John R. Ellis, Kai Li, and Andrew Appel. Real-time concurrent collection on stock multiprocessors. Technical Report 25, DEC Systems Research Center, Palo Alto, CA, February 1988.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... ML [11, 12] and for the Java programming language [14, 13] Baker invented an incremental copying collector [5] which was implemented in hardware on Lisp machines [25] Ellis, Li, and Appel implemented this idea on stock hardware with a virtual memory based barrier, and added true concurrency [23]. However, most of these have little relevance to the present work beyond the fact they involve garbage collection and concurrency. Two other collection approaches are similar to the current work in that they use log based write barriers whose output is processed by a concurrent thread devoted to ....

John R. Ellis; Kai Li; and Andrew W. Appel. Realtime concurrent collection on stock multiprocessors. Technical Report 25, Digital Equipment Corporation Systems Research Center, February 1988.

....and Neville Harris. Garbage collection in an object oriented, distributed, persistent environment. In Jul and Juul [Jul and Juul1990] Ellis and Detlefs, 1993] John R. Ellis and David L. Detlefs. Safe, efficient garbage collection for C . Technical report, Xerox PARC, Palo Alto, CA, 1993. [Ellis and Olson, 1988] C. S. Ellis and T. J. Olson. Algorithms for parallel memory allocation. International Journal of Parallel Programming, 17(4) 303 345, 1988. Ellis and Stroustrup, 1990] Margaret A. Ellis and Bjarne Stroustrup. The Annotated C ....

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report DEC--SRC--TR--25, DEC Systems Research Center, Palo Alto, CA, February 1988.

....algorithm. Halstead [6] describes a multiprocessor GC for Multilisp. Each processor has its own local heap, and they use lock bits for moving and updating forwarding pointers. Load balancing is done statically rather than dynamically. Many collectors operate concurrently with mutator activity [9, 4, 5, 8]. This kind of concurrency is orthogonal to the style of parallel collection we describe in this paper. A collector might combine both: some concurrent collectors have stop world phases that might be performed in parallel, and collectors with concurrent GC threads might use several such threads ....

John R. Ellis; Kai Li; and Andrew W. Appel. Real-time Concurrent Collection on Stock Multiprocessors. Technical Report 25, Digital Equipment Corporation Systems Research Center, February 1988.

.... was initiated by Steele and Dijkstra, et al. 27, 28, 8] and continued in a series of papers [9, 14, 3, 4, 20, 21] culminating in the Doligez Leroy Gonthier (DLG) collector [11, 10] The advantage of an on the y collector over a parallel collector and other types of concurrent collectors [1, 13, 24], is that it avoids the operation of stopping all the program threads. Such an operation can be costly. Usually, program threads cannot be stopped at any point; thus, there is a non negligible wait until the last (of many) threads reaches a safe point where it may stop. The drawback of on the y ....

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report DEC-SRC-TR-25, DEC Systems Research Center, Palo Alto, CA, February 1988.

.... was initiated by Steele and Dijkstra, et al. 30, 9] and continued in a series of papers [10, 18, 5, 6, 23, 24] culminating in the Doligez Leroy Gonthier (DLG) collector [13, 12] The advantage of an on the fly collector over a parallel collector and other types of concurrent collectors [3, 16, 28, 7], is that it avoids the operation of stopping all the program threads simultaneously. Usually, program threads cannot be stopped at any point; thus, there is a non negligible wait until the last (of many) threads reaches a safe point where it may stop. On the fly collectors also satisfy the ....

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report DEC-SRC-TR-25, DEC Systems Research Center, Palo Alto, CA, February 1988.

.... occasionally wait for garbage collection during access to previously allocated objects is 500 sec [6] for applications that are somewhat restricted in their use of dynamic memory.More general garbage collection systems promise looser bounds, ranging from several to several hundred milliseconds [7, 8]. These delays are too large to be tolerated by manyreal time applications. Furthermore, available garbage collection systems offer no guarantees of minimum time separation between consecutive events that require abnormal delays in program execution. Yetanother shortcoming of manyexisting garbage ....

....sizeisnot an exact multiple of the subregion size, round the sizeup. 3. add 1 because of ProbeOffset al..ignments #define NumSubRegions(nw) nw) 1) SubRegionSize 1) SubRegionSize) 1) int ProbeOffset = 16; byte offset at which subregions are aligned. static int nxtprobes[8] = ProbeOffset is changed for each pass of the garbage collector. 3, 5, 6, 7, 2, 0, 1, 4, #define nextProbe(oldprobe) nxtprobes[oldprobe BytesPerWord] BytesPerWord) The final phase of garbage collection is to postprocess control blocks, carving each of the controlled regions into ....

