J. P. Singh, W-D. Weber and A. Gupta, SPLASH: Stanford Parallel Applications for Shared-Memory, Technical Report, Department of Computer Science, Stanford University, 1991.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....data locality. Even though shared memory machines have support for coherence, good performance requires locality of reference because of the memory hierarchy. Synchronization and true sharing must be minimized [1] Efficient parallelized codes synchronize infrequently and have little true sharing [22]. Therefore, a good parallelization requires no communication whenever possible. Using the data decomposition, which allows lower communication cost, also improves the data locality. The numerical results in Section 5 are obtained in a shared memory architecture. The performance of the parallel ....

J. Singh, W. Weber, and A. Gupta, SPLASH: Stanford parallel applications for shared-memory, Comput. Arch. News 20, 5 (1992).

....debugging. 3.1 Benchmarks Methodology We used five di#erent shared memory parallel applications for our study. The description of the applications is given in Table 1. Out of these, FFT and LU are from SPLASH2 benchmark applications suite [22] Mp3d and Water are from SPLASH benchmark suite [21], and SOR is from TreadMarks benchmark applications [9] The inputs used to generate the execution traces are also shown in Table 1. To collect execution traces, we ran each of these applications on RSIM 1.0 simulator [8] RSIM is an executiondriven simulator that models shared memory ....

J.P. Singh, W-D. Weber, and A. Gupta. SPLASH: Stanford parallel applications for shared-memory. ACM SIGARCH Computer Architecture News, 20(1), 1992.

....at the core of many important scienti c computations that simulate physical systems. The importance of such applications is likely to increase in the future. Later in this section we brie y describe the applications that we have used. The actual descriptions of the applications can be found in [20, 22, 23, 26]. 3.1 Problem Size Problem size is a very important issue. Generally, the larger the problem size the lower the frequency of synchronisation relative to computation. On one hand, using large problem sizes will therefore make synchronisation operations seem less important. On the other hand, ....

....6 Radiosity computes the equilibrium distribution of light in a scene using the iterative hierarchical di use radiosity method [5] The structure of the computation and the access patterns to data objects are highly irregular. Water Nsquared is an improved version of the Water program in SPLASH [23]. This application evaluates forces and potentials that occur over time in a system of water molecules. A process updates a local copy of the particle accelerations as it computes them, and accumulates into the shared copy once at the end. Water Spatial solves the same problem as Water Nsquared, ....

J. P. Singh, W. D. Weber and Anoop Gupta, SPLASH: Stanford Parallel Applications for Shared-Memory, Computer Architecture News, 20(1), pp. 2-12, March 1992.

....in the design of all layers of protocols. Studies of the typical distribution of TCP IP Internet traffic indicate that bi modal traffic is a common trait. The traffic typically consists of either small packets or big packets with little traffic in between [22,23] The studies based on the SPLASH [24] multiprocessor benchmarks for typical Distributed Shared Memory (DSM) applications also indicate similar bi modal traffic patterns [8] Based on this, we define two classes of traffic: Class A: small amounts of data, such as control information. Class B: large amounts of data. In the case of DSM ....

....by the cache coherence protocol and the amount of sharing in the application programs. A simulation study of snooping based and directory based cache coherence schemes was used to estimate fl. The FFT and SPEECH multiprocessor traces [25] were used with traces from the SPLASH benchmark suite [24] to estimate fl. Snooping based coherence schemes, which generally require a broadcast facility for requests and invalidations, used with the mp3d and water SPLASH benchmarks had a value of fl 0:3. Directory based coherence schemes, studied with the FFT and SPEECH traces, usually use explicit ....

J. P. Singh, W. D. Weber, and A. Gupta, "Splash: Stanford parallel applications for shared memory," Tech. Rep. CSL-TR-91-469, Stanford University, 1991.

....for any potential stopper. Therefore, if a program has not any potential stopper (which is a relatively simple task) we can say it is deadlock free. Nevertheless, despite there are many potential stopper free programs for which this last result is relevant (see for instance the SPLASH testbed [9]) there are other programs for which this situation do not apply. In that case, it is necessary to check whether the potential stoppers are in fact stoppers by identifying if the op operations in Def. 4 are contemporary, and this is not a simple task. At this point, we are currently focusing ....

