R. Blumofe and P. Lisiecki. Adaptive and Reliable Parallel Computing on Networks of Workstations. In USENIX 1997.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....on host machines. Indeed, administration even of clusters is a challenge [20] Charlotte [5] was the first research project, to our knowledge, that was Java centric. Charlotte used eager scheduling, introduced by the Charlotte team, and implemented a full distributed shared memory. Cilk NOW [7], based on Cilk 2, provides for well structured computations (a strict subset of dag structured computations, where dag means directed acyclic graph) It uses work stealing and checkpointing (to a shared filesystem, such as NFS) for adaptively parallel computations (i.e. computations hosted by ....

R. D. Blumofe and P. A. Lisiecki. Adaptive and Reliable Parallel Computing on Networks of Workstations. In Proc. USENIX Ann. Technical Symposium, Jan. 1997. Anaheim.

....If another node, B, runs out of work while nodeA is still running the child task, it can steal the parent task from nodeA and continue to execute it. This workstealing algorithm has been shown to be provably efficient and fault tolerant. Cilk has been implemented in C for NOWs in Cilk NOW [18], and a proof of concept system using Java command line applications (not browser based applets) was implemented in ATLAS [11] Performance and scalability Finally, for volunteer computing to be truly useful, it must ultimately provide its users with speedups better than, or at least comparable ....

R.D. Blumofe and P.A. Lisiecki. Adaptive and Reliable Parallel Computing on Networks of Workstations, in Proc. of the USENIX 1997.

....we do not support the use of either fine grained, highcommunication threads or shared memory, both for performance reasons and for failure tolerance. At the same time, our model of parallelism should allow for inexpensive scheduling. We therefore base our model of parallelism on that of Cilk NOW [3] and, more generally, dataflow computer architectures [11] In our model, programs are split into segments called cords, on which we impose certain invariants to simplify scheduling tasks. First, once a cord is ready to run, it is able to execute continuously to completion without waiting for ....

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997.

....succinct but robust low level interface that ConCert provides. 2 Programming Model We seek to give the developer of grid applications access to the parallelism from within his application in a convenient and simple way. Drawing on Multilisp futures [Hal85] and a similar construct in Cilk NOW [BL97] our primitive notion The ConCert Project is supported by the National Science Foundation under grant ITR SY SI 0121633: Language Technology for Trustless Software Dissemination . on the grid is a task, a piece of code that runs to produce a result. Grid programs are composed of a number of ....

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133-147, Anaheim, California, 1997.

....succinct but robust low level interface that ConCert provides. 2 Programming Model We seek to give the developer of grid applications access to the parallelism from within his application in a convenient and simple way. Drawing on Multilisp futures [Hal85] and a similar construct in Cilk NOW [BL97] our primitive notion The ConCert Project is supported by the National Science Foundation under grant ITR SY SI 0121633: Language Technology for Trustless Software Dissemination . on the grid is a task, a piece of code that runs to produce a result. Grid programs are composed of a number of ....

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133--147, Anaheim, California, 1997.

....we do not support the use of either ne grained, highcommunication threads or shared memory, both for performance reasons and for failure tolerance. At the same time, our model of parallelism should allow for inexpensive scheduling. We therefore base our model of parallelism on that of CilkNOW [BL97] and, more generally, data ow computer architectures [CSG99] In our model, programs are split into segments called cords, on which we impose the three invariants. Although we do not at present enforce these invariants via certi cation, they are necessary for clients to write correct and ....

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133-147, Anaheim, California, 1997.

....we do not support the use of either fine grained, highcommunication threads or shared memory, both for performance reasons and for failure tolerance. At the same time, our model of parallelism should allow for inexpensive scheduling. We therefore base our model of parallelism on that of CilkNOW [BL97] and, more generally, dataflow computer architectures [CSG99] In our model, programs are split into segments called cords, on which we impose the three invariants. Although we do not at present enforce these invariants via certification, they are necessary for clients to write correct and ....

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133--147, Anaheim, California, 1997.

