L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Dept. of Computer Science, Princeton University, June 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....data distribution, where each node is assigned a two dimensional subgrid, has less communication overhead than its one dimensional counterpart. Unfortunately, two dimensional distributions are not efficiently supported in software distributed shared memory (SDSM) systems [14] In an SDSM [11, 3, 10, 7], twodimensional distributions can cause a large amount of excess communication between nodes, even when using relaxed memory consistencies such as eager write shared [3] lazy write shared [10, 7, 8] or a hybrid [1] The primary problem is that two dimensional distributions force nodes to share ....

.... efficiently supported in software distributed shared memory (SDSM) systems [14] In an SDSM [11, 3, 10, 7] twodimensional distributions can cause a large amount of excess communication between nodes, even when using relaxed memory consistencies such as eager write shared [3] lazy write shared [10, 7, 8], or a hybrid [1] The primary problem is that two dimensional distributions force nodes to share data in two dimensions, but the data is organized in either a row or column major layout. Consequently, nodes share both rows and columns, meaning that every element in a page must be write shared, ....

[Article contains additional citation context not shown here]

L. Iftode. Home-Based Shared Virtual Memory. PhD thesis, Princeton University, June 1998.

....as a runtime component, the GOS cannot rely on the programmer to provide the sharing information. Scientific applications exhibit diverse execution patterns. To execute these applications e#ciently in software DSM systems, many cache coherence protocols have been proposed. Home based protocols [19] assign a home node to each shared data object from which all copies are derived. It is widely believed that home based protocols are more scalable than homeless protocols [20] for the reason that the former has less memory consumption and can eliminate diff accumulation. The home in a home based ....

....data object from which all copies are derived. It is widely believed that home based protocols are more scalable than homeless protocols [20] for the reason that the former has less memory consumption and can eliminate diff accumulation. The home in a home based protocol can be either fixed [19] or mobile [21] There is also variation for the coherence operations, such as a multiple writer protocol, or a single writer protocol. The multiple writer protocol introduced in Munin [17] supports concurrent writes on di#erent copies using the di# technique. It may however incur heavy di# ....

L. Iftode, Home-based Shared Virtual Memory, Ph.D. thesis, Princeton University (August 1998).

....to implement the Java memory model. We believe that adaptive protocols are superior to non adaptive ones due to their adaptability to applications object access patterns. Many cache coherence protocols have been proposed in the field of software distributed shared memory. Home based protocols [18] assign a home node to each shared data object from which all copies are derived. It is widely believed that home based protocols are more scalable than homeless protocols [20] for the reason that the former kind has less memory consumption and eliminates di# accumulation. The home in a ....

....data object from which all copies are derived. It is widely believed that home based protocols are more scalable than homeless protocols [20] for the reason that the former kind has less memory consumption and eliminates di# accumulation. The home in a home based protocol can be either fixed [18] or migratable [9] There are also choices for the coherence operations, such as that between a multiple writer protocol or a single writer protocol. The multiple writer protocol introduced in Munin [8] supports concurrent writes on di#erent copies using the di# techniques. It may however incur ....

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Princeton University, August 1998.

....node that is responsible for maintaining and distributing the data stored on the page to other nodes of the system. In this section we describe the features of home based software DSM protocols that are important for our purposes a comprehensive discussion can be found in Iftode s dissertation [8]. Program operation on each node is divided into intervals delimited by synchronization operations, such as locks and barriers. All synchronization operations are classified as either an acquire or a release. Lock acquire and release operations are classified in the obvious manner. A barrier ....

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Princeton University, 1998.

....of multiple workstations to form a large shared address space. JIAJIA is implemented entirely as a user level library and currently runs on many mainstream UNIX platforms. Multiple writer technique is employed to alleviate false sharing. JIAJIA implements the scope consistency (ScC) memory model[7]. The unique characteristic of JIAJIA is the use of home concept not only for data information, but also for coherence information, such as write notices. Generally, home based software DSM systems[7] propagate data information and apply them at homes eagerly, but this information is fetched by ....

