Prince Kohli, Mustaque Ahamad, and Karsten Schwan. Indigo: User-level support for building distributed shared abstractions. Concurrency: Practice and Experience, 8(10), 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... (a) the design and development of object based distributed operating systems [14] b) the development of consistency models for maintaining shared state in parallel and distributed systems [52, 34, 61, 6] and (c) efficient implementations of distributed shared memory systems on cluster machines [53, 61, 62, 31]. In an attempt to provide benchmark statistics to feed a scalability simulation for future media server capabilities (see 2.3.4) we have also performed measurements on the traffic between the electronic whiteboard and a central server machine (a uniprocessor SunSparc workstation) In a typical ....

Prince Kohli, Mustaque Ahamad, and Karsten Schwan. Indigo: User-level support for building distributed shared abstractions. Concurrency: Practice and Experience, 8(10), 1996.

....during the execution of a program on weak memory. Gibbons and Korach [45] give a method for online testing of a shared memory implementation. Systems supporting multiple consistency conditions were described by Agrawal et al. 10, 9] Carter et al. 30] Keleher et al. 64] and Kohli et al. [65]. Attiya et al. considered the eoeect of letting a single node issue multiple shared memory operations in parallel [16] A description of the DEC Alpha s memory consistency model is given by Attiya and Friedman in [18] Chandra et al. describe a language for writing memory consistency protocols ....

Prince Kohli, Mustaque Ahamad, and Karsten Schwan. Indigo: User-level support for building distributed shared abstractions. Concurrency: Practice and Experience, 1997. To appear.

....of the TOP C programming model. For shared memory machines, a utility, msbarrier( is offered. This is useful in the context of an updateenvironment( since a slave might be accessing data at the same time that updateenvironment( is causing it to be modified. Some systems, such as Indigo [17], provide the applications programmer with finer control over sharing of memory, and this could provide a more efficient alternative to msbarrier( The generality of the system can be further augmented by allowing arbitrary messages directly between slaves. This eliminates the inefficiency of ....

....in this paper uses the same mechanism for both memory models, it should be ideally suited to such an undertaking. Further, by sharing the distributed memory at the task level, it should present a higher level alternative to the userlevel sharing of distributed memory that is discussed in [17]. 9 Acknowledgements The author gratefully acknowledges Boston University for the use of the SPARC 1000 and SGI Power Challenger Array. ....

P. Kohli, M. Ahamad and K. Schwan, "Indigo: User-level Support for Building Distributed Shared Abstractions", Proc. of High Performance Distributed Computing '95 (HPDC-4), IEEE, 1995.

....swap space, or network global memory [6] This policy provides address equivalence for the global address space throughout the cluster the same address refers to the same datum. This approach is similar to the one taken in Shasta DSM system [28] The other approach taken by several DSM systems [3, 13, 14] is to have name equivalence. Our approach has the merit of saving the overhead of maintaining name tables required in supporting name equivalence, and making pointers valid uniformly across nodes. The only OS level support relied upon in achieving this partitioning of shared address space is ....

....the coherence granularity at which the cache and memory directory information is maintained, and the communication granularity at which data transfers take place. Having an application specific granularity relieves the need and overhead of twinning and diffing present in most software DSM systems [3, 13, 14]. A configurable access granularity has also been implemented in the Shasta DSM system [28] The more important aspect of access granularity is the access control mechanism. It is necessary to track potential violations of coherence and trigger the memory subsystem to take appropriate consistency ....

[Article contains additional citation context not shown here]

P. Kohli, M. Ahamad, and K. Schwan. Indigo: User-level Support for Building Distributed Shared Abstractions. In Fourth IEEE International Symposium on High-Performance Distributed Computing (HPDC-4), August 1995.

....of MD and the minimal impact that different synchro33 nization strategies have on MD performance indicate that design for distributed memory is a successful approach to creating efficient shared memory applications. Recently, it has also enabled easy porting of MD to distributed memory platforms[KAS95] Acknowledgments and Availability We would like to thank Bill Ribarsky and Uzi Landman for their help in getting this project started, Jian Ouyang for his work on the force calculations, and Tingkang Xia and Charles Cleveland for their help in debugging the physics of the sample system. The ....

Prince Kohli, Mustaque Ahamad, and Karsten Schwan. Indigo: User-level support for building distributed shared abstractions. Technical Report GIT-CC-95-36, College of Computing, Georgia Institute of Technology, Atlanta, GA 30332-0280, 1995. 34

....as well as the extensions that are necessary to develop a failure resilient version of the application. We then evaluate the performance of both versions of the application. 5.2. 1 Implementing TSP The TSP algorithm implemented by us is similar to the one used by Bal et al.[2] as well as Kohliet al..[40]. The application creates a statically defined job queue that is shared between tasks. The initial queue is composed of partial tours that are generated by unraveling possible paths through the cities to a fixed depth. Locks are utilized to dequeue jobs from the queue as well as to update the ....

....As illustrated in the Figure, completion times decrease as the number of worker tasks is increased. Thus, significant speedup is achievable with the system. Furthermore, the graph of completion times for this problem size of TSP is remarkably similar to graphs of completion times of other systems[40] that implement a 106 95000 100000 105000 110000 115000 120000 125000 130000 0 2 4 6 8 10 Time (milliseconds) Number of Disseminations Execution Time for 8 Node TSP Figure 34: Completion times of 8 Node TSP (variable dissemination) in milliseconds. distributed shared memory TSP ....