J. R. Ellis, K. Li and A. W.Appel, "Real-time Concurrent Collection on Stock Multiprocessors", ACMSIGPLAN Notices Conference on Programming Language Design and Implementation,June 1988.

....by first copying 2 the referenced object into to space. Proposed System Architecture Testing for memory operations that require special handling in software significantly impairs system performance [2] By using hardware, the overhead of normal memory operations can be significantly reduced [6]. The hardware we propose monitors each memory operation and stalls read and write requests that refer to memory locations that have not yet been scanned or copied. In figure 4, the only memory operations that require special handling are operations that refer to memory found between scanned and ....

J. R. Ellis, K. Li and A. W. Appel, Real-time Concurrent Collection on Stock Multiprocessors, SRC Report 25, Palo Alto, CA, Feb 1988.

....2. Also, no provision is made for invoking destructors of collected objects, violating property 5. In previous work [6] I extended Bartlett s work by modifying a compiler to produce the pointer location information, and allowing concurrent mutator activity in the style of Appel, Ellis, and Li [19]; however, a collection strategy requiring a modified compiler violates property 1. Ferreira [11] presents a C collector framework that allows explicitly managed storage and calls destructors. Collectibility is determined on a per class basis; programmers must invoke a macro in the definition ....

John R. Ellis; Kai Li; and Andrew W. Appel. Real-time Concurrent Collection on Stock Multiprocessors. Research Report 25, Digital Equipment Corporation Systems Research Center, February 1988.

....response during those collections. I chose instead to investigate concurrent collection, where the mutator can continue working during collection. In particular, I use the virtual memory protection technique of Ellis, Li, and Appel to synchronize mutator and collector access to heap pages [EA88]. 4 Concurrent, Atomic Collection 4.1 Overview At a very high level, my algorithm can be seen as combining Kolodner s algorithm with the concurrency control technique of Ellis, Li, and Appel. However, on more detailed inspection, the semantics of transactions complicate things a good deal. My ....

Kai; Ellis, John R.; Li and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report 25, DEC Systems Research Center, February 1988.

....literature. O Toole [19] replicates the heap so that the application can access one replica while the collector performs a standard stop and copy on the other replica. New cells and updates in the application s replica are logged. Kolodner s algorithm [14] is based on the Ellis Appel Li collector [5], but so that all changes done to the heap by the collector are logged: for example copying a single cell results in a log record containing the old and new addresses of the cell and the value written over by the forwarding pointer. Most persistent garbage collectors based on partitions are ....

J. R. Ellis, K. Li, and A. W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report DEC{SRC{TR{25, DEC Systems Research Center, Palo Alto, CA, Feb. 1988.

.... was initiated by Steele and Dijkstra, et al. 27, 28, 8] and continued in a series of papers [9, 14, 3, 4, 20, 21] 1 culminating in the Doligez Leroy Gonthier (DLG) collector [11, 10] The advantage of an on the y collector over a parallel collector and other types of concurrent collectors [1, 13, 24], is that it avoids the operation of stopping all the program threads. Such an operation can be costly. Usually, program threads cannot be stopped at any point; thus, there is a non negligible wait until the last (of many) threads reaches a safe point where it may stop. The drawback of on the y ....

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report DEC-SRC-TR-25, DEC Systems Research Center, Palo Alto, CA, February 1988.

....(with garbage collection) and Icon to our dlx environment. In spite of the current absence of experimental data to support our claims, we expect that, on the average, performance of our system is comparable or favorable to the stock hardware assisted garbage collection algorithm described in [5]. In comparison with that algorithm, ours is more general in that it supports garbage collection of arbitrarily large records, stacks, and overlapping slices. Finally,our algorithm offers much lower worst case response times both for allocation of new memory and access to previously allocated ....

J. R. Ellis, K. Li and A. W.Appel, Real-time Concurrent Collection on Stock Multiprocessors, SRC Report 25, Palo Alto, CA, Feb 1988.

....node faults, causing another replica to be created and the writer s copy to be changed back to read only. Along with copy on write and distributed virtual memory, other operating system functions are being overloaded on virtual memory protection bits as well: these include garbage collection [Ellis et al. 88] checkpointing [Li et al. 90] recoverable virtual memory [Eppinger 89] and transaction locking [Radin 82] Because these functions often are implemented at the run time level, their implementations are simplified by user level handling of page faults and efficient modification of TLB or page ....

