X. Martorell, J. Labarta, J.I. Navarro, and E. Ayguad'e. Analysis of several scheduling algorithms under the nano--threads programming model. In 11th IEEE International Parallell Processing Symposium (IPPS'97), 1997.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....or directly by the programmer, who may provide scheduling hints to the runtime system. This approach increases flexibility, since user level scheduling is often dependent on the application characteristics. Several scheduling algorithms for parallel loops implemented in Nthlib, can be found in [Mart97, Poly97]. 3 Thread Management Alternatives In this section we introduce two thread management techniques which we have implemented in NthLib. The first technique performs efficient memory management in the runtime system, by exploiting the affinity of nano threads for processors. The proposed mechanism ....

....the generated loops locally in the machine is beneficial. The same holds for tasks T 2 ; T 3 ; T 5 ; and T 8 . Repeated execution of the task graph allows tasks to establish affinity for certain processors. For the parallelization of loops we use the adaptive size chunking algorithm proposed in [Mart97]. This algorithm executes loops with bursts of nano threads. In the implementation for the Base and Stack aff alternatives, loop parallelization is done using all the available processors. In the Hier version, parallelization is done by partitioning the machine and correspondingly the queues ....

X. Martorell, J. Labarta, N. Navarro and E. Ayguade, Analysis of Several Scheduling Algorithms under the Nano-threads Programming Model, Proceedings of the 11th International Parallel Processing Symposium, pp. 281--287, Geneva, Switzerland, April 1997.

....endwhile endwhile end main 14 7. Kernel Interaction NthLib is tightly coupled with the operating system in order to adapt at user level when any kernellevel scheduling event alters the user level execution. NthLib uses the kernel interface described in [M1D197] and evaluated in [Poly98] and [Mart97] in order to maintain as many user level threads running as the number of physical processors assigned by the operating system. The interaction with the operating system is done mainly by NthLib at every user level scheduling point, when a nano thread terminates execution and the package knows ....

....sections have been also parallelized. It has been automatically parallelized using a modified version of the Parafrase 2 compiler, which accepts a set of PCF like directives with some extensions to express multi level parallelism. The compiler has allowed us to test several scheduling algorithms [Mart97], such as interleave and dynamic [Mark92] Results of the experiments are presented in Table 6. In the table, the execution time of the application is shown for the sequential FORTRAN version, the FORTRAN parallelized version (labeled MP) and the nano threads versions using the interleave and ....

Xavier Martorell, Jesús Labarta, Nacho Navarro, Eduard Ayguadé, "Analysis of Several Scheduling Algorithms under the Nano-Threads Programming Model", Proc. of the 11th International Parallel Processing Symposium (IPPS `97), Switzerland, April 1-5, 1997.

....for the parallel loop using the interleaved scheduling policy. different scheduling policies (uniform size and guided self scheduling) assuming 3 processors. Experimental evaluation of different scheduling algorithms and their adaptability to varying system conditions has been presented in [MLNA97] Figure 4.c shows the generation of code for one of the loops in the working example. The nthf burst create and nthf burst wait primitives are used to control the start and the end of the loop parallelism when a barrier synchronization is required. The barrier at the end of the loop introduced ....

X. Martorell, J. Labarta, J.I. Navarro, and E. Ayguad'e. Analysis of several scheduling algorithms under the nano--threads programming model. In 11th IEEE International Parallell Processing Symposium (IPPS'97), 1997.

....of system resources. Figure 5 shows the generation of threads for two different scheduling policies (uniform size and guided self scheduling) assuming 3 processors. Experimental evaluation of different scheduling algorithms and their adaptability to varying system conditions has been presented in [MLNA97] Figure 4.b shows the generation of code for one of the loops in the same example. The nthf burst create and nthf burst wait primitives are used to control the start and the end of the loop parallelism when a barrier synchronization is required. The loop end barrier created by Nthf burst create ....

X. Martorell, J. Labarta, J.I. Navarro, and E. Ayguad'e. Analysis of several scheduling algorithms under the nano--threads programming model. In 11th IEEE International Parallell Processing Symposium (IPPS'97), 1997.

....Figure 4 shows the generation of threads for two different scheduling policies (uniform sized chunk scheduling and guided self scheduling) assuming 3 processors. Experimental evaluation of different scheduling algorithms and their adaptability to varying system conditions has been presented in [Mart97]. Figure 3.c shows the generation of Fortran code with calls to NthLib starting from user directives for the doubly nested loop in the third section of our working example. The scheduling of iterations corresponds to STATIC with a fixed chunk size of 16. The nthf burst create and nthf burst wait ....

X. Martorell, J. Labarta, J.I. Navarro, and E. Ayguad'e, Analysis of several scheduling algorithms under the nano--threads programming model, In Proceedings of 11th IEEE International Parallel Processing Symposium, 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