....the protected node, the node migrates and all other speculative protection nodes are garbage collected. Low level protocols to support, what is in effect, remote node invocation has been investigated elsewhere: a Triple Manager scheduling PVM processes [9] and a traditional dataflow system [13]. Investigating suitable domain scheduling heuristics is a topic for future research. Example 9. Condensed Graphs are used to exploit parallelism in a computation and the Triple Manager(s) can schedule the computation across networks of workstations [10] Figure 15 gives an example of a graph ....

R.D. Blumofe, P. L. Adaptive and reliable parallel computing on networks of workstations. Proceedings of the USENIX 1997 Annual Technical Symposium (January 1997).

....a provably good work stealing scheduling algorithm and follows a work rst principle. Cilk concentrates on minimizing overheads that contribute to work, even at the expense of overheads that contribute to the critical path [8] CilkNOW is an implementation of Cilk for networks of workstations [1, 3]. It transparently manages resources, provides transparent fault tolerance, and implements adaptive parallelism which allows a Cilk application to run on a set of workstations that may grow and shrink throughout program execution. Cilk s underlying programming model is limited to ....

R.D. Blumofe and P.A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Symposium, California, 1997.

....5. RELATED WORK We discussed load balancing in Satin, a divide and conquer extension of Java. Satin is designed for wide area systems, without shared memory. Many divide and conquer systems are based on the C language. Among them, Cilk [7] only supports sharedmemory machines, CilkNOW [6] and DCPAR [11] run on localarea, distributed memory systems, but do not support shared data. SilkRoad [20] is a version of Cilk for distributed memory that uses a software DSM to provide shared memory to the programmer, targeting at small scale, local area systems. Alice [9] and Flagship [25] ....

R. Blumofe and P. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, Anaheim, California, 1997.

....parallel programs, whereas our technique targets object based multithread programs. Computing systems that adaptively use idle machines. Various systems have been developed to allow parallel programs to execute on a network of workstations with a variable number of nodes [SLGZ99, CFGK95, BL97] In these systems, computing nodes join the computation when they become idle, and withdraw when their users need them. The way in which these systems adjust the number of processors is based on the availability of idle computers and not based on performance or the amount of resource ....

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and Reliable Parallel Computing on Networks of Workstations. In Proceedings of the USENIX 1997 Annual Technical Symposium, pages 133-- 147, Anaheim, January 1997.

....Piranha, and Persistent Linda. such as Piranha and Persistent Linda offer a trade off between high level constructs and efficiency. However, Persistent Linda achieves fault tolerance and heterogeneity with lower overhead. 2.5 Other Parallel Computing Systems on NOW 2.5. 1 Cilk NOW Cilk NOW [6] is a runtime system designed to execute Cilk [5] program in parallel on a network of UNIX workstations. Cilk (pronounced silk ) is a parallel multi threaded extension of the C language, and all Cilk runtime systems employ a provably efficient thread scheduling algorithm. A Cilk program contains ....

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and Reliable Parallel Computing on Networks of Workstations. In Proceedings of the USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133-147, Anaheim, California, January 6-10, 1997.

....without shared memory. The Java classes presented in [9] can also be used for divide and conquer algorithms. However, they are restricted to shared memory systems. Most other divide and conquer systems are based on the C language. Among them, Cilk [4] only supports shared memory machines, CilkNOW [3] and DCPAR [7] run on local area, distributed memory systems, however without support for shared data. SilkRoad [12] is a version of Cilk for distributed memory that uses a software DSM to provide shared memory to the programmer, targeting at small scale, localarea systems. Alice [6] and Flagship ....

R. Blumofe and P. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, Anaheim, California, 1997.

....the node then executes. If another node, B, runs out of work while NodeA is still running the child task, it can steal the parent task from NodeA and continue to execute it. This work stealing algorithm has been shown to be provably efficient and fault tolerant. Cilk has been implemented for NOWs [30], and has been implemented using Java applications (but not applets) in ATLAS [31] Project Bayanihan aims to develop a framework that does not limit the programmer to using only one form of adaptive parallelism, but instead makes it easy to implement any of these forms and even develop new ones. ....

