B. Lubachevsky. Synchronization barrier and related tools for shared memory parallel programming. In Proc. 1989.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....down to all of the leaves. On a machine with coherent caches and unlimited replication, we could replace the wakeup phase of our algorithm with a spin on a global ag. We explore this alternative on the Sequent in section 4.4. Tournament Barriers Hensgen, Finkel, and Manber [17] and Lubachevsky [26] have devised tree style tournament barriers. Conceptually, to achieve a barrier using these tournament algorithms, processors start at the leaves of a binary tree, as in a combining tree barrier with fan in two. One processor from each node continues up the tree to the next round of the ....

....where i 2 (mod 2 k 1 ) and j = i 2 . Processor i then drops out of the tournament and waits on a global ag for notice that the barrier has been achieved. Processor j participates in the next round of the tournament. Processor 0 sets a global ag when the tournament is over. Lubachevsky [26] presents a CREW (concurrent read, exclusive write) tournament barrier that uses a global ag for wakeup, similar to that of Hensgen, Finkel, and Manber. He also presents an EREW (exclusive read, exclusive write) tournament barrier in which each processor spins on separate ags in a binary wakeup ....

B. Lubachevsky. Synchronization barrier and related tools for shared memory parallel programming. In Proc. 1989.

....but we observe in [18] that simple reads and writes suffice. A combining tree barrier still spins on non local locations, but causes less contention than a centralized counter barrier, since at most k 1 processors share any individual variable. Hensgen, Finkel, and Manber [121 and Lubachevsky [17] have devised tree style tournament barriers. In their algorithms, processors start at the leaves of a binary tree. One processor from each node continues up the tree to the next round of the tournament. The winning processor is statically determined; there is no need for fetch and . In ....

B. Lubachevsky. Synchronization barrier and related tools for shared memory parallel programming. In Proceedings of the 1989.

....to obtain the lock. Each processor spins on a local memory location and the running processor passes the lock by modifying the test memory location of its successor. On traditional multiprocessors, efficient algorithms that achieve barrier synchronization in O(logP ) time have been developed [6, 10, 11]; on distributed memory systems, the number of messages required is at least P , where P is the number of processors participating in the barrier. Representative descriptions of several such algorithms may be found in [4, 6] and, for message passing systems, in [5] However, all the algorithms ....

B. Lubachevsky, Synchronization barrier and related tools for shared memory parallel programming. in Proceedings of the 1989 International Conference on Parallel Processing, Aug., 1989, II-175--II-179.

....machine and somewhat less efficient on a single bus architecture. Its performance is better than that of the combining tree barrier on a hierarchical bus architecture because of the reduction in the number of reads and writes performed. The MCS barrier does worse than the tournament barrier [46, 56] when the hierarchy contains more than two levels, because of sharing across clusters that cannot be avoided (see Figure 5.10) When the hierarchy contains less than two levels, the smaller execution path of the MCS algorithm results in better performance. The tournament barrier uses a binary 0 20 ....

B. Lubachevsky. Synchronization Barrier and Related Tools for Shared-Memory Parallel Programming. In Proceedings of the 1989 International Conference on Paralle l Processing, pages II--175--II--179, August 1989.

.... (which guarantee that no process continues past a given point in a computation until all other processes have reached that point) Of particular interest in recent years have been scalable synchronization algorithms, which employ backoff or distributed data structures to minimize contention [1, 9, 11, 19, 20, 25, 26, 30, 31, 32, 38, 44, 45]. The purpose of backoff is to reduce the frequency with which spinning processes access a common synchronization variable. The purpose of distributed data structures is to allow each process to spin on a separate, locally accessible variable. Unfortunately, busy waiting in user level code tends ....

....running time of O(p) for the barrier algorithm, which becomes unacceptable as the number of processors p grows large. Several researchers have shown how to solve these problems by building scalable barriers, with log depth tree or FFT like patterns of point to point notifications among processes [3, 11, 20, 25, 30, 32, 38, 45]. Unfortunately, the deterministic notification patterns of scalable barriers may require that processes run in a different order from the one chosen by the scheduler. The problem is related to, but more severe than, the preemption while waiting problem in FIFO locks. With a lock the scheduler may ....

