C. A. Ruggiero. Throttle Mechanisms for the Manchester' Dataflow Machine. PhD thesis, University of Manchester, Manchester M13 9PL, England, July 1987. 22  Home/Search   Document Not in Database   Summary   Related Articles   Check

....localizes 50 of the accesses. Observe that the computational load is quite even with this policy. Caching results in a five fold speedup. 4. 3 K bounds Several techniques have been proposed in the literature for controlling parallelism, including k bounded loops[6] and activation tree throttling[20]. For the most part, these approaches have been validated through idealized simulations and execution on a small number of processors. Here we present data on the effects of k bounded loops on 64 processors. In this case, we study Gamteb under a fixed problem size (8092 particles) on a fixed ....

C. A. Ruggiero. Throttle Mechanisms for the Manchester' Dataflow Machine. PhD thesis, University of Manchester, Manchester M13 9PL, England, July 1987. 22

....all tokens go through the global token matching unit. Loops are also unfolded in nature if any restriction is not made. Since the system resource is limited, however, the mechanisms to throttle massive parallelism in loops have been studied as k bounded loop[5] or the intelligent token queue[19]. In multithreaded architectures, the situation is different. Since an execution unit, called thread, is coarser than an instruction, we have a chance to exploit the locality of computation within a thread through registers or caches. Note that a thread is defined as a set of instructions which ....

C. A. Ruggiero, Throttle Mechanisms for the Manchester Dataflow Machine, PhD thesis, Univ. of Manchester, July 1987.

....threaded execution models discussed in the literature; TAM was developed to address these issues directly. The key observation is that the activation tree and the continuation pool are typically quite large, except on toy programs. This has been demonstrated empirically for programs in Id[8] Sisal[23], and Multilisp[18] Minimizing the activation tree size while exposing sufficient parallelism is an active area of research, but even with advances in this area we cannot expect the entire activation tree or the entire continuation pool to be maintained in high speed processor storage. Therefore, ....

C. A. Ruggiero. Throttle Mechanisms for the Manchester Dataflow Machine. PhD thesis, University of Manchester, Manchester M13 9PL, England, July 1987.

....of the top level matching store, the synchronization cost increases dramatically, since some form of overflow store must be used[12, 6] In dataflow models the problem is made more severe by the lack of distinction between long lived call frame storage and short lived register storage. Throttling[17] and k bounding[9] were introduced to constrain the parallelism in dataflow programs so that the VPs would fit within a modest matching store. The Hybrid proposal[14] makes a clear distinction between call frame and register storage, and places the requirement that registers must be vacant at the ....

....exists, however, it remains an open question how this high level direction should be applied in a consistent manner. Acceleration of the local, naive dynamic scheduling mechansism through hardware enhancements has no impact on this fundamental problem. Work on k bounded loops[9] and throttling[17] is a step in the right direction, but neither of these approaches has fully considered that scheduling of computation must compete for resources with the computation. If compilers are not able to meet this challenge, the programmer will be forced to express the scheduling in the program, by ....

C. A. Ruggiero. Throttle Mechanisms for the Manchester Dataflow Machine. PhD thesis, University of Manchester, Manchester M13 9PL, England, July 1987.

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