Blumofe, R. D., Lisiecki, P. A.: Adaptive and Reliable Parallel Computing on Networks of Workstations. Proceedings of the USENIX 1997 Annual Technical Symposium. (Jan. 1997)

....is represented as a directed acyclic graph (DAG) Each node of the graph denotes a task. Each arc of the graph denotes an item produced by one task and used by another. In other presentations of the dataflow model, a task is named an actor [Models] Streams] a node [Architectures] a thread [Cilk NOW][MIMDStyle ] TAM] or another name. The names task and item are used here for consistency with previous presentations of TSIA. A task consists of items: ins, inouts and outs. An in is an item used by the task. An inout is an item modified by the task. An out is an item produced by the task. An ....

....obeys the dependencies between tasks, an application execution has a determinate result. 4 The State of an Executing Application As introduced in the previous sections, an executing application is represented as a directed acyclic graph (DAG) Because the graph is acyclic, a task executes once [Cilk NOW][MIMD Style] Monsoon] TAM] This approach is pursued here. In contrast, some previous dataflow models introduced special items and tasks for cycles. Parts of the graph then may execute more than once in the application execution [Models] Architectures ] 5 A graph conveniently represents the ....

[Article contains additional citation context not shown here]

Robert D. Blumofe and Philip A. Lisiecki. "Adaptive and Reliable Parallel Computing on Networks of Workstations", in Proceedings of the USENIX 1997 Annual Technical Symposium, Anaheim, California, USA, January 6-10, 1997. supertech.lcs.mit.edu/cilk

....by using load balancing: Systems such as DOME [3] Dataparallel C [7, 13] Charm [8, 9] and various versions of PVM [10] can adjust the load per node on partially available workstations, but the processor set for the computation is fixed, once started, as opposed to our system. Cilk NOW [4] and Piranha [5] support adaptive parallel computation on NOWs in the sense that the pool of processors used can vary during the computation, as in our system. However, the Cilk NOW system is restricted to func tional programs, and Piranha requires the adoption of the Linda tuple space as a ....

R.D. Blumofe and P.A. Lisiecki. Adaptive and reliable parallel computing on network of workstations. In Proceedings of the USENIX 1997 Annual Technical Symposium, pages 133--147, January 1997.

....5 Related Work We discussed Satin, a divide and conquer extension of Java. Satin has been designed for distributed memory machines, while most divide and conquer systems use shared memory machines (e.g. Cilk [6] There is also a version of Cilk for distributed memory machines, called CilkNOW [5], but it only supports functional Cilk programs (without shared memory) and it does not make a deep copy of the parameters to spawned methods. Our own previous work on parallel divide and conquer [9] was based on the C language while having similar restrictions as CilkNOW. Alice [7] and Flagship ....

R. Blumofe and P. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In In Proceedings of the USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, Anaheim, California, 1997.

....of a task may change the graph and may use or produce the items of the task. The dataflow model thus is a form of graph reduction. For example, ALICE [ALICE] is an implementation of graph reduction with many similarities to the implementation of dataflow by Cilk NOW and its precursors [Cilk 1][Cilk NOW] [PCM] 3 In the succession of dataflow models, that of Cilk NOW and its precursors is the closest model to that of TSIA. Cilk NOW provides a subset of the C programming language with a transparent reliable, adaptive and parallel execution. In Cilk NOW and its precursors, an application executes ....

....task part and concerns calls to routines. The TSIA language of this presentation is designed to help clearly present TSIA and to demonstrate its feasibility. A crude implementation is described in the next section. A TSIA language with a sophisticated implementation is described elsewhere [Cilk 1][Cilk NOW][PCM] For a routine written in a TSIA language, the execution proceeds as usual, except that calls to other routines yield the replacement tasks on the stack, as briefly described in the previous section. This alternative implementation of routines is known as delegation in other presentations ....