J. Singh, W. Weber, and A. Gupta. SPLASH: Stanford parallel applications for shared memory. Computer Architecture News, 20(1):5-44, March 1992.

....on system performance and in identifying the performance advantages of either protocol. A simulation study of snooping based and directory based cache coherence scheme was used to estimate fl. The FFT and SPEECH multiprocessors traces [17] were used with traces from the SPLASH benchmark suite [18] used to estimate fl. Snooping based coherence schemes, which generally require a broadcast facility for requests and invalidations, used with the mp3d and water SPLASH benchmarks had a value of fl 0:3. Directory based coherence schemes, studied with the FFT and SPEECH traces, usually use ....

J. P. Singh, W. D. Weber, and A. Gupta, "Splash: Stanford parallel applications for shared mermory," Tech. Rep. CSLTR -91-469, Stanford University, 1991.

....the performance, scalability, and resource usage of full map and hierarchical directory based cache coherence protocols for parallel applications on COMA machines with up to 256 processors. Comparisons are made using a mixedmode simulator with applications from the Stanford SPLASH benchmark suite [16]. Our results are more accurate than those found in [17] as we model the contention for directory service and dynamically execute synchronization references with feedback from the memory system. Our results show that the choice between full map and hierarchical schemes for COMA machines is not ....

....and Full Map as the base protocols. After experimenting, we picked the delay1 function to be 5(N Gamma how early) and delay2 as 2 num failures . 4.2 Applications The three applications we use in our study are MP3D, Water, and LocusRoute programs. They are taken from the SPLASH benchmark suite [16]. All three programs exhibit significant parallelism of the type that fits our programming model. 4.2.1 MP3D The MP3D [10, 16] application is a three dimensional particle simulator that is used in aeronautics to solve a problem in rarefied fluid flow simulation. The program studies the motions ....

[Article contains additional citation context not shown here]

J. P. Singh, W. Weber, and A. Gupta, "SPLASH: Stanford parallel applications for shared-memory," Computer Architecture News, pp. 5 -- 44, March 1992.

....impact on system performance and in identifying the performance advantages of either protocol. A simulation study of snooping based and directory based cache coherence schemes was used to estimate . The FFT and SPEECH multiprocessor traces [4] were used with traces from the SPLASH benchmark suite [12] to estimate . Snooping based coherence schemes [16] which generally require a broadcast facility for requests and invalidations, used with the mp3d and water SPLASH benchmarks had a value of 0:3. Directory based coherence schemes [16] studied with the FFT and SPEECH traces, usually use ....

SINGH, J. P., WEBER, W. D., AND GUPTA, A. Splash: Stanford parallel applications for shared memory. Tech. Rep. CSL-TR-91-469, Stanford University, 1991.

....the LRC model which is somewhat different from the EC or the ScC model. The application programs used for performance comparison are Barnes, Integer Sort (IS) Traveling Salesman Problem (TSP) 3 dimensional FFT, WATER 2N, and WATER SPATIAL. Barnes, WATER2N and WATER SPATIAL are from SPLASH suite [28, 29]; and the rest of the application programs are 16 from TreadMarks test suite [30] Table 2 shows the input parameters and the synchronization operations used for each application program. The number of locks do not include the local lock transfers. All the application programs have at least 7 ....

J.P. Singh et al., "SPLASH: Stanford parallel applications for shared-memory," in Tech. Rep. CSLTR -91-469, Stanford University, Apr. 1991.

....on system performance and in identifying the performance advantages of either protocol. A simulation study of snooping based and directory based cache coherence scheme was used to estimate fl. The FFT and SPEECH multiprocessor traces [13] were used with traces from the SPLASH benchmark suite [14] used to estimate fl. In the study later, we examine protocol performance for fl 2 f0:3; 0:9g. Transmission in HRP is organized into cycles where each cycle consists of a control phase and a data phase. The control phase operates in a broadcast environment. A control packet sent during the RCB ....

J. P. Singh, W. D. Weber, and A. Gupta, "Splash: Stanford parallel applications for shared memory," Tech. Rep. CSLTR -91-469, Stanford University, 1991.