....is employed to alleviate false sharing. JIAJIA implements the scope consistency (ScC) memory model[7] The unique characteristic of JIAJIA is the use of home concept not only for data information, but also for coherence information, such as write notices. Generally, home based software DSM systems[7] propagate data information and apply them at homes eagerly, but this information is fetched by others lazily. On the other hand, these systems propagate coherence information and apply them either eagerly or lazily. In our lock based cache coherence protocol, coherence information is processed in ....

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Dept. of Computer Science, Princeton University, June 1998.

....in TreadMarks[15] a diff is not created for a modified page until a processor requests that diff in order to update its copy of the page. Each processor keeps diffs created by itself until garbage collection. When compared, home based protocols have the following advantages over homeless ones[12]. 1) Each access fault requires only one round trip to the home of the fault page. 2) Writes to home pages do not produce twins and # The work of this paper is supported partly by National Climbing Program of China and National Natural Science Foundation of China (Grant No. 69703002 and ....

....and twin pointer of the cache page. With the cache replacement scheme and the simplicity of data structure, JIAJIA is able to support shared memory that is larger than physical memory of one machine. 2. 2 Lock Based Cache Coherence Protocol JIAJIA implements the scope memory consistency (ScC)[12] model. Multiple writer technique is employed to reduce false sharing. In JIAJIA, the coherence of cache pages is maintained through write notices kept on the lock. In the protocol, each page has a fixed home and can be cached by a non home processor in one of three states: Invalid (INV) ....

[Article contains additional citation context not shown here]

L. Iftode, "Home-based Shared Virtual Memory", Ph. D. Thesis, Princeton University, Aug. 1998.

....for parallelism within a node. This enables the automatic use of hardware shared memory within a node for better eciency. There is a wealth of related work on ne grain parallelism and software distributed shared memory systems; two representative papers that cover these topics are [TTY99] and [Ift98] We cannot further cover research in these areas due to space limitations. Finally, as described in previous sections, HyFi is a combination of Shared Filaments [LFA98] and Distributed Filaments [LFA96] 6 Conclusion This paper has presented a library package called HyFi, which provides a ....

Liviu Iftode. Home-Based Shared Virtual Memory. PhD thesis, Princeton University, June 1998.

....reference is made, a page fault occurs, and the SDSM system implicitly resolves the reference by sending a message to the node that owns the page. Note that we use a multi threaded SDSM, so multiple page faults can be outstanding. Further information on SDSM systems can be found in, for example, [7]. A key issue involved in building an SDSM system is solving the memory coherence problem, which is typically done through a page consistency protocol ; in this paper, we use the write invalidate protocol [11] where each page has either a single writer or multiple readers. 2.2 SUIF Adapt ....

....also which locations within those pages are written. Then, the DFA describing page states would need to be modi ed to describe write shared behavior. For example, writing to a read shared page no longer causes an invalidation. The exact DFA modi cations depend on the write shared update strategy [7]. We will discuss three such strategies: eager, lazy, and home based. The easiest version of write shared to support would be home based, because it is conceptually the simplest. Using the DRSDs, SUIF Adapt could know which pages are sent to the home and run isolated experiments to determine the ....

L. Iftode. Home-Based Shared Virtual Memory. PhD thesis, Princeton University, June 1998.

....programs as opposed to more complicated message passing ones. When a non local reference is made, a page fault occurs, and the SDSM system implicitly resolves the reference by sending a message to the node that owns the page. Further information on SDSM systems can be found in, for example, [9]. A key issue involved in building an SDSM system is solving the memory coherence problem, which is typically done through a page consistency protocol ; in this paper, we use the write invalidate protocol [13] where each page has either a single writer or multiple readers. 2.2 SUIF Adapt ....

....also which locations within those pages are written. Then, the DFA describing page states would need to be modi ed to describe write shared behavior. For example, writing to a read shared page no longer causes an invalidation. The exact DFA modi cations depend on the write shared update strategy [9]. We will discuss three such strategies: eager, lazy, and home based. The easiest version of write shared to support would be 8 Jacobi Iteration with One Competing Process (Iteration 20) 0 20 40 60 80 100 2 nodes 4 nodes 8 nodes SUIF Adapt (with best redistribution) No ....

