J. Mattson Jr. An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction. PhD thesis, University of California, San Diego, Feb. 1993.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....Hence, computations of similar duration have similar chances of being selected. In contrast many other implementations assign priorities to speculative threads, typically lower priority than mandatory threads. Others have variable priorities often based on the depth of nested speculation, e.g. [13, 16]. Synchronous: ambND blocks until a result is in weak head normal form, whereas for the operators that allow nesting (such as ambNDND) it blocks while the nested ambND operators are evaluating. State preserving: The initial temptation may be to use the IO monad to carry the results of the ....

....admit, their main contribution lies in controlling the resource consumption in the presence of speculative computation. Performance measurements are reported on the RED graph reducer running on a 32processor nCube machine. Other closely related implementations of speculative parallelism are [13, 16]. Both implement lazy functional languages using a parallel graph reduction machine and use a priority scheduler with annotations to mark expressions for which speculative threads should be generated. The aim is to generate additional local parallelism to hide the communication latency in a system ....

J. Mattson Jr. An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction. PhD thesis, University of California, San Diego, Feb. 1993.

....the grade of parallelism, and so on, and to allow the programmer to analyse formally the consequences of some program decisions. This is particularly interesting for languages like Eden, where speculative parallelism is controled by the programmer, in contrast with other approaches like [Mat93] where it is done automatically by the compiler, as an optimization for the program execution. Besides, we intend to use the semantics to infer other properties inherent to parallel and concurrent programs. For instance, we can detect deadlock situations involving communications. For example, the ....

James S. Mattson Jr. An effective speculative evaluation technique for parallel supercombinator graph reduction. PhD thesis, University of California, San Diego, 1993.

....graph reduction on GRIP [5] however, we expect that the mechanisms described here are also appropriate for other distributed memory architectures. The details of the implementation are derived from a prototype developed for unrestricted speculative graph reduction on a shared memory machine [2]. 2 Terminology The program graph is represented by closures in the heap. A closure consists of a code pointer and a (possibly empty) set of arguments. A spark is a request for the concurrent reduction of a subgraph rooted at a particular closure. A spark is either mandatory or speculative, ....

....work available anyway. 8 Conclusions Previously, we argued that speculative threads should be at least as large as conservative threads, because they would suffer the same communication costs as equivalent conservative threads and higher costs for the overhead of speculative thread management [2]. While this is true for speculative threads that will be executed on remote processors, it does not apply to fine grain (childless) speculative threads that are permanently bound to the local processor. Local speculation actually encourages fine grain threads. For distributed graph reduction, ....

James S. Mattson Jr. An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction. PhD thesis, Department of Computer Science and Engineering, University of California, San Diego, Mail Code 0114, 9500 Gilman Drive, La Jolla, CA, USA, February 1993. Technical Report CS93--282.

....of communication, like that of distribution, is influenced by the underlying physical machine architecture. In shared memory machines, the tasks selected for reduction are often placed in a centralised global store, usually called a task pool, from where idle processors request work. Mattson [Matt93] has shown that 3.2.6 Controlling parallelism 34 having a centralised task pool is a bad design decision which may lead to the creation of hot spots. He therefore suggests the use of distributed task pools (like those maintained by the GRIP IMUs) since they avoid the problem of hot spots. In ....

JS Mattson Jt., An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction. PhD thesis, University of California, San Diego, 1993.

.... appropriate to both architectures, and hence it s name: Graph reduction for a Unified Machine model (GUM) The machine model chosen is message passing, which is the obvious choice for MPPs, and our earlier experience constructing runtime systems on SMPs lead us to believe that it suits them too [9]. To achieve an architecture independent runtime system we elected to extend our sequential runtime system, already written in portable gnu C [ with the standard PVM message passing library [11] The result is a parallel graph reduction engine. The price of architecture independence is ....

Mattson JS, An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction, PhD thesis, Dept. of Computer Science and Engineering, University of California, San Diego, (1993).

.... hence it s name: Graph reduction for a Unified Machine model (GUM) The machine model chosen is message passing, which is the obvious choice for massively parallel machines (MPPs) and our earlier experience constructing runtime systems on SMPs lead us to believe that it is also suitable for them [Mat93] To achieve an architecture independent runtime system we elected to extend our sequential multi threaded runtime system, already written in portable C, with a message passing library. We initially selected the PVM library [PVM93] The resulting system is a parallel graph reduction engine. The ....

J.S. Mattson Jr., An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction, PhD thesis, Dept. of Computer Science and Engineering, University of California, San Diego, 1993.