....CHAN ID RSVD CHAN ID Figure 2: Format of control and data packets. cache control traffic. A simulation study of snooping based and directory based cachecoherenceschemeswas used to estimate fl. The FFT and SPEECH multiprocessor traces [13] were used with traces from the SPLASH benchmark suite [14] to estimate fl. Snoopingbased coherenceschemes,which generally require a broadcast facility for requests and invalidations, used with the mp3d and water SPLASH benchmarks had a value of fl 0:3. Directory based coherence schemes, studied with the FFT and SPEECH traces, usually use explicit ....

J. P. Singh, W. D. Weber, and A. Gupta, "Splash: Stanford parallel applications for shared mermory," Tech. Rep. CSL-TR-91-469, Stanford University, 1991.

....these stages, instead of using two separate input and output array, we interleave these arrays to avoid large number of conflict misses. Though, this algorithm is not optimal for cache based systems [16] but for given problem size, number of processors and cache size, it performs fairly. MP3D [17] is a three dimensional particle simulator used in rarefied fluid flow simulation. We used 16000 molecules with the default geometry provided with SPLASH [17] which uses a 14 Theta 24 Theta 4 (2646 cell) space containing a single flat sheet placed at an angle to the free stream. The simulation ....

....is not optimal for cache based systems [16] but for given problem size, number of processors and cache size, it performs fairly. MP3D [17] is a three dimensional particle simulator used in rarefied fluid flow simulation. We used 16000 molecules with the default geometry provided with SPLASH [17] which uses a 14 Theta 24 Theta 4 (2646 cell) space containing a single flat sheet placed at an angle to the free stream. The simulation was done for 5 time steps. In this application, the active space is divided into cells and the molecules interact only with the molecules in the same cell. ....

J. P. Singh, W.-D. Weber, and A. Gupta, "SPLASH: Stanford Parallel Applications for Shared-Memory," ACM SIGARCH Computer Architecture News, vol. 20, no. 1, , March 1992. 22

....including the cache coherence protocol and the network. This simulator is an extension of Proteus, which is an execution driven simulator [11] Five different application programs are used to evaluate the performance of the schemes under study. These are MP3D and WATER from SPLASH benchmarks [12], and matrix multiplication, Floyd Warshall s all pair shortest path algorithm and FFT developed by us. The results indicate that the proposed scheme performs close to or even better than the full map directory scheme for all the applications. The rest of the paper is organized in the following ....

....hand,sharing is distributed over the whole range in Floyd Warshall. In MP3D and Water, too, most of the writes are to the blocks that are shared among a few processors. But in MP3D the number of widely shared blocks is higher and also percentage of sharing is much higher in MP3D than in Water [12]. Matrix multiplication has no sharing for shared writable blocks. In our implementation, the output matrix is divided into rectangular blocks and one block is assigned to each processor. The processors share the rows and columns of the input matrices but there are no writes to it, so, it does not ....

[Article contains additional citation context not shown here]

J. P. Singh, W. D. Weber, and A. Gupta, "Splash: Stanford parallel applications for shared memory," ACM SIGARCH Computer Architecture News, vol. 20, , March 1992.

....of the dynamic sub block scheme to reduce the false sharing miss ratio, extensive simulations were done on benchmark applications using the complete block scheme, the fixed size sub block scheme and the dynamic sub block scheme. Two of these applications, MP3D and Water are from the SPLASH [13] benchmark suite. These applications were chosen to evaluate the performance of the dynamic sub block schemes on applications that exhibit large amounts of true and false sharing. Floyd and Jacobi [14] were chosen to evaluate the the performance of the dynamic sub block scheme when read sharing is ....

J.P.Singh, W-D. Weber and A. Gupta "SPLASH: Stanford Parallel Application for Shared Memory. ", Technical Report, Dept. Of Computer Science, Stanford University, April 1991.

....release operations. After releasing the lock, the processor waits a minimum random interval before proceeding to ensure another processor has an opportunity to acquire the lock before a successive local lock re acquire, thus reducing unfairness. We use barnes, cholesky, and mp3d from SPLASH [34] and radiosity, water nsq, ocean cont, and raytrace from SPLASH2 [39] A locking version of mp3d is used to study the impact of TLR on a lock intensive benchmark [16] This version of mp3d does frequent synchronization to largely uncontended locks and lock access latencies cannot be hidden by a ....