[Article contains additional citation context not shown here]

Robert D. Blumofe and Philip A. Lisiecki. "Adaptive and Reliable Parallel Computing on Networks of Workstations", in Proc. of the USENIX 1997 Annual Technical Symposium, Anaheim, California, USA, January 6-10, 1997, pp. 133-147. supertech.lcs.mit.edu/cilk

....of the TSIA is the fact that many TSIAs exist. Examples include TSIAs or TSIA like approaches for each of the following application areas: a reliable, adaptive, heterogeneous, parallel application for simulation [Funnel] a reliable, adaptive, parallel subset of the C programming language [Cilk NOW] adaptive master worker parallelism [Linda Piranha] coarse grain parallelism [Jade] computational fluid dynamics on shared memory multiprocessor computers [CFD] using the otherwise idle cycles of thousands of computers on the Internet to process data from the search for ....

....to have been created largely independently. Secondly, the similarities between existing systems were not recognized. Thus these systems were not grouped together; though there are some exceptions [SISA] Thirdly, systems intended to support a variety of applications have not been further pursued [Cilk NOW] Linda Piranha] Jade] There seem to be at least two possible reasons why TSIAs previously were not recognized. Firstly, the achievement of a high level definition and its benefits were not recognized. Secondly, TSIAs were naively assumed to be restricted to bag of tasks applications. This ....

[Article contains additional citation context not shown here]

Robert D. Blumofe and Philip A. Lisiecki. "Adaptive and Reliable Parallel Computing on Networks of Workstations", in Proceedings of the USENIX 1997 Annual Technical Symposium, Anaheim, California, USA, January 6-10, 1997. (This and other Cilk information is available at http://theory.lcs.mit.edu/~cilk/).

....definition and on the item architecture (IA) of TSIA. 2.1 The Precursors of the Name Task The name task has had many precursors. These include the task [Fx] Jade] the independent task [Spawn] the thread (not to be confused with the thread as known in this thesis and elsewhere in computing) Cilk NOW] Coarse Grain Dataflow] the regular Mentat object [Mentat] the upcall [RTU] SUMO] the filter predicate [Packet Filter] the event [168 E] Funnel] the soft instruction and others [SISA] Precursors of the task are cited throughout this presentation. By mentioning a few original names, the ....

....in section 1.1, a TS is built on top of the underlying environment. The TS thus is perhaps best imagined as a run time service which is only loosely coupled to the environment. Loosely coupled means that the implementation of the TS requires little to no modification of an existing environment [Cilk NOW] Funnel] Mentat] The possibility of such a loosely coupled TS supports the feasibility of a TS. In addition, imagining a TS as a loosely coupled system helps maintain the conceptual distinction between the TS and the environment. Of course a TS need not be only loosely coupled to the ....

[Article contains additional citation context not shown here]

Robert D. Blumofe and Philip A. Lisiecki. "Adaptive and Reliable Parallel Computing on Networks of Workstations", in Proceedings of the USENIX 1997 Annual Technical Symposium, Anaheim, California, USA, January 6-10, 1997. (This and other Cilk information is available at http://theory.lcs.mit.edu/~cilk/). 234

....tasks . out1000. Before execution, these thousand tasks are in the task pool of Figure 1. The TSIA does the simple management of each task. For an application which executes in terms of tasks, these are some of the transparent executions that can be provided by a TSIA: reliable execution [Cilk NOW] DNA] After the failure of a computer, its task executes on the restarted computer or on another computer. parallel execution [Cilk NOW] DNA] Linda Piranha] Each task executes on one of multiple computers. distributed execution [DNA] Internet Computing] A task executes on a remote ....

....For an application which executes in terms of tasks, these are some of the transparent executions that can be provided by a TSIA: reliable execution [Cilk NOW] DNA] After the failure of a computer, its task executes on the restarted computer or on another computer. parallel execution [Cilk NOW] DNA] Linda Piranha] Each task executes on one of multiple computers. distributed execution [DNA] Internet Computing] A task executes on a remote computer. heterogeneous execution [DNA] Internet Computing] A task executes on a different kind of computer. adaptive execution ....