L. Iftode. Home-Based Shared Virtual Memory. PhD thesis, Princeton University, June 1998.

....programs as opposed to more complicated message passing ones. When a non local reference is made, a page fault occurs, and the SDSM system implicitly resolves the reference by sending a message to the node that owns the page. Further information on SDSM systems can be found in, for example, Ift98] A key issue involved in building an SDSM system is solving the memory coherence problem, which is typically done through a page consistency protocol ; in this paper, we use the write invalidate protocol [LH89] where each page has either a single writer or multiple readers. 2.2 Computational ....

Liviu Iftode. Home-Based Shared Virtual Memory. PhD thesis, Princeton University, June 1998. 15

....overhead than its one dimensional counterpart. Unfortunately, two dimensional distributions are not eciently supported in software distributed shared memory (DSM) systems. Such systems provide an attractive shared memory programming model on a distributed machine [LH89, CBZ91, KDCZ94, Ift98, LCD 97, CDLZ97, KT97] In a software DSM, two dimensional distributions can cause a large amount of excess communication This work was supported by National Science Foundation CAREER Grants CCR 9733063 and CCR 9876073. y Department of Computer Science, The University of Georgia, Athens, ....

....Georgia, Athens, GA 30602. Email:dkl cs.uga.edu z Department of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN 46556. Email: vin cse.nd. edu 1 between nodes, even when using relaxed memory consistencies such as eager write shared [CBZ91] lazy write shared [KDCZ94, Ift98, IPL96] or a hybrid [ACL 99] The primary problem is that two dimensional distributions force nodes to share data in two dimensions, but the data is organized in either a row or column major layout. Consequently, nodes share both rows and columns, meaning that every element in a page must ....

[Article contains additional citation context not shown here]

Liviu Iftode. Home-Based Shared Virtual Memory. PhD thesis, Princeton University, June 1998.

....only requires the last interval of the releaser to be visible to the acquiring processor on an acquire. The unique characteristic of JIAJIA is the use of home concept not only for data information, but also for coherence information, such as write notices. In many home based software DSM systems[14, 10], data information are processed in intermediate approach, that is to be propagated and applied to home eagerly, and fetched by others lazily. However, coherence information are propagated and applied either eagerly or lazily [15] In lock based cache coherence proto6 N P 3 P 2 P 1 P ....

....4. Encoding and decoding diffs do not take so much time as we may imagine. For all eight applications, T diff is lower than 1 of total time except Barnes(1. 2 ) Home based memory organization contributes much to this, since the write operation in the home node does not incur any diff operation[14]. On the other hand, the shared data is uniformly distributed among the processors, which favors some applications greatly. For example, in SOR and LU, there is no diffs to be generated at all. In order to have a clear understanding about the diff related time, we allocate all shared data on host ....

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Dept. of Computer Science, Princeton University, June 1998.

....home migration scheme, which is the unique home based software DSM system that implements this idea. The unique characteristic of JIAJIA is the use of home concept not only for data information, but for coherence information, such as write notices. In many home based software DSM systems[6, 17], data information is processed in intermediate approach, that is to be propagated and applied to home eagerly, and fetched by others lazily. However, coherence information are 8 propagated and applied either eagerly or lazily [7] In lock based cache coherence protocol, coherence information are ....

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Dept. of Computer Science, Princeton University, June 1998.

....3. Section 4 discusses methods for parallelizing sequential programs with the real applications as examples. Section 5 presents running results. The conclusion of this paper is drawn in Section 6. 2 The JIAJIA Software DSM JIAJIA is a home based software DSM which supports scope consistency[6]. It has two distinguishing features compared to other recent software DSM systems. First, it combines physical memories of multiple computers to form a larger shared space. Second, it implements the lock based cache coherence protocol with many optimization methods. 2.1 Programming Interface ....