J. R. Ellis, K. Li, and A. W. Appel. Real-time concurrent collection on stock multiprocessors. In Proceedings of the ACM SIGPLAN '88 Conference on Programming Language Design and Implementation, pages 11--20, June 1988.

....very short. In one test of alternatives, a mostly copying collector was implemented for Modula2 . A mostly copying collector [1] behaves much like a copying collector, but does not relocate objects that the program may hold pointers to. The mostly copying collector used VM synchronization [6] to allow it to run concurrently. With VM synchronization, the overhead for language primitives can approach zero. However, the experimental collector s working set was significantly larger than the reference counting collector s, sometimes provoking thrashing when reference counting would not. ....

John R. Ellis, Kai Li, and Andrew W. Appel, Real-Time Concurrent Collection on Stock Multiprocessors. Research Report 25, Digital Equipment Corporation Systems Research Center, February 1988.

....involved in intercepting memory requests for inconsistent objects has been the primary stumbling block in developing garbage collectors that satisfy tight real time constraints. Compilergenerated range checking of each memory reference more than doubles the cost of executing typical applications [4, 5, 11]. Performing such checks in microcode as in certain CISC based Lisp machines is less costly, but still bears an overhead cost of roughly 30 [5, 19] and does not apply to modern RISC architectures. Using virtual memory protection to generate page faults for inconsistent objects [5] produces ....

....tight real time constraints. Compilergenerated range checking of each memory reference more than doubles the cost of executing typical applications [4, 5, 11] Performing such checks in microcode as in certain CISC based Lisp machines is less costly, but still bears an overhead cost of roughly 30 [5, 19], and does not apply to modern RISC architectures. Using virtual memory protection to generate page faults for inconsistent objects [5] produces better results, but the high cost of page faults and the coarseness of page boundaries result in worst case memory latencies in the tenths of seconds. By ....

[Article contains additional citation context not shown here]

J. R. Ellis, K. Li, and A. W. Appel, Realtime concurrent collection on stock multiprocessors, ACM SIGPLAN '88 Conference on Programming Language Design and Implementation, 1988, pp. 11--20.

....up after multiple concurrent mutators. This algorithm, as Pixley notes, behaves incorrectly in the presence of certain race conditions, which Pixley explicitly assumes do not occur. Our algorithm introduces multiple versions to avoid precisely these kinds of problems. The Ellis, Li, and Appel [6] describe the design and implementation of a multi mutator, single collector copying garbage collector. This algorithm is blocking, since processes synchronize via locks, and flipping the from and to spaces requires halting the mutators and inspecting and altering their registers. 7 Conclusions ....

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report 25, Digital Systems Research Center, 130 Lytton Avenue, Palo Alto, CA 94301, February 1988.

....large. This overhead may be exacerbated by an increased address space. Coupled with this is the cost of allocation and freeing of secondary storage. A further consideration is distributed garbage collection. Garbage collection mechanisms that are time independent of the size of the address space [Kol92, ELA88] can be complex. The complexity is increased when the address space is distributed over several nodes. A review of distributed collectors in [AMR92] highlights the problems and indicates that very few fault tolerant collectors exist. distributed stability and recovery. The issues of distributed ....

Ellis, J.R., Li, K. & Appel, A.W. "Real-time Concurrent Collection on Stock Multiprocessors". Technical Report 25, Digital Equipment Corporation (1988).

....mechanism to form and disband a distributed group is quite complex and will not scale well. We avoid this problem considering only a single site when by setting up groups of regions [45] Previous work on garbage collection in DSM (without support for persistence) deals either with multiprocessors [16, 15] or with a small scale systems [27] These algorithms make strong coherence assumptions, therefore they unsuitable for a large scale network. In contrast, Larchant makesno coherence assumption whatsoever [18] C.4 Application behaviour and heuristics evaluation Our TC.3 studies application ....

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report 25, Digital Research Center, Palo Alto, CA (USA), February 1988.

....Khayri Ali. Both the strategy, its implementation, and performance has been described in the thesis by Galal [5] this section will only be a brief overview. 1 Except during unification. This order need not be preserved. 120 The Implementation Parallel garbage collectors have a long history [27, 6, 20, 60]. The one that come closest to our implementation is the implementation by Imai and Tick [30] They parallelized a stop and copy garbage collection scheme for shared memory multiprocessors. The system was implemented within a concurrent logic programming system called VPIM, a parallel KL1 ....

J. R. Ellis, K. Li, and A. W. Appel. Real-time concurrent collection on stock multiprocessors. In SIGPLAN' 88 Conference on Programming Language Design and Implementation, pages 11--20, June 1988. Also Digital SRC Research Report number 25.