.... hence it s name: Graph reduction for a Unified Machine model (GUM) The machine model chosen is message passing, which is the obvious choice for massively parallel machines (MPPs) and our earlier experience constructing runtime systems on SMPs lead us to believe that it is also suitable for them [Mat93] To achieve an architecture independent runtime system we elected to extend our sequential multi threaded runtime system, already written in portable C, with a message passing library. We initially selected the PVM library [PVM93] The resulting system is a parallel graph reduction engine. The ....

J.S. Mattson Jr., An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction, PhD thesis, Dept. of Computer Science and Engineering, University of California, San Diego, 1993.

....As a consequence, while the implementation did achieve real speedups, they were less than might have been expected: a factor of 5 to 11 on a 16 processor configuration. Mattson observed similar problems with a similar sharedmemory implementation of the STG Machine on a 64processor BBN Butterfly [26]. In this case the best overall performance was achieved using approximately 16 processors, and adding more processors had a negative effect on performance. The root cause of these problems is that a shared memory system is capable of supporting only a limited number of accesses before the bus or ....

Mattson JS, An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction, PhD thesis, Dept. of Computer Science and Engineering, University of California, San Diego, (1993).

....just that induced by user interface actions. Of course, this goal amounts to some form of speculative evaluation. We considered two approaches to this. One was using an implementation language with a speculative evaluation mechanism built in, as has been described by Partridge [21] and Mattson [19], among others. Another was to use a lazy language but to provide demand to the lazy system by some external program, written in a conventional, eager language. This program would mediate between the user interface and the lazy system, by anticipating the user s demands. While we preferred the ....

James S. Mattson Jr. An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction. PhD thesis, Department of Computer Science and Engineering, University of California, San Diego, February 1993.

....form. 3 Implementation A prototype implementation containing these and other ideas has been developed for the BBN TC2000 butterfly . It is based on the Spineless Tagless G Machine model of graph reduction [4] Few modifications to the STG Machine are necessary to support speculative evaluation [3]. Conservative task synchronization is handled with the aid of special black hole nodes. When a task enters a updatable node, it locks the node by replacing it with a black hole. Any task that enters a black hole is suspended and added to the black hole s blocking queue. When a black hole is ....

James S. Mattson Jr. An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction, PhD thesis, Department of Computer Science and Engineering, University of California, San Diego, February 1993.

....the software is no longer hardware dependent (e.g. the GRIP microcode) Although performance results are not yet available, after tuning we hope that this system will prove to be practically useful. We may need to exploit techniques that have not yet been tried on GRIP (e.g. speculative evaluation [Par92,Mat93a]) in order to achieve good performance. It will probably also be necessary to use asynchronous message passing rather than the synchronous communication used on GRIP. ....

Mattson JS, "An Effective Speculative Evaluation Technique for Parallel Graph Reduction", PhD Thesis University of California at San Diego, February, 1993.

....and garbage collection requires inter processor synchronisation. As a consequence absolute speedups achieved were a factor of 5 to 11 on a 16 processor configuration. Mattson observed similar problems with a similar shared memory implementation of the STG Machine on a 64 processor BBN Butterfly [18]. The more recent GAML implementation is an attempt to address some of the shortcomings of the original ; G Machine. GAML introduces the notion of possibly shared nodes, which are the only nodes that must be locked. It also uses a linked list of stack chunks similar to those we use in GUM. ....

Mattson JS, An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction, PhD thesis, Dept. of Computer Science and Engineering, University of California, San Diego, (1993).

....concurrent variant of Haskell, which is typically used to implement user interfaces as a network of communicating sequential processes [6] An obvious extension is to integrate concurrent and parallel Haskell, creating networks of communicating parallel processes. Secondly, speculative evaluation [15] might also be supported. Some interesting programs rely on speculation to achieve performance, so this could open up some new applications. Care must be taken, however, to avoid the high overheads that are often associated with speculative techniques. Concurrency In Concurrent Haskell, threads ....

....is speculative. Such speculation is unwise under GUM if e 1 consumes many resources, or sparks additional tasks. A speculative RTS may terminate speculative tasks if they are discovered to be unnecessary, and may revert the graph into the state prior to the execution of the speculative task [15, 18]. Some of the machinery required to support speculation is already in place in GUM. For example, speculative threads could generate revertable black holes rather than normal blackholes. These can then be reverted to their original state if the thread which is evaluating them is found to be ....

Mattson JS, "An Effective Speculative Evaluation Technique for Parallel Supercombinator Graph Reduction", PhD thesis, Dept. of Computer Science and Engineering, University of California, San Diego, 1993.

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