L. Iftode, "Home-based Shared Virtual Memory", Ph. D. Thesis, Princeton University, Aug. 1998.

....and consequently improve performance significantly in some benchmarks. The extra time overhead of the write vector technique is negligible in the tested benchmarks. 1 Introduction Software DSMs can be categorized into homeless and homebased systems according to the way shared memory is organized [8]. In multiple writer protocols which implement relaxed memory consistency model, both homeless and homebased methods detect modifications to shared pages with twin and diff. However, they differ in the way modifications are kept and propagated. In a home based protocol, each shared page has a ....

....number of locally cached remote pages is larger than the maximum number allowed, some aged cache pages must be replaced to its home to make room for the new page. This allows JIAJIA to support shared memory that is larger than physical memory of one machine. JIAJIA implements the scope consistency[8] programming model. Multiple writer technique is employed to reduce false sharing. In JIAJIA, the coherence of cached pages is maintained through write notices kept on the lock. On a release, the releaser performs a comparison of all cached pages written in this critical section with their twins ....

L. Iftode, "Home-based Shared Virtual Memory", Ph. D. Thesis, Princeton University, Aug. 1998.

....Software DSMs Communication requirements are intimately related to coherence protocols. For example, in homeless system, such as CVM[16] and TreadMarks[18] scatter and gather operations are of great importance to propagate and apply diffs, while in home based system, such as JIAJIA[14] and HLRC[15], this requirement is unnecessary. Fortunately, homeless and home based software DSMs have very similar communication requirements except scatter gather operations. In the rest of this chapter we therefore take the home based JIAJIA system as our target software DSM system. Most software DSM ....

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Dept. of Computer Science, Princeton University, June 1998.

....processors fetch the same diff. Memory and coherence related complexity also limits the scalability of the LRC protocol. The work of this paper is supported partly by National Climbing Program of China and National Natural Science Foundation of China (Grant No. 69703002) Home based software DSMs[8] provide a simple, effective, and scalable way to build software DSMs. In a homebased protocol, each shared page has a designated home to which all writes are propagated and from which all copies are derived. In a multiple writer protocol which implements relaxed memory consistency model, diffs of ....

....according to the application sharing pattern in a home based software DSM called JIAJIA[7] JIAJIA is char acterized by its capability of combining physical memories of multiple computers to form a larger shared space. It implements the lock based cache coherence protocol for scope consistency[8]. The protocol is lock based because it totally eliminates directory and maintains coherence through accessing write notices kept on the lock. In our home migration scheme, pages that are written by only one processor between two barriers are recognized by the barrier manager and their homes are ....

[Article contains additional citation context not shown here]

L. Iftode, "Home-based Shared Virtual Memory", Ph. D. Thesis, Princeton University, Aug. 1998.

....of multiple workstations to form a large shared address space. JIAJIA is implemented entirely as a user level library and currently runs on many mainstream UNIX platforms. Multiple writer technique is employed to alleviate false sharing. JIAJIA implements the scope consistency (ScC) memory model[4]. The unique characteristic of JIAJIA is the use of home concept not only for data information, but also for coherence information, such as write notices. Generally, home based software DSM systems[4] propagate data information and apply them at homes eagerly, but this information is fetched by ....

....is employed to alleviate false sharing. JIAJIA implements the scope consistency (ScC) memory model[4] The unique characteristic of JIAJIA is the use of home concept not only for data information, but also for coherence information, such as write notices. Generally, home based software DSM systems[4] propagate data information and apply them at homes eagerly, but this information is fetched by others lazily. On the other hand, these systems propagate coherence information and apply them either eagerly or lazily. In our lock based cache coherence protocol, coherence information is processed in ....

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Dept. of Computer Science, Princeton University, June 1998.

....of locally cached pages is larger than the maximum number allowed, some aged cache pages must be replaced to its home to make room for the new page. This allows JIAJIA to support shared memory that is larger than physical memory of one machine. JIAJIA implements the scope memory consistency (ScC)[4] model. Multiple writer technique is employed to reduce false sharing. In JIAJIA, the coherence of cached pages is maintained through write notices kept on the lock. On a release, the releaser performs a comparison of cached pages written in this interval with their twins to produce diffs. These ....