....supports dynamic load balancing mechanism, that is any worker can garbage collect objects in any memory. This implemented by a global stack for maintaining garbage collection work that can be taken by any worker. Their calculated garbage collection speedups are expected. Ellis, li, and Appel [8] proposed also the design and the implementation of a concurrent copying garbage collector on a shared memory multiprocessors. One processor reclaims all the garbage, while others proceed with their normal computation. Heap area organized in pages. They using a memory protections facilities ....

....scheme. They used a shared memory multiprocessors simulator for implementation of their scheme. Heap area is represented by several generation and the garbage collection of each generation can take place in parallel. The preformance of their scheme evaluated by comparison with Ellis, li, and Appel [8] scheme. We are implemented the scheme proposed by Ali [1] which is a parallelization of a sequential stop and copy garbage collection scheme based on traversing active data in a depth first manner. The heap area is noncontiguous memory blocks with no extra space overhead. The non contiguous ....

J. R. Ellis, K. Li, and A. W. Appel. Real-time Concurrent Collection on Stock Multiprocessors. SIGPLAN' 88 Conference on Programming Language Design and Implementation, pages 11-20, June 1988. Also Digital SRC Research Report number 25.

....table. Our implementation instead uses the trail for pointing out references from the old to the new generation. Huelsbergen and Larus [103] developed a concurrent copying garbage collector for shared memory with two processes; collector and mutator. Ellis, 1.6. Related Work 23 Li and Appel [73] propose a similar design with several mutators and one collector. Rojemo [161] extended the collector by Ellis et al. for the hv; Gimachine [13] a parallel version of the G machine [14] Chapter 2 Reform Prolog: The Language and its Implementation Johan Bevemyr, Thomas Lindgren, Hakan ....

J. R. Ellis, K. Li, A. W. Appel, Real-time Concurrent Collection on Stock Multiprocessors, Tech. Rep. 25, Digital Systems Research Center, Palo Alto, 1988. --23

....a space just uses a memory address. Inter space references (Section 3.5) are easy to track, because they are passed explicitly in messages. A space can be composed of a small set of cooperating processes, possibly on different sites. This is the model used by many multiprocessor collectors [9, 10, 19]. As long as the set uses a single addressing scheme, and assuming no independent failures of one of the processes and reliable communication, this is very similar to the classical model. However, the costs are different because of message passing between the processes, of synchronization of the ....

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report 25, Digital Research Center, Palo Alto, CA (USA), February 1988.

....boundary so that the collector can find the beginning of the object. Although his system supports fast inline allocation, doing so is trivial since only two allocation pointers are required to handle the two metatypes his system supports. The page based incremental collector described by Ellis [12] also uses a boundary crossing table to enable the collector to find the beginning of an object that crosses page boundaries. This can be seen as a simple use of metatyping. ....

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Systems Research Center 25, Digital Equipment Corp., February 1988.

....is good enough for many C applications [Yip 91, Zorn 92] Both families of collectors use similar technology for satisfying our requirements. Low latency (short interruptions) is achieved using virtual memory synchronization to implement generational and concurrent collection [Shaw 87, Ellis 88] Interior pointers are handled in the Boehm collectors by ensuring that all objects on a page have the same size and by using a table to map page numbers to object sizes. The Bartlett collectors use a bitmap per page indicating the starting offsets of objects. Cross heap pointers from the ....

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Systems Research Center Report 25, Digital Equipment Corporation, 1988.

No context found.

John R. Ellis, Kai Li, and Andrew Appel. Real-time concurrent collection on stock multiprocessors. Technical Report 25, DEC Systems Research Center, Palo Alto, CA, February 1988.

No context found.

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report DEC--SRC--TR--25, DEC Systems Research Center, Palo Alto, CA, February 1988.

No context found.

John R. Ellis, Kai Li, and Andrew W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report DEC--SRC--TR--25, DEC Systems Research Center, Palo Alto, CA, February 1988.

No context found.

J. R. Ellis, K. Li, and A. W. Appel. Real-time concurrent collection on stock multiprocessors. Technical Report DEC{SRC{TR{25, DEC Systems Research Center, Palo Alto, CA, Feb. 1988.

No context found.

J. R. Ellis, K. Li, and A. W. Appel, Real-time Concurrent Collection on Stock Multiprocessors, In SIGPLAN'88 Conference on Programming Language Design and Implementation, June 1988. Also Digital SRC Research Report number 25

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC