David E. Culler. Resource management for the tagged token dataflow architecture. Master's thesis, Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, January 1980. Also: MIT Laboratory for Computer Science Technical Report MIT/LCS/TR-332.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....109] though they did not make this observation until the later work of Zhang and Ortynski [108] Our results specialize to match theirs. Likewise, most of the systems oriented work on multithreading has ignored the space issue. Notable exceptions include the k bounded loops of Culler and Arvind [34, 35, 36] and the throttling mechanism of Ruggiero and Sargeant [90] These techniques and others [56, 57] have met with some success, though none have any algorithmic foundation. In algorithmic work that considers space requirements or communication costs, most prior work has focused on cases like ....

....While Culler and Arvind argued convincingly that the observed useless parallelism is in fact useless, they came short of a proof, and they left open the possibility of a clever scheduler that might be able to exploit this parallelism without using excessive amounts of space. With loop bounding [34, 35, 36] techniques, they were able to eliminate the useless parallelism with only a small decrease in the average parallelism. Their applications Some of the research reported in this chapter is joint work with Charles Leiserson of MIT s Laboratory for Computer Science and was first published in [13] and ....

David E. Culler. Resource management for the tagged token dataflow architecture. Master's thesis, Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, January 1980. Also: MIT Laboratory for Computer Science Technical Report MIT/LCS/TR-332.

....storage to be allocated. Since the matching store is finite, it is possible for it to become full, and cause a resource deadlock. Techniques have been developed to more or less solve this problem, these are: Throttling of tasks from University of Manchester [RS87] and K Bounding of loops from MIT [Cul85] 2. Synchronization is done at the instruction level. This causes some problems by itself. While in the case of von Neumann based parallel processors the processor itself is a critical resource, in the traditional dataflow model, the matching unit becomes the critical unit. If we consider all of ....

....multithreaded architectures. However, the problem of high degree of parallelism has resurfaced in multithreaded architectures as problem sizes have grown larger. The two main parallelism control mechanisms proposed for fine grain architectures are: Throttling of tasks [RS87, Sne93] and K bounding [Cul85] In addition Egan et al. [EWB91] have proposed methods of slicing the iteration space. Teo and Bohm [TB91] have proposed a method of chunking on fine grain machines by the use of iterative instructions. Throttling is a pure run time method of controlling parallelism. New activation requests may ....

D.E. Culler. Resource management for the tagged token data flow architecture. Technical Report TR-332, Laboratory for Computer Science, MIT, January 1985.

....it is interesting to note that Zhang s work stealing algorithm for backtrack search [45] actually gives at most linear expansion of space, but he does not mention this fact. The problem of storage management for multithreaded computations has been a growing concern among practitioners [13, 22]. To date, most existing techniques for controlling storage requirements have consisted of heuristics to either bound storage SCHEDULING MULTITHREADED COMPUTATIONS 25 use by explicitly controlling storage as a resource or reduce storage use by modifying the scheduler s behavior. We are aware of ....

D. E. Culler, Resource management for the tagged token dataflow architecture, master's thesis, Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Jan. 1980. Also: MIT Laboratory for Computer Science Technical Report MIT/LCS/TR-332.

....it is interesting to note that Zhang s work stealing algorithm for backtrack search [25] actually gives at most linear expansion of space, but he does not mention this fact. The problem of storage management for multithreaded computations has been a growing concern among practitioners [6, 12]. To date, most existing techniques for controlling storage requirements have consisted of heuristics to either bound storage use by explicitly controlling storage as a resource or reduce storage use by modifying the scheduler s behavior. We are aware of no prior scheduling algorithms for ....

David E. Culler. Resource management for the tagged token dataflow architecture. Master's thesis, Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, January 1980.

....is essential for high speed computing, it introduces a resource management problem. When the available program parallelism exceeds the machine parallelism, tokens that are ready for execution will saturate the matching store and other data buffers. In the worst case this can result in deadlock [7]. ffl Communication Overhead: fine granularity potentially increases the amount of data communication in the system. ffl Structure Store Overhead: because the execution model is fine grain and asynchronous, the structure store is designed to cope with asynchronous element wise access patterns ....

D.E. Culler. Resource management for the tagged token data flow architecture. Technical Report TR-332, Laboratory for Computer Science, MIT, January 1985.

....context can be used by another process. Processes have to be fairly large, otherwise they lead to too much throttle overhead; however, processes that are too large could have large internal parallelism and cause resource deadlocks. It is therefore important to find a proper balance. K bounding [Cul85] is a method that controls loop parallelism in Id. As originally designed, the compiler analyzes the code and determines the maximum resource usage for a loop cycle. At run time the hardware decides the number of loop cycles that can be allowed to execute in parallel based on the activity level ....

.... In addition, TAM is based on the code block model where each code block represents a semantically distinguishable code segment such as a non nested loop or function body [Ian90] A storage segment, called frame, is allocated for each code block instance or set of instances (as in k bounded loops [Cul85] All the data values pertinent to a given code block are stored in the corresponding frame, these include: synchronization slots, temporary 49 Integer Value Data Pointer Integer Value Array Elements Array Descriptor Integer Value Size Lower Bound Shift Value (a) Array Array Pointer Record ....

D.E. Culler. Resource management for the tagged token data flow architecture. Technical Report TR-332, Laboratory for Computer Science, MIT, January 1985.

....grain multithreaded architectures. However, the problem of high degree of parallelism has resurfaced in multithreaded architectures as problem sizes have grown larger. The two main parallelism control mechanisms proposed for fine grain architectures are: Throttling of tasks [16, 17] and K bounding [18]. In addition Egan et al. [19] have proposed methods of slicing the iteration space. Teo and Bohm [20] have proposed a method of chunking on fine grain machines by the use of iterative instructions. Throttling is a pure run time method of controlling parallelism. New activation requests may be ....

D.E. Culler. Resource management for the tagged token data flow architecture. Technical Report TR332, Laboratory for Computer Science, MIT, January 1985.

.... In addition, TAM is based on the code block model where each code block represents a semantically distinguishable code segment such as a non nested loop or function body [17] A storage segment, called frame, is allocated for each code block instance or set of instances (as in k bounded loops [8]) All the data values pertinent to a given code block are stored in the corresponding frame; these include synchronization slots, temporary storage area, loop constants, and etc. The Tera [1] provides hardware support for thread switching, replicated processor states, and split phase ....

D.E. Culler. Resource management for the tagged token data flow architecture. Technical Report TR-332, Laboratory for Computer Science, MIT, January 1985.

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