Prince Kohli, Mustaque Ahamad, and Karsten Schwan. Indigo: User-level support for building distributed shared abstractions. Concurrency: Practice & Experience, 10(1):1--29, January 1998.

.... of arbitrary type hierarchies in object access[4] Simultaneously, researchers have investigated the efficient implementation of DSM, including the development of efficient consistency maintenance protocols[5, 6, 7, 8, 9] experimentation with alternative representations of shared memory pages[10] and with alternative methods for dealing with specific implementation issues, like false sharing[11] This paper focuses on the required consistency of logically shared state information in complex distributed applications that range from irregular parallel scientific codes to distributed ....

P. Kohli, M. Ahamad, and K. Schwan, "Indigo: Userlevel support for building distributed shared abstractions, " in Fourth IEEE International Symposium on HighPerformance Distributed Computing, Aug 1995.

.... hierarchies in object access[17, 34] Furthermore, many researchers have investigated the efficient implementation of DSM, including the development of efficient consistency maintenance protocols[7, 4, 8, 21, 6, 14, 20, 28] experimentation with alternative representations of shared memory pages[22] and with alternative methods for dealing with specific implementation issues, like false sharing[9] Ongoing DSM research is developing hardware[27, 25] and operating system[23] support for efficient DSM implementations. Two particular issues addressed by such work are (1) the levels of ....

....can exploit their knowledge of such patterns to improve program performance. S DSO does not offer a single consistency protocol, nor does it implement some particular lock management scheme for use in synchronization (when desired by programmers) Instead, it builds on the generality of the Indigo[22] system to present to developers low level primitives with which they may construct exactly the shared object functionality and consistency semantics they desire. In addition, using attributes, consistency maintenance may exploit application level semantics specified by end users. Last, and in ....

[Article contains additional citation context not shown here]

Prince Kohli, Mustaque Ahamad, and Karsten Schwan. Indigo: User-level support for building distributed shared abstractions. In Fourth IEEE International Symposium on High-Performance Distributed Computing (HPDC-4), August 1995.

....etc. to our current low level communication calls, distributed shared memory (DSM) abstractions may be implemented and compared with alternative representations within the existing DSA library. Such an implementation and performance results attained on a cluster of workstations are described in [27]. Third, ongoing research on distributed objects is contributing language and compiler support for describing objects and then compiling object interactions into efficient runtime invocations, using custom communication protocols[3] and or exploiting active message paradigms[25, 48, 21] We share ....

.... machines[41] In contrast to those implementations, the layering of DSA objects on a basic remote invocation mechanism has resulted in library portability to various target platforms, including the aforementioned shared memory platforms and a recently completed implementation on a network platform[27]. Last, shared abstractions are easily instrumented, evaluated[41, 37, 26] and even dynamically adjusted, us2 ing custom[35] or library provided[19] mechanisms for on line program monitoring and without exposing such instrumentation to application programs[35] The remainder of this paper first ....

[Article contains additional citation context not shown here]

Prince Kohli, Mustaque Ahamad, and Karsten Schwan. Indigo: User-level support for building distributed shared absractions. In Fourth IEEE International Symposium on High-Performance Ditributed Computing (HPDC-4), August 1995.

....provides an overview of the implementation of TSP for network computing using Mocha as well as the extensions that are necessary to develop a failure resilient version of the application. 5.2. 1 Implementing TSP The TSP algorithm is similar to those used by Bal et al.[2] as well as Kohliet al..[21]. The application creates a statically defined job queue that is shared between tasks. The initial queue is composed of partial tours that are generated by unraveling possible paths through the cities to a fixed depth. Locks are utilized to dequeue jobs from the queue as well as to update the ....

....As illustrated in the Figure, completion times decrease as the number of worker tasks is increased. Thus, significant speedup is achievable with the system. Furthermore, the graph of completion times for this problem size of TSP is remarkably similar to graphs of completion times of other systems[21] that implement a distributed shared memory TSP application. Another well known behavioral characteristic of TSP applications implemented using distributed shared memory is a decrease in completion times when an update or push based dissemination implementation is utilized. This is due to the fact ....

Prince Kohli, Mustaque Ahamad, and Karsten Schwan. Indigo: User-level support for building distributed shared abstractions. Concurrency: Practice & Experience, 10(1):1-- 29, January 1998.

....detailed in Section 5 and we conclude the paper in Section 6. 2 Issues in Object Caching This section deals with the problems that need to be addressed in building an object caching system and the various possible solutions. Some of these problems also arise in software based implementations [3, 6, 8, 7] of distributed shared memories (DSM) We adapt some of the techniques used in DSM systems to address the problems for sharing objects instead of memory pages in the context of a distributed object system. 1. Object Faulting and Access Detection: In a system where caching is not used, the ....

P. Kohli, M. Ahamad, and K. Schwan. Indigo: User Level Support for Building Distributed Shared Abstractions, Fourth Symp. on High Performance Dist. Computing, August 1995.

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