B. Lubachevsky. Synchronization Barrier and Related Tools for Shared Memory Parallel Programming. In Proceedings of the 1989 International Conference on Parallel Processing, pages II:175--179, August 1989.

....is being brought to the machine and will allow performance studies on hardware accelerators for the different asynchronous protocols. In particular, we study the impact of global controller on adaptive protocols. Other work has begun on the bounded lag algorithm described by Lubachevsky in [6]. A first analysis shows an important use of global operations. With the previously presented global accelerators and considering their performance, we expect that fpgas will help to improve the simulation by an interesting ratio. The reconfigurability enables application specific hardware ....

Lubachevsky, B. Synchronization Barrier and Related Tools for Shared Memory Parallel Programming,. In Proceedings of Intl. Conference on Parallel Processing, (Pen State, USA, Aug. 1989), pp. "II--175--179".

....wait for their peers to arrive as well, even if they have other work they could be doing that does not depend on the arrival of those peers. To improve asymptotic latency, several barriers have been developed that run in time O (log P) Most use some form of tree to gather and scatter information [6, 10, 14, 16]; the butterfly and dissemination barriers of Brooks [2] and of Hensgen, Finkel, and Manber [6] use a symmetric pattern of synchronization operations that resembles an FFT or parallel prefix computation. The butterfly and dissemination barriers perform a total of O (P log P) writes to shared ....

....version. 1 ############################# 1 In the butterfly and dissemination barriers, no process knows that the barrier has been achieved until the very end of the algorithm. In a static tree barrier [14] and in the tournament barriers of Hensgen, Finkel, and Manber [6] and Lubachevsky [10], static synchronization orderings force some processes to wait for their peers before announcing that they have reached the barrier. In all of the tree based barriers, processes waiting near the leaves cannot discover that their peers have reached the barrier until processes higher in the tree ....

B. Lubachevsky, "Synchronization Barrier and Related Tools for Shared Memory Parallel Programming," Proceedings of the 1989 International Conference on Parallel Processing, August 1989, pp. II:175-II:179.

....Marks itself as present at the barrier (entry phase) 2. Waits for all other participating processors to arrive at the barrier. 3. After all participating processors have arrived, it proceeds past the barrier (exit phase) Many algorithms exist for performing barrier synchronization in software [81, 21, 58]. Careful implementation of some of these algorithms are found to scale well to large scale multiprocessors without the contention for synchronization operations, referred to as barrier interference, becoming a significant problem [89] Barrier algorithms can be distinguished [8] by three ....

B.D. Lubachevsky. Synchronization barrier and related tools for shared memory parallel programming. In Proceedings of the 1989 International Conference on Parallel Processing, volume 2, pages 175 -- 179, August 1989.

....dissemination barrier with only local spinning. and without coherent caches, the shared allnodes array would be scattered statically across the memory banks of the machine, or replaced by a scattered set of variables. 3. 4 Tournament Barriers Hensgen, Finkel, and Manber [19] and Lubachevsky [30] have also devised tree style tournament barriers. The processors involved in a tournament barrier begin at the leaves of a binary tree, much as they would in a combining tree of fan in two. One processor from each node continues up the tree to the next round of the tournament. At each stage, ....

....then drops out of the tournament and busy waits on a global flag for notice that the barrier has been achieved. Processor j participates in the next round of the tournament. A complete tournament consists of dlog 2 Pe rounds. Processor 0 sets a global flag when the tournament is over. Lubachevsky [30] presents a CREW (concurrent read, exclusive write) tournament barrier that uses a global flag for wakeup, similar to that of Hensgen, Finkel, and Manber. He also presents an EREW (exclusive read, exclusive write) tournament barrier in which each processor spins on separate flags in a binary ....