L. Iftode, "Home-based Shared Virtual Memory", Ph. D. Thesis, Princeton University, Aug. 1998.

....technique is employed to alleviate false sharing. JIAJIA implements the scope consistency (ScC) 7] The unique characteristic of JIAJIA is the use of home concept not only for data information, but also for coherence information, such as write notices. Generally, home based software DSM systems[8, 9] propagate data information and apply them at homes eagerly, but this information is fetched by others lazily. On the other hand, these systems propagate coherence information and apply them either eagerly or lazily. In our lock based cache coherence protocol, coherence information is processed in ....

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Dept. of Computer Science, Princeton University, June 1998.

....propagation of writes, a word by word comparison is performed between the modified page and its twin to produce diff about this page. Multiple diffs of the same page from concurrent writers are then combined to form the up to date version of the page. Write detection of home based software DSMs[8] is more complex than write detection of COMA like software DSMs because detecting of writes to home pages has to be taken into account. In home based software DSMs, write detection of cached pages are similar to that of COMA like software DSMs. Write detection of home pages, on the other hand, is ....

....memory allocation function of JIAJIA allows the programmer to allocate a certain size of shared memory block by block cyclically across all processors. The processor from which the allocation starts can also be indicated. 2. 2 Cache Coherence Protocol JIAJIA implements the scope consistency (ScC)[8] memory consistency model. Multiple writer technique is employed to reduce false sharing. The distinguishing characteristic of JIAJIA s cache coherence protocol is its lock based feature, i.e. it does not rely on any directory information to maintain coherence. Instead, coherence is maintained ....

[Article contains additional citation context not shown here]

L. Iftode, "Home-based Shared Virtual Memory", Ph. D. Thesis, Princeton University, Aug. 1998.

....complexity also limits the scalability of the LRC protocol. The work of this paper is supported partly by National Climbing Program of China, National Science Foundation of China, and President Young Creation Foundation of the Chinese Academy of Sciences. Home based software DSMs[8] provide a simple, effective, and scalable way to build software DSMs. In a home based protocol, each shared page has a designated home to which all writes are propagated and from which all copies are derived. In a multiple writer protocol which implements relaxed memory consistency model, diffs ....

....to the application sharing pattern in a home based software DSM system called JIAJIA[7] JIAJIA is characterized by its capability of combining physical memories of multiple computers to form a larger shared space. It implements the lock based cache coherence protocol for scope consistency[8]. The protocol is lock based because it totally eliminates directory and maintains coherence through accessing write notices kept on the lock. In our home migration scheme, pages that are written by only one processor between two barriers are recognized by the barrier manager and their homes are ....

[Article contains additional citation context not shown here]

L. Iftode, "Home-based Shared Virtual Memory", Ph. D. Thesis, Princeton University, Aug. 1998.

....checkpointing with a state ofthe art DSM protocol in order to provide a faulttolerant yet high performing shared memory program0 7803 9802 5 2000 10.00 c 2000 IEEE. 1 ming environment on commodity clusters. Speci cally, we show how a Home based Lazy Release Consistency (HLRC) DSM protocol [16] can be extended with independent checkpointing in order to eciently tolerate single fault failures. Our fault tolerance algorithms are completely distributed, do not force processes to synchronize, and only make use of available protocol information. The choice of the base DSM coherence protocol ....

....operations. The local logical time of a process is de ned as a counter of local intervals. A local vector of logical times called vector timestamp [20] keeps track of the (partial) ordering of synchronization events occurring at any process. In Home based Lazy Release Consistency protocol (HLRC) [16], every shared page has an assigned home process which maintains the most recent version of the page. A process that has an invalid page fetches it from its home at page fault time. The protocol supports multiple writers: updates to a page are detected and propagated from a writer to its home as a ....

