works independently. The curves of GenMS in the third graph show how filtering out small nursery collections helps to stabilize heap size. This experiment still focus on how each JVM adapts to dynamic memory allocation. Although our mechanism effectively prevents each of them from paging, the memory allocation to each JVM may still be unfair. We leave to future work how best to divide memory among multiple competing JVMs.


6 Related Work


Table 2: A comparison of approaches to dynamic heap sizing.
Grows Shrinks Static Dynamic Collector Needs OS
Heap Heap Allocation Allocation Neutral Support Responds to
Alonso et al.[2] $\surd$ $\surd$ $\surd$ $\surd$ memory allocation
Brecht et al.[11] $\surd$ $\surd$ pre-defined rules
Cooper et al.[14] $\surd$ $\surd$ user supplied target
BC [18] $\surd$ $\surd$ $\surd$ $\surd$ page swapping
JRockit [7] $\surd$ $\surd$ $\surd$ $\surd$ throughput or pause time
HotSpot [19] $\surd$ $\surd$ $\surd$ throughput and pause time
MMTk [9] $\surd$ $\surd$ $\surd$ $\surd$ live ratio and GC load
CRAMM/AHS [28] $\surd$ $\surd$ $\surd$ $\surd$ $\surd$ $\surd$ memory allocation


We now discuss the work most closely related to CRAMM, first discussing work related to the CRAMM VM and then addressing GC-based approaches to sizing the heap.

6.1 Virtual Memory

The CRAMM VM computes stack distances, which were originally designed for trace analysis. Mattson et al. introduced a one-pass algorithm, based on stack distances, that analyzes a reference trace and produces cache misses for caches of any size [22]. This algorithm was later adapted by Kim and Hsu to handle highly-associative caches [21]. However, these algorithms compute a stack distance in linear time, making them too slow to use inside a kernel. Subsequent work on analyzing reference traces used more advanced dictionary data structures [1,8,17,23,26]. These algorithms calculate a stack distance in logarithmic time, but do not maintain underlying referenced blocks in order. This order is unnecessary for trace processing but crucial for page eviction decisions. The CRAMM VM maintains pages in a list that preserves potential eviction order, and uses a separate AVL tree to calculate a stack distance in logarithmic time.

Zhou et al. present a VM system that also tracks LRU reference curves inside the kernel [29]. They use Kim and Hsu's linear-time algorithm to maintain LRU order and calculate stack distances. To achieve reasonable efficiency, this algorithm requires the use of large group sizes (e.g., 1024 pages) that significantly degrade accuracy. They also use a static division between the active and inactive lists, yielding an overhead of 7 to 10%. The CRAMM VM computes the stack distance in logarithmic time, and can track reference histograms at arbitrary granularities. Furthermore, its inactive list size adjustment algorithm allows it to collect information accurately from the tail of miss curves while limiting reference histogram overhead to 1%.

6.2 Garbage Collection

Researchers have proposed a number of heap sizing approaches for garbage collection; Table 2 provides a summary. The closest work to CRAMM is by Alonso and Appel, who also exploit VM system information to adjust the heap size [2]. Their collector periodically queries the VM for the current amount of available memory and adjusts the heap size in response. CRAMM differs from this work in several key respects. While their approach shrinks the heap when memory pressure is high, it does not expand and thus reduce GC frequency when pressure is low. It also relies on standard interfaces to the VM system that provide a coarse and often inaccurate estimate of memory pressure. The CRAMM VM captures detailed reference information and provides reliable values.

Brecht et al. adapt Alonso and Appel's approach to control heap growth via ad hoc rules for two given static memory sizes [11]. Cooper et al. dynamically adjust the heap size of an Appel-style collector according to a user-supplied memory usage target [14]. If the target matches the amount of free memory, their approach adjusts the heap to make full use of it. However, none of these approaches can adjust to dynamic memory allocations. CRAMM automatically identifies an optimal heap size using data from the VM. Furthermore, the CRAMM model captures the relationship between working set size and heap size, making its approach more general and robust.

Our research group previously presented the bookmarking collector (BC), a garbage-collection algorithm that guides a lightly modified VM system to evict pages that do not contain live objects and installs ``bookmarks'' in pages in response to eviction notifications [18]. These bookmarks allow BC to collect the heap without touching already evicted pages, which CRAMM must. One shortcoming of BC is that it currently cannot grow the heap because it responds only to page eviction notifications. CRAMM both shrinks and grows the heap to fit, and can be applied to a wide range of existing garbage collection algorithms.


7 Conclusion

We present CRAMM, a new system designed to support garbage-collected applications. CRAMM combines a new virtual memory system with a garbage-collector-neutral, analytic heap sizing model to dynamically adjust heap sizes. In exchange for modest overhead (around 1-2.5% on average), CRAMM improves performance dramatically by making full use of memory without incurring paging. CRAMM allows garbage-collected applications to run with a nearly-optimal heap size in the absence of memory pressure, and adapts quickly to dynamic memory pressure changes, avoiding paging while providing high CPU utilization.

8 Acknowledgements

This material is based upon work supported by the National Science Foundation under grant number CCR-0085792 and CAREER Award CNS-0347339. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the NSF.

Bibliography

1
G. Almasi, C. Cascaval, and D. A. Padua.
Calculating stack distances efficiently.
In ACM SIGPLAN Workshop on Memory System Performance, pages 37-43, Berlin, Germany, Oct. 2002.

2
R. Alonso and A. W. Appel.
An advisor for flexible working sets.
In Proceedings of the 1990 SIGMETRICS Conference on Measurement and Modeling of Computer Systems, pages 153-162, Boulder, CO, May 1990.

3
B. Alpern, C. R. Attanasio, J. J. Barton, M. G. Burke, P. Cheng, J.-D. Choi, A. Cocchi, S. J. Fink, D. Grove, M. Hind, S. F. Hummel, D. Lieber, V. Litvinov, M. F. Mergen, T. Ngo, V. Sarkar, M. J. Serrano, J. C. Shepherd, S. E. Smith, V. C. Sreedhar, H. Srinivasan, and J. Whaley.
The Jalepeño virtual machine.
IBM Systems Journal, 39(1):211-238, Feb. 2000.

4
A. Appel.
Simple generational garbage collection and fast allocation.
Software: Practice and Experience, 19(2):171-183, Feb. 1989.

5
O. Babaoglu and D. Ferrari.
Two-level replacement decisions in paging stores.
IEEE Transactions on Computers, C-32(12):1151-1159, Dec. 1983.

6
H. G. Baker.
The Treadmill: Real-time garbage collection without motion sickness.
ACM SIGPLAN Notices, 27(3):66-70, March 1992.

7
BEA WebLogic.
Technical white paper JRockit: Java for the enterprise.
https://www.bea.com/content/news_events /white_papers/BEA_JRockit_wp.pdf.

8
B. T. Bennett and V. J. Kruskal.
LRU stack processing.
IBM Journal of R & D, 19(4):353-357, 1975.

9
S. M. Blackburn, P. Cheng, and K. S. McKinley.
Oil and Water? High Performance Garbage Collection in Java with MMTk.
In 26th International Conference on Software Engineering, pages 137-146, May 2004.

10
S. M. Blackburn, R. Garner, C. Hoffman, A. M. Khan, K. S. McKinley, R. Bentzur, A. Diwan, D. Feinberg, D. Frampton, S. Z. Guyer, M. Hirzel, A. Hosking, M. Jump, H. Lee, J. E. B. Moss, A. Phansalkar, D. Stefanovic, T. VanDrunen, D. von Dincklage, and B. Wiedermann.
The DaCapo benchmarks: Java benchmarking development and analysis.
In OOPSLA '06: Proceedings of the 21st annual ACM SIGPLAN conference on Object-Oriented Programing, Systems, Languages, and Applications, New York, NY, USA, Oct. 2006. ACM Press.

11
T. Brecht, E. Arjomandi, C. Li, and H. Pham.
Controlling garbage collection and heap growth to reduce the execution time of Java applications.
In Proceedings of the 2001 ACM SIGPLAN Conference on Object-Oriented Programming, Systems, Languages & Applications, pages 353-366, Tampa, FL, June 2001.

12
R. W. Carr and J. L. Henessey.
WSClock - a simple and effective algorithm for virtual memory management.
In Proceedings of the Eighth ACM Symposium on Operating Systems Principles (SOSP), pages 87-95, Dec. 1981.

13
W. W. Chu and H. Opderbeck.
The page fault frequency replacement algorithm.
In AFIPS Conference Proceedings, volume 41(1), pages 597-609, Montvale, NJ, 1972. AFIPS Press.

14
E. Cooper, S. Nettles, and I. Subramanian.
Improving the performance of SML garbage collection using application-specific virtual memory management.
In Conference Record of the 1992 ACM Symposium on Lisp and Functional Programming, pages 43-52, San Francisco, CA, June 1992.