[Article contains additional citation context not shown here]

Robert D. Blumofe and Philip A. Lisiecki. "Adaptive and Reliable Parallel Computing on Networks of Workstations", in Proceedings of the USENIX 1997 Annual Technical Symposium, Anaheim, California, USA, January 6-10, 1997. supertech.lcs.mit.edu/cilk .

....5 Related work We discussed Satin, a divide and conquer extension of Java. Satin has been designed for distributed memory machines, while most divide and conquer systems use shared memory machines (e.g. Cilk [5] There is also a version of Cilk for distributed memory machines, called CilkNOW [4], but it only supports functional Cilk version 2 programs (without shared memory) and it does not make a deep copy of the parameters to spawned methods. Our own previous work on parallel divide and conquer [10] was based on the C language while having similar restrictions as CilkNOW. Another ....

R. Blumofe and P. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In In Proceedings of the USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, Anaheim, California, 1997.

....of existing computing practice and research. For example, part of an application may use a TSIA, while the other part does not. For some such applications, the role of the TSIA is similar to that of a coordination language [Coordination ] One of the most general of the existing TSIAs is Cilk NOW. It provides a transparent reliable, adaptive and parallel execution [Cilk NOW] Since it supports a subset of the C programming language, Cilk NOW allows for applications far beyond a bag of tasks application. As outlined elsewhere [MIMD Style] also some other systems for parallel computing ....

....may use a TSIA, while the other part does not. For some such applications, the role of the TSIA is similar to that of a coordination language [Coordination ] One of the most general of the existing TSIAs is Cilk NOW. It provides a transparent reliable, adaptive and parallel execution [Cilk NOW] Since it supports a subset of the C programming language, Cilk NOW allows for applications far beyond a bag of tasks application. As outlined elsewhere [MIMD Style] also some other systems for parallel computing execute an application in terms of tasks. In these systems, a task is known as a ....

[Article contains additional citation context not shown here]

Robert D. Blumofe and Philip A. Lisiecki. "Adaptive and Reliable Parallel Computing on Networks of Workstations", in Proceedings of the USENIX 1997 Annual Technical Symposium, Anaheim, California, USA, January 6-10, 1997. supertech.lcs.mit.edu/cilk

....perhaps most, can have a convenient definition which executes in terms of tasks. Also the second assumption is wrong. Tasks allow for a very performant application execution. For example, Cilk NOW provides a subset of the C programming language with a performant execution in terms of tasks [Cilk NOW]. Thus applications beyond bag of tasks applications are provided by Cilk NOW with a performant and transparent reliable, adaptive and parallel execution. Another example is the easy implementation of partial execution in a TSIA, as demonstrated in this presentation. As presented elsewhere [TSIA] ....

....items. For example, in the execution illustrated in Figure 3, the execution of the task tpow(0,1.6; u) yields u=1.0. Similarly, the execution of mult(2.56,1.6; s) yields s=4.096. An implementation of an execution in terms of tasks, including delegation and the task pool, is described elsewhere [Cilk NOW]. 5 The Partial Execution of a Task When all its items are available, a task can have a complete execution. For example, the previous section describes how the complete execution of the task tpow(3,1.6; s) yields s=4.096. Complete execution, same for partial execution described below, refers to ....

Robert D. Blumofe and Philip A. Lisiecki. "Adaptive and Reliable Parallel Computing on Networks of Workstations", in Proceedings of the USENIX 1997 Annual Technical Symposium, Anaheim, California, USA, January 6-10, 1997. supertech.lcs.mit.edu/cilk

....to migratory behavior was first suggested by Cox and Fowler [9] and Stenstrom et al. 24] in the context of hardware shared memory machines. Another form of adaptivity that is important in networks of workstations is adapting to environmental characteristics such as processor and network load [6], 7] This form of adaptivity is orthogonal to the one discussed in this paper. VIII. Conclusions We have described software DSM protocols that automatically adapt, on a per page basis, to the access patterns in the application. The protocols dynamically choose between single and ....

R.D. Blumofe and P.A. Lisiecki. Adaptive and reliable parallel computing on network of workstations. In Proceedings of the USENIX 1997 Annual Technical Symposium, January 1997.

....to allow sequential computations to use idle time on networked nodes. These systems include, among others, Butler [20] Condor [17] and many process migration systems such as, e.g. Sprite [8] Our work distinguishes itself from these systems by its support for parallel computations. Cilk NOW [5], Dataparallel C [19] Piranha [6] and various migration based systems (e.g. Millipede [12] or versions of PVM [15] support adaptive parallel computation on NOWs. Blumofe and Lisiecki [5] describe the Cilk NOW system for adaptive and reliable parallel execution of functional Cilk programs on ....

....[8] Our work distinguishes itself from these systems by its support for parallel computations. Cilk NOW [5] Dataparallel C [19] Piranha [6] and various migration based systems (e.g. Millipede [12] or versions of PVM [15] support adaptive parallel computation on NOWs. Blumofe and Lisiecki [5] describe the Cilk NOW system for adaptive and reliable parallel execution of functional Cilk programs on networks of workstations. Programs need to be written in the Cilk language. Only functional Cilk programs are supported by Cilk NOW. Nodes may join and leave an ongoing computation at any ....

R.D. Blumofe and P.A. Lisiecki. Adaptive and Reliable Parallel Computing on Network of Workstations. In Proceedings of the USENIX 1997 Annual Technical Symposium, pages 133--147, January 1997.

....include the following: ffl Use the higher performance networks available on the Whitney cluster. We are using UDP on the ethernet. We hope to try these experiments on a machine with a high performance system area network. ffl We hope to merge the fault tolerant Cilk of Blumofe and Lisiecki [BL97] with our system so that TreeCilk will be able to tolerate processor failures. In the longer term, internet applications such as TreeCilk will become more common, and the internet will need to provide a way to control the bandwidth consumption of these applications. It is becoming clear how one ....

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Symposium, Anaheim, California, 6--10 January 1997. (ftp://theory.lcs.mit.edu/pub/cilk/USENIX97.ps.gz).

....current usage. This would especially be useful ffl Application Fault Tolerance. The ultimate goal is to write applications to the River interface that not only have robust performance, but also can continue operation under machine failure, similar to work in other dynamic programming environments [8, 43]. Some form of automatic check pointing may be the solution, as suggested in [5] We also believe River is well suited to a large class of external, distributed applications, including traditional scientific codes and perhaps multimedia programs as well. Some evidence for this exists in the ....

R. D. Blumofe and P. A. Lisiecki. Adaptive and Reliable Parallel Computing on Networks of Workstations. In USENIX, editor, 1997 Annual Technical Conference, January 6--10, 1997. Anaheim, CA, pages 133--147, Berkeley, CA, USA, Jan. 1997. USENIX.

....the node then executes. If another node, B, runs out of work while Node A is still running the child task, it can steal the parent task from NodeA and continue to execute it. This work stealing algorithm has been shown to be provably efficient and fault tolerant. Cilk has been implemented for NOWs [28], and has been implemented using Java applications in ATLAS [29] Project Bayanihan aims to develop a framework that does not limit the programmer to using only one form of adaptive parallelism, but instead makes it easy to implement any of these forms and even develop new ones. Fault Tolerance. ....

Robert D. Blumofe, and Philip A. Lisiecki. Adaptive and Reliable Parallel Computing on Networks of Workstations. In Proceedings of the USENIX 1997 Annual Technical Symposium, Anaheim, California, January 6-10, 1997.

