R. Gupta and C. R. Hill, "A Scalable Implementation of Barrier Synchro- nization Using An Adaptive Combining Tree," International Journal of Parallel Programming, 18(3):161-180, June 1989.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....3(1) 101 137, Winter 1990. Squillante and Lazowska, 1990] M. S. Squillante and Eward D. Lazowska, Using Processor Cache Affinity Information in Shared Memory Multiprocessor Scheduling, Technical Report 89 06 01, Computer Science Department, University of Washington, February 1990. Tucker and Gupta, 1989] A. Tucker and A. Gupta, Process Control and Scheduling Issues for Multiprogrammed Shared Memory Multiprocessors, Proceedings of the 12th Symposium on Operating Systems Principles, pages 159 166, December 1989. Yew et al. 1987] P. C. Yew, N. F. Tzeng, and D. H. Lawtie, Distributing Hot Spot ....

R. Gupta and C. R. Hill, "A Scalable Implementation of Barrier Synchro- nization Using An Adaptive Combining Tree," International Journal of Parallel Programming, 18(3):161-180, June 1989.

....in distributed systems. For example, counting networks are used to implement ecient Fetch Increment counters [4] as well as linearizable counters [15] Furthermore, smoothing networks solve load sharing problems [25] while threshold networks provide solutions to barrier synchronization problems [11, 12]. For applications of balancing networks, see [4, 15, 16, 18, 21, 23] A limitation of balancing networks is that they can be accessed by tokens only. Recall that a token can be thought of as an increment operation issued by the process inserting the token into the network. Using tokens only, the ....

R. Gupta and C. R. Hill, \A Scalable Implementation of Barrier Synchronization Using an Adaptive Tree," International Journal of Parallel Programming, Vol. 18, No. 3, pp. 161-180, June 1989.

....is the current counter value and w is a fixed constant. Thus, the Read operation returns the approximate value of the counter to within the constant w. Threshold counters have a variety of potential uses, most obviously software barrier synchronization (see, for example, 11, Section 4.2. 5] or [6, 7]) Threshold counters are interesting because they can sometimes be implemented more efficiently than exact counters. The most obvious way to implement a shared counter, whether threshold or exact, is to use a single shared variable protected by a lock. However, such centralized data structures ....

R. Gupta and C. R. Hill, "A Scalable Implementation of Barrier Synchronization Using an Adaptive Tree," International Journal of Parallel Programming, Vol. 18, No. 3, pp. 161--180, June 1989.

....messages from the root to the member nodes, but the performance can be significantly improved by reducing the number of messages. For regular interconnection networks, many research efforts have been devoted to develop efficient implementations of barrier synchronization with either software [13, 14, 15, 16] or hardware support [17, 18, 19, 20, 21, 22] For instance, Xu, et al. 15] proposed a software tree approach for barrier synchronization in 1 Nodes, in this paper, actually mean PCs or workstations in a cluster system. 2 wormhole routed hypercube multicomputers. Tree based schemes perform ....

R. Gupta and C. R. Hill, "A Scalable Implementation of Barrier Synchronization Using an Adaptive Combining Tree," International Journal of Parallel Programming, Vol. 18, No. 3, pp. 161-180, June 1989.

.... Trivially masking Eventually correctable Masking 2 Stabilizing Uncorrectable Fail safe Intolerant Table 1 : Classification of faults and appropriate tolerances to them in barrier synchronization Finally, we note that our program can be systematically extended to deal with fuzzy barriers [1]. In particular, the transition from execute to success is the same as entering the barrier, and the transition from ready to execute is the same as leaving the barrier. It is therefore possible to allow a process can perform useful work between these two state transitions, which captures the ....

R. Gupta and C. R. Hill. A scalable implementation of barrier synchronization using an adaptive combining tree. International Journal of Parallel Programming, 18(3), 1989.

....Furthermore, decentralized algorithms for dynamically changing tree size (see for example the Reactive Diffracting Trees of Della Libera and Shavit [7] tend to be complex and require significant tuning efforts. This is not to say that there are no known adaptive combining schemes. Gupta and Hill [13] and Mellor Crummey and Scott [24] have devised adaptive combining tree barriers that support changing the layout of a combining tree on the fly. However, their constructions cannot be readily applied to implement general data structures. The reason is that in order to work they require two key ....

R. Gupta and C.R. Hill. A Scalable Implementation of Barrier Synchronization Using an Adaptive Combining Tree. Proceedings of the Third International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS III), pp. 54--63, 1989.

