S. Orlando and R. Perego. SUPPLE: an efficient run-time support for non-uniform parallel loops. Research Report CS-96-17, Dip. di Matematica Applicata ed Informatica, Universita Ca'Foscari di Venezia, December 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....effects. Load balancing The mapping of threads to processors can also be used to improve the application load balance. Some research focuses on statically finding a good mapping [11, 35, 8] It take place in the world of task scheduling and only works for (statically) predictable task graphs. [26] dynamically maps such threads to balance the load with a work stealing technique. But, it only works for a particular kind of RR n3207 8 Christian Perez Raymond Namyst HPF nested loops. In [28] we have studied the feasibility of our approach. The paper focuses on HPF and shows that good ....

S. Orlando and R. Perego. SUPPLE: an efficient run-time support for non-uniform parallel loops. Research Report CS-96-17, Dip. di Matematica Applicata ed Informatica, Universita Ca'Foscari di Venezia, December 1996.

....of task scheduling. When the task graph is (statically) predictable, one can look for an optimal mapping of the abstract processors on the nodes [6, 25] For certain kinds of HPF loops, it is possible to dynamically map the abstract processors to balance the load with a work stealing technique [17]. In contrast, we propose to let abstract processors migrate dynamically among the nodes within their lifetime. ffl Take advantage of multi processor nodes where available, just through the native process scheduling [19] ffl As an extra bonus, distributing the data among a larger number of ....

S. Orlando and R. Perego. SUPPLE: an efficient runtime support for non-uniform parallel loops. Research Report CS-96-17, Dip. di Matematica Applicata ed Informatica, Universita Ca'Foscari di Venezia, Dec. 1996.

....and time sharing of resources. Here we restrict our view to data parallel computations expressed by means of parallel loops. In previous works we considered imbalances introduced by non uniform data parallel computations to be run on homogeneous, distributed memory, MIMD parallel systems [12, 15, 13, 14]. We devised a novel compiling technique for parallel loops and a related run time support (SUPPLE) This paper shows how the same support can also be utilized to implement loops in all the cases where load imbalance is not a characteristic of the user code, but is caused by variations in ....

....regular problems and homogeneous space shared architectures, workload imbalance due to the irregularities of either the problem or the target machine may worsen the final performance. Much research has been conducted in the field of run time supports and compilation methods for irregular problems [18, 16, 10, 12, 15, 13, 14]. In our opinion, many of these techniques can be also adopted when load imbalance derives from the use of a time shared or heterogeneous parallel system. Other techniques have been specifically devised to solve the problem of load imbalance when this is introduced into uniform data parallel ....

[Article contains additional citation context not shown here]

S. Orlando and R. Perego. SUPPLE: an Efficient Run--Time Support for Non--Uniform Parallel Loops. Technical Report TR-17/96, Dipartimento di Mat. Appl. ed Informatica, Universit`a di Venezia, Dec. 1996.

....collected at run time during a previous simulation time step. This means that, the first time a given unbalanced loop is executed, the distribution is the statically chosen block one, since no information has yet been collected; ffl the second hybrid approach exploits the SUPPLE support [13, 14, 15], which always adopts a statically fixed block distribution. During execution, chunks of iterations and associated data may be dynamically moved toward underloaded processors. In this case both the time instants when a possible migration may occur, and the actual migration plan, are decided ....

....thus jeopardizing part of the benefits of subsequent dynamic load balancing actions if the workload is very unbalanced and the parallel loop is executed only a few times. The SUPPLE approach SUPPLE is a run time support for the implementation of both uniform and non uniform parallel loops [13, 14, 15]. It only supports block distributed arrays to exploit the locality that derive from regular stencil references. The innovative feature of SUPPLE is its efficient support for non uniform loops such as the Reaction loop in the flame simulation code. For this purpose, SUPPLE exploits an innovative ....

S. Orlando and R. Perego. SUPPLE: an Efficient Run--Time Support for Non--Uniform Parallel Loops. Technical Report TR-17/96, Dipartimento di Mat. Appl. ed Informatica, Universit`a di Venezia, Dec. 1996.

....with on both parallel and distributed systems: the problem of system load imbalance due to heterogeneity and time sharing of resources. In previous works we considered imbalances introduced by non uniform data parallel computations run on homogeneous, distributed memory, MIMD parallel systems [16, 19, 17, 18]. Here we restrict our view once more to data parallel computations expressed by means of parallel loops, but we treat the dual problem where load imbalance is not a characteristic of the user code, but is caused by variations in capacities of processing node. Architectural heterogeneity is usual ....

....are proposed. These proposals focus on distributed memory architectures, but exploit centralized managers which can result in bottlenecks when many processors are used. Also the hybrid scheduling technique discussed in this paper has been previously applied to non uniform parallel loops [16, 19, 17, 18]. It is asynchronous and does not introduce centralization points, and can be used even if the parallel loop is executed only once. There are some other techniques that have been explicitly devised to solve the problem of load imbalance when this is introduced into uniform data parallel ....

[Article contains additional citation context not shown here]

S. Orlando and R. Perego. SUPPLE: an Efficient Run--Time Support for Non--Uniform Parallel Loops. Technical Report TR-17/96, Dipartimento di Mat. Appl. ed Informatica, Universit`a di Venezia, Dec. 1996.

....time step before entering the Reaction phase. However, the actual binding of redistributed data and computations is decided dynamically, on the basis of load information collected at run time. SUPPLE. The second hybrid approach exploits the SUPPLE support (SUPport for Parallel Loop Execution) [12, 14]. It adopts a statically fixed block distribution, but, during execution, chunks of iterations and associated data may be dynamically migrated toward underloaded processors. In this case both the time instants when a possible migration may occur, and the actual migration plan, are decided at ....

....a static schedule, thus jeopardizing part of the benefits of subsequent dynamic load balancing actions if the workload is very unbalanced and the parallel loop is executed only a few times. SUPPLE. SUPPLE is a run time support for the implementation of both uniform and non uniform parallel loops [12, 14]. It only supports block distributed arrays to exploit the locality that derives from regular stencil references. The innovative feature of SUPPLE is its efficient support for non uniform loops such as the Reaction loop in the flame simulation code. For this purpose, SUPPLE exploits a hybrid ....

S. Orlando and R. Perego. SUPPLE: an Efficient Run--Time Support for Non--Uniform Parallel Loops. Technical Report TR-17/96, Dip. di Mat. Appl. ed Informatica, Universit`a di Venezia, Dec. 1996.

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