J. P. Singh, W.-D. Weber, and A. Gupta. SPLASH: Stanford parallel applications for shared memory. Computer Architecture News, 20(1):5--44, Mar. 1992.

....Table 1: Characteristics of the applications used in the evaluation. we interleave these arrays to avoid large number of conflict misses. Though, this algorithm is not optimal for cache based systems [20] but for given problem size, number of processors and cache size, it performs fairly. MP3D [21] is a three dimensional particle simulator used in rarefied fluid flow simulation. We used 16000 molecules with the default geometry provided with SPLASH [21] which uses a 14 Theta 24 Theta 4 (2646 cell) space containing a single flat sheet placed at an angle to the free stream. The simulation ....

....is not optimal for cache based systems [20] but for given problem size, number of processors and cache size, it performs fairly. MP3D [21] is a three dimensional particle simulator used in rarefied fluid flow simulation. We used 16000 molecules with the default geometry provided with SPLASH [21] which uses a 14 Theta 24 Theta 4 (2646 cell) space containing a single flat sheet placed at an angle to the free stream. The simulation was done for 5 time steps. In this application, the active space is divided into cells and the molecules interact only with the molecules in the same cell. ....

J. P. Singh, W.-D. Weber, and A. Gupta, "SPLASH: Stanford Parallel Applications for Shared-Memory," ACM SIGARCH Computer Architecture News, vol. 20, no. 1, , March 1992. 22

....virtual memory mode implied that all pages were loaded before computation started (roll in) and there was no disk access thereafter. Application characteristics: Three types of applications, whose traces were obtained from the literature, were simulated: WATER, MP3D and CHOLESKY(SPLASH benchmark [15]) We consider two characteristics: PD access pattern and PD locality of reference. Access pattern is the percentage of PD references among all references. Locality is the probability that the next access to PD will be to the same page as the previous one. The WATER problem simulates the ....

....of reference. Access pattern is the percentage of PD references among all references. Locality is the probability that the next access to PD will be to the same page as the previous one. The WATER problem simulates the evolution of a system of water molecules, done through short range N body [15]. WATER presents no locality in terms of PD. Access to PD corresponds to 18 of all references [1] MP3D simulates rarefied fluid flow, done through particle in cell, Monte Carlo methods [15] MP3D presents racing conditions, some locality, and its access rate to PD is 40 [1] CHOLESKY ....

[Article contains additional citation context not shown here]

J. P. Singh, W.-D. Weber, and A. Gupta. SPLASH: Stanford Parallel Applications for Shared-Memory. SIGArch Computer Architecture News, 20(1), March 1992.

....study. In order to evaluate the benefits of our proposals, we have selected several scientific applications for which cache tocache transfer misses constitute an important percentage of the total miss rate (more than 25 in all the cases) MP3D and Water are from the SPLASH benchmark suite [22], FFT and Ocean are from SPLASH 2 benchmark suite [23] EM3D is a shared memory implementation of the Split C benchmark. Unstructured is a computational fluid dynamics application that uses an unstructured mesh. All experimental results reported in this paper are for the parallel phase of these ....

J. Singh, W.-D. Weber and A. Gupta. "SPLASH: Stanford Parallel Applications for Shared-Memory". Computer Architecture News, 20:5--44, March 1992.

....mainly for costly acquisition decisions, and has therefore been practiced for a long time. However, such evaluations typically focus on the computational aspects of the system. They typically use a small set of benchmark applications, and measure the performance of these applications in isolation [2, 3, 7, 1]. The results re ect the performance of the processor, the memory hierarchy, the interconnection network, and the relationship between these factors. ESP is di erent it targets the system level performance rather than the hardware [8] Issues include the eciency of the scheduling, its ....

J. P. Singh, W-D. Weber, and A. Gupta, \SPLASH: Stanford parallel applications for shared-memory". Comput. Arch. News 20(1), pp. 5-44, Mar 1992.

....tunnel and a solid object placed inside the tunnel. The tunnel is represented as a 3D space array of unit sized cells. Particles move through the space array and can only collide with particles occupying the same cell in the same time step. A complete description of this program can be found in [21]. Ricardo Bianchini has done a study on the performance of large parallel applications on Alewife. Ricardo s study includes experimentation with multiple implementations of Mp3d. In this chapter, we have used three different implementations of Mp3d and run each on a 16 node Alewife machine, with ....

J. P. Singh, W. D. Weber, and A. Gupta. SPLASH: Stanford Parallel Applications for Shared-Memory. Computer Architecture News, 20(1):5--44, March 1992.

....3 Iterations Matrix Multiply 384 x 384 Matrices Gauss 512 x 512 Matrix FFT 32K Elements TSP 10 CityTour Water 343 Molecules, 2 Iterations Barnes Hut 2K Bodies, 3 Iterations Unstructured 2800 Nodes, 17377 Edges, 1 Iteration Table 1: List of applications and their problem sizes. SPLASH [12] benchmark suite. Barnes Hut is also taken from the SPLASH benchmark suite. Finally, Unstructured is a computation over an unstructured mesh from the University of Wisconsin, Madison, and the University of Maryland, College Park [10] Figure 8 through Figure 15 present performance results of the ....

J.P. Singh, W.-D. Weber, and A. Gupta. SPLASH: Stanford Parallel Applications for Shared-Memory. Technical Report CSL-TR-92-526, Stanford University, June 1992.

....generated topologies are not likely to capture specific behaviors without careful intervention and design in the topology generator. Such generators will be useful, however, in testing problems of scale. It is tempting to look to existing architectural benchmarks such as SPEC [25] or SPLASH [24] as a model. However, although such benchmarks have been proposed in the past (c.f. 16] there are several compelling reasons why a network simulation suite should not be a benchmark intended to yield a single number. First, unlike processor design where the end result must be a single ....

J. P. Singh, W.-D. Weber, and A. Gupta. SPLASH: Stanford parallel applications for shared-memory. Technical Report CSL-TR-91-469, Stanford University, April 1991.

....independently, also implemented data oriented statistics. Data oriented statistics are especially useful in cases in which a particular data structure may constitute a memory bottleneck, but accesses to it are distributed across several procedures. For example, in Pthor (a SPLASH benchmark [12]) the EleraentArray is responsible for more of the program s cache misses than any other variable, but these misses are distributed across several procedures. Code oriented output cannot emphasize Eleraont.lrray s performance problems as well as data oriented output can, because no single section ....

....MemSpy has also been used to tune several other programs as well. For example, it has identified performance bugs due to: i) false sharing and a vestigial (incremented but unused) variable in Locusl oute, a SPLASH benchmark, ii) self interference in the ElementArray in Pthor [12], iii) poor spatial locality in a sequential volume rendering program, Vrender [7] and (iv) shared accesses to a private variable in a parallel version of Vrender. Together these experiences have confirmed the utility of these detailed, data oriented statistics. 5 MemSpy Implementation In ....

J.P. Singh, W.-D. Weber, and A. Gupta. SPLASH: Stanford Parallel Applications for Shared-Memory. Computer Architecture News, 20(1):5-44, March 1992.

No context found.

J. P. Singh, W-D. Weber and A. Gupta, SPLASH: Stanford Parallel Applications for Shared-Memory, Technical Report, Department of Computer Science, Stanford University, 1991.

No context found.

J. P. Singh, W.-D. Weber, and A. Gupta, "SPLASH: Stanford parallel applications for shared-memory," Computer Systems Laboratory, Stanford University, Technical Report CSL-TR-91-469, April 1991.

No context found.

J.P. Singh, W.-D. Weber, and A. Gupta. SPLASH: Stanford Parallel Applications for Shared-Memory. Stanford University, Technical Report CSL-TR-91-469, April 1991.

No context found.

J. P. Singh, W. Weber, and A. Gupta. SPLASH: Stanford Parallel Applications for SharedMemory. Technical Report, Computer Systems Laboratory, Stanford University, 1991.

No context found.

P. Singh, W. Weber, and A. Gupta. SPLASH: Stanford parallel applications for sharedmemory. Computer Architecture News, 20(1):5-44, March 1992.

No context found.

J. P. Singh, W. Weber, and A. Gupta. SPLASH: Stanford Parallel Applications for Shared-Memory. Computer Architecture News, 1992.

No context found.

J. P. Singh, W. Weber, and A. Gupta. Splash: Stanford parallel applications for shared-memory. Computer Architecture News, 20(1).

No context found.

Singh, J.P., Weber, W.-D., and Gupta, A. "SPLASH: Stanford Parallel Applications for Shared-Memory," in Computer Architecture News, 20(1):5-44, March 1992.

No context found.

J. Singh, W. Weber, and A. Gupta. SPLASH: Stanford parallel applications for shared memory. Computer Architecture News, 20(1):5--44, March 1992.

No context found.

J.P. Singh, W.-D. Weber and A. Gupta. SPLASH: Stanford Parallel Applications for Shared-Memory. Computer Architecture News, vol. 20, pp. 5-44, March 1992.

No context found.

J. P. Singh, W.-D. Weber, and A. Gupta. SPLASH: Stanford parallel applications for shared memory. Computer Architecture News, 20(1):5--44, Mar. 1992.

No context found.

J. Singh, W. Weber, and A. Gupta. SPLASH: Stanford parallel applications for shared memory. Computer Architecture News, 20(1):5--44, March 1992.

No context found.

Singh, J. P.; Weber, W.; Gupta, A,; "SPLASH: Stanford Parallel Applications for SharedMemory " Computer Architecture News, vol. 20, no. 1

No context found.

J.P. Singh, W. Weber, A. Gupta. Splash: Stanford Parallel Applications for Shared Memory. Computer Architecture News, 20(1), 1992.

No context found.

J. P. Singh et al. SPLASH: Stanford Parallel Applications for Shared-Memory. Computer Architecture News, March 1992.

No context found.

J. Singh, W.-D. Weber and A. Gupta. "SPLASH: Stanford Parallel Applications for Shared-Memory". Computer Architecture News, 20:5--44, March 1992.

No context found.

J. P. Singh, W. Weber, A. Gupta, "SPLASH: Stanford Parallel Applications for Shared-Memory", Computer Architecture News, vol. 20, no. 1, March 1992. pp. 20(1):5-44.

No context found.

J. P. Singh, W. D. Weber, and A. Gupta. "SPLASH: Stanford Parallel Applications for Shared Memory", Computer Architecture News, 20(1):5-44, March 1992.

No context found.

J.P. Singh, W.-D. Weber and A. Gupta, \SPLASH: Stanford Parallel Applications for SharedMemory ", Computer Architecture News, Vol. 20, No. 1, pp. 5-44, March, 1992.

No context found.

Singh, J. P., Weber, W., and Gupta, A., "SPLASH: Stanford parallel applications for shared-memory," Computer Architecture News, Vol. 20, pp. 5--44, March 1992.

No context found.

P.Singh, W. Weber and A. Gupta. SPLASH: Stanford Parallel Applications for Shared-memory. Computer Architecture News, 20(1):5--44, March 1992.

No context found.

J. P. Singh, W.-D. Weber, and A. Gupta. Splash: Stanford parallel applications for shared-memory. CAN, 20(1):pp. 5--44, March 1992.

No context found.

J. P. Singh, W.-D. Weber, and A. Gupta. SPLASH: Stanford parallel applications for shared-memory. Technical Report CSL-TR-91-469, Stanford University, April 1991.

No context found.

P. Singh, W. Weber and A. Gupta. SPLASH: Stanford Parallel Applications for Shared-memory. Computer Architecture News, 20(1):5-44, March 1992. 13

No context found.

P. Singh, W. Weber, and A. Gupta. SPLASH: Stanford parallel applications for shared-memory. Computer Architecture News, 20(1):5--44, Mar. 1992.

No context found.

J. Singh, W. Weber, and A. Gupta, "SPLASH: Stanford Parallel Applications for Shared-Memory", ACM SIGARCH Computer Architecture News, 22(4), 1992, pp. 544.

No context found.

Singh, J., Weber, W., and Gupta, A.: SPLASH: Stanford Parallel Applications for Shared Memory. Computer Architecture News, Vol. 20, No. 1 (1992) 5-44

First 50 documents  Next 50

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