....of message passing protocols. Since dedicated hardware barrier trees are intrinsically parallel and have very low latency, they are usually an order of magnitude faster than software barriers. There exist numerous algorithms and methods in the literature for improved software barriers, including [1, 7, 10, 13, 14, 12, 20], but these are improvements on a mechanism that is inherently slow. Methods for masking the latency of barriers have also been proposed [5, 6] These methods hide the synchronization overhead as well as the time spent waiting for other processors to reach the barrier. They depend on being able to ....

R. Gupta and C. R. Hill. A scalable implementation of barrier synchronization using an adaptive combining tree. International Journal of Parallel Programming, 18:161--180, 1989.

....for Research and Technology Hellas. Preprint submitted to Elsevier Science 15 August makes use of a synchronization barrier, a distributed data structure ensuring that no process advances beyond a particular point in a computation until all processes have arrived at that point (see, e.g. [3,7,9]) Conventional multiprocessors mostly use centralized barriers, which often become the network bottleneck or hot spots in the shared memory. In a seminal paper, Aspnes et al. 2] suggest a completely different approach to such synchronization problems in the context of balancing networks, a new ....

R. Gupta and C. R. Hill, "A Scalable Implementation of Barrier Synchronization Using an Adaptive Tree," International Journal of Parallel Programming, Vol. 18, No. 3, pp. 161--180, June 1989.

....of message passing protocols. Since dedicated hardware barrier trees are intrinsically parallel and have very low latency, they are usually an order of magnitude faster than software barriers. There exist numerous algorithms and methods in the literature for improved software barriers, including [1, 5, 7, 11, 12, 10, 17], but these are improvements on a mechanism that is inherently slow. Methods for masking the latency of barriers have also been proposed [3, 4] These methods hide the synchronization overhead as well as the time spent waiting for other processors to reach the barrier. They depend on being able to ....

R. Gupta and C. R. Hill. A scalable implementation of barrier synchronization using an adaptive combining tree. International Journal of Parallel Programming, 18:161--180, 1989.

....for Research and Technology Hellas. Preprint submitted to Elsevier Science 15 August makes use of a synchronization barrier, a distributed data structure ensuring that no process advances beyond a particular point in a computation until all processes have arrived at that point (see, e.g. [3,7,9]) Conventional multiprocessors mostly use centralized barriers, which often become the network bottleneck or hot spots in the shared memory. In a seminal paper, Aspnes et al. 2] suggest a completely different approach to such synchronization problems in the context of balancing networks, a new ....

R. Gupta and C. R. Hill, "A Scalable Implementation of Barrier Synchronization Using an Adaptive Tree," International Journal of Parallel Programming, Vol. 18, No. 3, pp. 161--180, June 1989.

....algorithms for busy wait barrier synchronization execute in time linear in the number of synchronizing processes. This time can be made logarithmic in the number of processes by adopting algorithms based on trees or FFT like synchronization patterns. As an additional improvement, Gupta and Hill [5] have proposed an adaptive combining tree barrier that exploits non uniformity in inter barrier computation times: processes begin to leave the barrier in time logarithmic in the number of processes when all processes arrive at once, but in constant time after the arrival of the last process when ....

....node of the tree to elect the process that will continue up to the parent. Once the arrival phase of the barrier is complete, however, the processes waiting on any particular path to the root must be awoken in LIFO order. To address the lack of amortization in logarithmic barriers, Gupta and Hill [5] introduced the concept of an adaptive combining tree barrier. Each process arriving at an adaptive combining tree barrier performs a local modification to the tree that allows later arrivals to start their work closer to the root. Given sufficient skew in the arrival times of processes, the last ....

R. Gupta and C. R. Hill, "A Scalable Implementation of Barrier Synchronization Using An Adaptive Combining Tree," International Journal of Parallel Programming 18:3 (June 1989), pp. 161-180.

....then a check is made for a copy in the enclosing set, and the process is repeated. Dealing with asymmetries in the request distribution is also possible, although it complicates the algorithm significantly. A similar approach for a dynamic combining tree barrier is described by Gupta and Hill in [GH89] None of the possible techniques were implemented in this CHAPTER 4. ALGORITHMS AND IMPLEMENTATIONS 33 thesis however. 4.4 Choosing Shared Region Sizes and Options Two of the most important decisions for the programmer when working with shared regions is choosing the size of the region and ....

R. Gupta and C. R. Hill. A scalable implementation of barrier synchronization using an adaptive combining tree. International Journal of Parallel Programming, 18(3):161--180, June 1989.

....algorithms for barriers rely on a software combining tree to achieve O(logN) barrier latency. However, if barrier arrival times are skewed, the use of a combining tree to accumulate barrier arrivals leads to a higher latency than a simple, centralized counter. This observation led Gupta and Hill [20] to propose an adaptive combining tree barrier that adapts the shape of the combining tree to the arrival patterns of the participating processes. They show that their algorithm leads to improved time complexities. However, their analysis of the algorithm ignores the run time overhead of ....