....systems. It is based on the tuple space abstraction introduced by Linda to which it adds a few run time hooks to support adaptive parallelism. It relies on a restrictive flat master worker model and requires programmers to explicitly code the actions required to release a machine. Cilk(v2) [Blumofe and Lisiecki, 1997] is an integrated language and system that only supports a very limited functional style of parallel computing. However, Cilk supports transparent adaptation and fault tolerance on the basis of automatic cancellation and restart of subtasks that have yet to report their result. 3 Communication ....

Blumofe, R D and Lisiecki, P A (1997). Adaptive and reliable parallel computing on networks of workstations. In Proceedings of the USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, Anaheim, California.

....destination a load balancing policy determines if and where it is sent. Another difference is that the transfer of a seed task from one processor to another is a delayed action, whereas the transfer of a remote procedure call takes place almost immediately. A seed task is also similar to Cilk s [18] closure concept. Cilk programs are written using a functional paradigm and a closure can be activated once all of its arguments are available. Cilk s closure concept is only suitable for a functional programming paradigm. We foresee that the seed task concept has the ability to be suitable for ....

R.D. Blumofe and P.A. Lisiecki, Adaptive and Reliable Parallel Computing on Networks of Workstations, Proceedings of the USENIX 1997 Annual Technical Symposium, Anaheim, California, January 1997.

....to allow sequential computations to use idle time on networked nodes. These systems include, among others, Butler [19] Condor [16] and many process migration systems such as, e.g. Sprite [7] Our work distinguishes itself from these systems by its support for parallel computations. Cilk NOW [4], Dataparallel C [18] Piranha [5] and various migration based systems (e.g. Millipede [11] or versions of PVM [14] support adaptive parallel computation on NOWs. Blumofe and Lisiecki [4] describe the Cilk NOW system for adaptive and reliable parallel execution of functional Cilk programs on ....

....[7] Our work distinguishes itself from these systems by its support for parallel computations. Cilk NOW [4] Dataparallel C [18] Piranha [5] and various migration based systems (e.g. Millipede [11] or versions of PVM [14] support adaptive parallel computation on NOWs. Blumofe and Lisiecki [4] describe the Cilk NOW system for adaptive and reliable parallel execution of functional Cilk programs on networks of workstations. Programs need to be written in the Cilk language. Only functional Cilk programs are supported by Cilk NOW. Nodes may join and leave an ongoing computation at any ....

R.D. Blumofe and P.A. Lisiecki. Adaptive and reliable parallel computing on network of workstations. In Proceedings of the USENIX 1997 Annual Technical Symposium, January 1997.

....unlike our global objects, and can be rather inefficient. A more recent project Mist [7] considers migration of full OS processes, rather than lightweight tasks; however this is very expensive compared with lightweight task migration. Unlike Gardens Mist addresses fault tolerance. Cilk(v2) [5] is an interesting project which supports efficient adaptive and fault tolerant parallel computation. However Cilk programs must be cast into a very restricted functional model of parallelism. A number of other attempts have been documented to integrate general process migration facilities into ....

R D Blumofe and P A Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In Proceedings of the USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, Anaheim, California, 1997.

No context found.

R. Blumofe and P. Lisiecki. Adaptive and Reliable Parallel Computing on Networks of Workstations. In USENIX 1997.

No context found.

Blumofe, R. D. and P. Lisiecki: 1997, `Adaptive and Reliable Parallel Computing on Networks of Workstations'. In: USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems. Anaheim, CA, pp. 133--147.

No context found.

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133--147, 1997.

No context found.

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997.

No context found.

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In Proceedings of the USENIX 1997.

No context found.

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133--147, 1997.

No context found.

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133--147, 1997.

No context found.

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133--147, 1997.

No context found.

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In USENIX 1997 Annual Technical Conference on UNIX and Advanced Computing Systems, pages 133--147, 1997.

No context found.

Robert D. Blumofe and Philip A. Lisiecki. Adaptive and reliable parallel computing on networks of workstations. In Proceedings of the USENIX 1997.

No context found.

R.D. Blumofe, P.A. Lisiecki. "Adaptive and Reliable Parallel Computing on Networks of Workstations". In proc. of USENIX 1997.

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