B. Lubachevsky. Synchronization barrier and related tools for shared memory parallel programming. In Proc. of the 1989 International Conference on Parallel Processing, pages II-- 175--II--179, Aug. 1989.

....refined version of hypothesis 1. The reformulated version of this strong hypothesis must examine the following assertion: The barrier code which implements an algorithm with complexity Omega Gamma N) must be replaced by a sub linear algorithm. In particular the O(logN) barrier algorithm given in [5] should be employed. The conclusion from this performance diagnosis case study is that the primary bottleneck for the performance problem is the granularity of parallel computation, and the secondary bottleneck is the implementation of the barrier code. 3 Chitra This section overviews the three ....

Lubachevsky, B., "Synchronization Barrier and Related Tools for Shared Memory Parallel Programming," International Journal of Parallel Programming, vol. 19, no. 3,226-250, July. 1990.

....have reached the barrier) then the barrier algorithm must dynamically identify one processor out of the subset of processors. Previous work on barrier synchronization has focusedon specific synchronization hardware [13, 2, 14, 8, 9, 11] and on algorithms for synchronizing common memory systems [1, 12, 6, 10, 15]. In this paper we investigate subset barrier synchronization in the context of private memory systems without specific synchronization hardware. We introduce two general models for communication systems: the bounded buffer broadcast model and the anonymous destination message passing model. We ....

LUBACHEVSKY, B. Synchronization barrier and related tools for shared memory parallel programming. International Journal of Parallel Processing 19, 3 (1990), 225--250.

.... (which guarantee that no process continues past a given point in a computation until all other processes have reached that point) Of particular interest in recent years have been scalable synchronization algorithms, which employ backoff or distributed data structures to minimize contention [2, 10, 11, 18, 19, 25, 26, 30, 31, 35, 42, 43]. Unfortunately, busy waiting in user level code tends to work well only if each process runs on a separate physical processor. If the total number of processes in the system exceeds the number of processors, them some processors will have to be multiprogrammed. The processes on a given processor ....

....based on centralized counters. As with centralized locks, contention for counters creates serious performance problems on large machines. Scalable barriers, generally based on log depth tree or FFT like patterns of point to point notifications among processes, have received significant attention [4, 11, 19, 25, 30, 35, 43]. Comparatively little work has addressed the interaction of scheduling and barriers, possibly because barrier based applications tend to be used with a one process per processor system model. Much of the work on spin then block techniques (e.g. that of Ousterhout [33] and of Lim and Agarwal [22] ....

LUBACHEVSKY, B. Synchronization Barrier and Related Tools for Shared Memory Parallel Programming. In Proceedings of the 1989 InternationalConference onParallel Processing, pages II:175--179, August 1989.

....for their peers to arrive as well, even if they have other work they could be doing that does not depend on the arrival of those peers. To improve asymptotic latency, several barriers have been developed that run in time O ( log P) Most use some form of tree to gather and scatter information [7, 13, 15, 19]; the butterfly and dissemination barriers of Brooks [2] and of Hensgen, Finkel, and Manber [7] use a symmetric pattern of synchronization operations that resembles an FFT or parallel prefix computation. The butterfly and dissemination barriers perform a total of O (P log P) writes to shared ....

....above has such an obvious fuzzy version. In the butterfly and dissemination barriers, no process knows that all other processes have arrived until the very end of the algorithm. In a static tree barrier [15] and in the tournament barriers of Hensgen, Finkel, and Manber [7] and Lubachevsky [13], static synchronization orderings force some processes to wait for their peers before announcing that they have reached the barrier. In all of the tree based barriers, processes waiting near the leaves cannot discover that the barrier has been achieved until processes higher in the tree have ....

B. Lubachevsky, "Synchronization Barrier and Related Tools for Shared Memory Parallel Programming," Proceedings of the 1989 International Conference on Parallel Processing II (August 1989), pp. 175-179.

No context found.

B. Lubachevsky, "Synchronization Barrier and Related Tools for Shared Mem- ory Parallel Programming," In Proceedings of the 1989 International Conference on Parallel Processing, pages II:175-179, August 1989.

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