....adaptivity. If processes arrive skewed in time, the length of time in between barrier episodes will be sufficiently long that the reduction in latency for detecting the last process is insignificant. The adaptive combining tree does however show a performance advantage when used as a fuzzy barrier [20]. In a fuzzy barrier, a process waiting at the barrier can perform some useful computation that does not rely on completion of the barrier. This shows that the main advantage of the adaptive combining tree barriers comes from allowing the waiting processes to perform more useful work while ....

Rajiv Gupta and Charles R. Hill. A Scalable Implementation of Barrier Synchronization Using an Adaptive Combining Tree. International Journal of Parallel Programming, 18(3):161--180, 1989.

....each type of fault. Based on the discussion in Section 7, the user could specify that the program should recover from correctable faults in an appropriate manner and abort in case of uncorrectable faults. Finally, we note that our program can be systematically extended to deal with fuzzy barriers [2]. In particular, the transition from execute to success is the same as entering the barrier, and the transition from ready to execute is the same as leaving the barrier. It is therefore possible to allow a process can perform some useful work between these two state transitions, which captures the ....

R. Gupta and C. R. Hill. A scalable implementation of barrier synchronization using an adaptive combining tree. International Journal of Parallel Programming, 18(3), 1989.

....Shared Memory Multiprocessors Michael L. Scott Computer Science Department University of Rochester Rochester, NY 14627 0226 scott cs.rochester.edu John M. Mellor Crummey Computer Science Department Rice University, P.O. Box 1892 Houston, TX 77251 johnmc cs.rice.edu Abstract In a previous article [5], Gupta and Hill introduced an adaptive combining tree algorithm for busy wait barrier synchronization on shared memory multiprocessors. The intent of the algorithm was to achieve a barrier in logarithmic time when processes arrive simultaneously, and in constant time after the last arrival when ....

....shared locations after the arrival of the last process before any process can continue. In a traditional centralized barrier, the last arriving process discovers that the barrier has been achieved in constant time (ignoring possible delay due to contention) To address this problem, Gupta and Hill [5] introduced the concept of an adaptive combining tree barrier. Each process arriving at an adaptive combining tree barrier performs a local modification to the tree that allows later arrivals to start their work closer to the root. Given sufficient skew in the arrival times of processes, the last ....

R. Gupta and C. R. Hill, "A Scalable Implementation of Barrier Synchronization Using An Adaptive Combining Tree," International Journal of Parallel Programming 18:3 (June 1989), pp. 161-180. 27

....among the intermediate values computed by the processes. A simple way to solve this problem makes use of a synchronization barrier, a distributed data structure ensuring that no process advances beyond a particular point in a computation until all processes have arrived at that point (see, e.g. [3,8,10]) Conventional multiprocessors mostly use centralized barriers, which often become the network bottleneck or hot spots in the shared memory. In a seminal paper, Aspnes et al. 2] suggest a completely different approach to such synchronization problems in the context of balancing networks, a new ....

R. Gupta and C. R. Hill, "A Scalable Implementation of Barrier Synchronization Using an Adaptive Tree," International Journal of Parallel Programming, Vol. 18, No. 3, pp. 161--180, June 1989.

....during as well as at the end of program segments. When synchronization only needs to ensure that no PE advances beyond a certain point and each PE knows in advance when it is going to be done with local and global computations, barrier style synchronization primitives can generally be applied [5, 11, 16, 21, 29]. However, in many applications synchronization points cannot be explicitly or statically placed in the program and the points in time local or global computations may require synchronization are data dependent. Termination detection achieves global synchronization for such scenarios. A major ....

R. Gupta and C.R. Hill. A scalable implementation of barrier synchronization using an adaptive combining tree. International Journal of Parallel Programming, 18:161--180, June 1989.

....to obtain its correct value. Although the optimistic execution of v causes extra work, all elements eventually attain their final values. The global synchronization for termination detection in line 3 of the optimistic algorithm should be implemented using a split phase or fuzzy barrier [5, 6, 10] so that work can continue while the barrier is being evaluated. This implementation imposes an O(1) time cost per barrier per processor, and a barrier latency of O(lg P ) We incur this latency penalty only when the algorithm terminates, which means that the total running time is only O(lg P ) ....

Rajiv Gupta and Charles R. Hill. A scalable implementation of barrier synchronization using an adaptive combining tree. International Journal of Parallel Programming, 18(3):161--180, June 89.

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