[Article contains additional citation context not shown here]

L. Iftode. Home-Based Shared Virtual Memory. Ph.D. Thesis, Princeton University, June 1998.

....on top of VIA, while preserving its performance bene ts. Recently, several message passing libraries over VIA, such as MPI [26] have been announced. In this paper, we describe an implementation of software distributed shared memory (DSM) over VIA, for a Linux based cluster of PCs. Software DSM [19] is available to applications as a runtime library that provides the abstraction of a shared address space across the cluster using message passing and virtual memory page protection. Given its low latency and overhead, as well as its capability to DMA directly into user address space of remote ....

....into user address space of remote memory without intermediate copies, VIA appears very promising for software DSM. To our best knowledge, ours is the rst implementation of a software DSM protocol on VIA. The protocol we have implemented on VIA is homebased lazy release consistency (HLRC) [37, 19]. Previous studies have shown that HLRC provides good scalability by reducing the number of messages and memory overhead compared to the homeless counterpart [37] Home based protocols have been previously implemented on other memory mapped interconnected clusters both for clusters of ....

[Article contains additional citation context not shown here]

Liviu Iftode. Home-based Shared Virtual Memory. PhD thesis, Princeton University, 1998.

....Since the sender can AU map a local address to only one remote bu#er, AU broadcast is not supported in the SHRIMP implementation of VMMC. 3 Protocols This section describes the two coherence protocols we have compared on the SHRIMP multicomputer system: Home base Lazy Release Consistency (HLRC) [29, 16] and Automatic Update Release Consistency (AURC) 12] We will first briefly review the lazy release consistency model. 3.1 Lazy Release Consistency Lazy Release Consistency (LRC) is a consistency model similar to release consistency [10] that is intended for software implementation. It delays ....

....triggered at global barriers to free up the di# storage, but this garbage collection is itself an expensive global operation that increases protocol overhead. 3. 2 Home Based LRC Protocols A home based multiple writer scheme is an alternative approach to supporting multiple concurrent writers [16]. This approach combines eager propagation of data at release time with lazy invalidations. For every shared page, a single node (usually the first writer) is designated as the page s home. The home s copy of the page is eagerly updated before the next release point. Non home copies are updated on ....

Liviu Iftode. Home-based Shared Virtual Memory. PhD thesis, Princeton University, 1998. Technical Report TR-58398.

....is how to eciently combine independent checkpointing and a software DSM protocol in order to provide a high performance fault tolerant sharedmemory programming environment on both local area and wide area clusters. Speci cally, we show how a Home based Lazy Release Consistency (HLRC) DSM protocol [18] can be extended with independent checkpointing in order to eciently tolerate single fault failures. While independent checkpointing avoids global coordination, it is susceptible to rollback propagation (i.e. the domino e ect [32] where a single failure may recursively force peer processes to ....

....failed peer processes may require state or log entries saved prior to the last local checkpoint. In this paper we present two algorithms, Lazy Log Trimming (LLT) and Checkpoint Garbage Collection (CGC) that can be used to build a fault tolerant Home based LRC DSM system. Home based LRC protocols [18, 47] are easier to implement relative to many other DSM protocols and have low communication and memory overhead, making them good candidates for large scale local area clusters as well as for meta clustering. We address two speci c questions: i) what are the rules for LLT and CGC, and (ii) how ....

[Article contains additional citation context not shown here]

L. Iftode. Home-Based Shared Virtual Memory. Ph.D. Thesis, Princeton University, June 1998.

No context found.

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Dept. of Computer Science, Princeton University, June 1998.

No context found.

L. Iftode. Home-Based Shared Virtual Memory. PhD thesis, Princeton University, Princeton, USA, Jun. 1998.

No context found.

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Princeton University, Dept of Computer Science, 1998.

No context found.

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Princeton University, Dept of Computer Science, 1998.

No context found.

L. Iftode. "Home-based Shared Virtual Memory". (PhD thesis), 1998.

No context found.

L. Iftode. Home-based Shared Virtual Memory. PhD thesis, Princeton University, August 1998.

No context found.

Liviu Iftode. Home-based shared virtual memory (thesis). Technical Report TR-583-98, Princeton University, Computer Science Department, June 1998.

No context found.

Liviu Iftode. Home-Based Shared Virtual Memory. PhD thesis, Princeton University, June 1998.

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