15
P. J. Denning.
The working set model for program behavior.
In Proceedings of the ACM Symposium on Operating System Principles, pages 15.1-15.12, Jan. 1967.

16
P. J. Denning.
Working sets past and present.
IEEE Transactions on Software Engineering, SE-6(1):64-84, Jan. 1980.

17
C. Ding and Y. Zhong.
Predicting whole-program locality through reuse distance analysis.
In ACM SIGPLAN 2003 Conference on Programming Language Design and Implementation, pages 245-257, San Diego, CA, June 2003.

18
M. Hertz, Y. Feng, and E. D. Berger.
Garbage collection without paging.
In Proceedings of the 2005 ACM SIGPLAN Conference on Programming Language Design and Implementaton, pages 143-153, Chicago, IL, June 2005.

19
JavaSoft.
J2SE 1.5.0 documentation: Garbage collector ergonomics.
https://java.sun.com/j2se/1.5.0/docs/guide/vm/ gc-ergonomics.html.

20
S. F. Kaplan, L. A. McGeoch, and M. F. Cole.
Adaptive caching for demand prepaging.
In Proceedings of the 2002 International Symposium on Memory Management, pages 114-126, June 2002.

21
Y. H. Kim, M. D. Hill, and D. A. Wood.
Implementing stack simulation for highly-associative memories.
In Proceedings of the 1991 SIGMETRICS Conference on Measurement and Modeling of Computer Systems, pages 212-213, San Diego, CA, 1991.

22
R. L. Mattson, J. Gecsei, D. R. Slutz, and I. L. Traiger.
Evaluation techniques for storage hierarchies.
IBM Systems Journal, 9(2):78-117, 1970.

23
F. Olken.
Efficient methods for calculating the success function of fixed space replacement policies.
Technical Report LBL-12370, Lawrence Berkeley Laboratory, 1981.

24
R. H. Patterson, G. A. Gibson, E. Ginting, D. Stodolsky, and J. Zelenka.
Informed prefetching and caching.
In Proceedings of the Fifteenth ACM Symposium on Operating Systems Principles, pages 79-95, New York, NY, USA, 1995. ACM Press.

25
Y. Smaragdakis, S. F. Kaplan, and P. R. Wilson.
The EELRU adaptive replacement algorithm.
Performance Evaluation, 53(2):93-123, July 2003.

26
R. A. Sugumar and S. G. Abraham.
Efficient simulation of caches under optimal replacement with applications to miss characterization.
In Measurement and Modeling of Computer Systems, pages 24-35, Santa Clara, CA, 1993.

27
P. R. Wilson, S. F. Kaplan, and Y. Smaragdakis.
The case for compressed caching in virtual memory systems.
In Proceedings of The 1999 USENIX Annual Technical Conference, pages 101-116, Monterey, California, June 1999. USENIX Association.

28
T. Yang, M. Hertz, E. D. Berger, S. F. Kaplan, and J. E. B. Moss.
Automatic heap sizing: Taking real memory into account.
In Proceedings of the ACM SIGPLAN International Symposium on Memory Management, pages 61-72, Vancouver, Canada, Oct. 2004.

29
P. Zhou, V. Pandy, J. Sundaresan, A. Raghuraman, Y. Zhou, and S. Kumar.
Dynamic tracking of page miss ratio curves for memory management.
In Proceedings of the 11th International Conference on Architectural Support for Programming Languages and Operating Systems, pages 177-188, Boston, MA, Oct. 2004.

About this document ...

CRAMM: Virtual Memory Support for Garbage-Collected Applications

This document was generated using the LaTeX2HTML translator Version 2002 (1.62)

Copyright © 1993, 1994, 1995, 1996, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999, Ross Moore, Mathematics Department, Macquarie University, Sydney.

The command line arguments were:
latex2html -split 0 -show_section_numbers -local_icons autoheapsizing.tex

The translation was initiated by Ting Yang on 2006-09-05


Footnotes

... (VM1
We use VM to denote virtual memory throughout this paper.
... allocation,2
The main memory allocation is not the same as the resident set size. The latter is the amount of main memory currently consumed by a process, while the former is the amount of main memory that the VM system is willing to let the process consume before evicting its pages.
....3
Notice that we refer to the histogram as an LRU reference histogram, but that our page lists are not in true LRU order, and so the histogram is really a SegQ reference histogram. Also, note that only references to the inactive and evicted lists are applicable here, since references to active pages occur without kernel intervention.
... suite.4
We could not compile and run some SPEC2000 Fortran programs, so we omit some of the FP benchmarks.

next_inactive up previous
Ting Yang 2006-09-05