J. Nieh and M. Levoy, Volume Rendering on Scalable Shared Memory MIMD Architectures, in Proceeding of the 1992.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....3.2 Application Description Ocean simulates eddy currents in an ocean basin [27] Both its inherent and induced (at page granularity) data referencing patterns generally involve one producer with one consumer. Volrend renders three dimensional volume data into an image using a ray casting method [17]. The volume data are read only. Its inherent data referencing pattern on data that are written (task queues and image data) is migratory, while its induced pattern at page granularity involves multiple producers with multiple consumers. Both the read accesses to the read only volume and the write ....

J. Nieh and M. Levoy, Volume Rendering on Scalable Shared Memory MIMD Architectures, in Proceeding of the 1992.

....rendering algorithms. Both papers report on parallelization, which is the main focus of [6] Challinger [6] also used an image tiling scheme for parallelization with very good results. Other parallel techniques on shared memory machines include the results by Williams [59] Nieh and Levoy [36], and Lacroute [27, 28] 2.3 Out Of Core Techniques We now briefly review the work on out of core techniques. For theoretical results on out of core algorithms for graphs and for computational geometry, we refer to the recent survey by Vitter [54] Teller et al. 50] described a system to ....

J. Nieh and M. Levoy. Volume rendering on scalable shared-memory MIMD architectures. In Workshop on Volume Visualization, pages 17--24. ACM Press, October 1992.

....3.2 Application Description Ocean simulates eddy currents in an ocean basin [27] Both its inherent and induced (at page granularity) data referencing patterns generally involve one producer with one consumer. Volrend renders three dimensional volume data into an image using a ray casting method [17]. The volume data are read only. Its inherent data referencing pattern on data that are written (task queues and image data) is migratory, while its induced pattern at page granularity involves multiple producers with multiple consumers. Both the read accesses to the read only volume and the write ....

J. Nieh and M. Levoy, Volume Rendering on Scalable Shared Memory MIMD Architectures, in Proceeding of the

.... model of the geology between two oil wells [Harris et al. 1990] Search: A program that simulates the interaction of electron beams with solids [Browning et al. 1994; Browning et al. 1995] Volume Rendering: A program that renders a three dimensional volume data set for graphical display [Nieh and Levoy 1992]. Panel Cholesky: A program that factors a sparse positive definite matrix [Rothberg 1993] Ocean: A program that simulates the role of eddy and boundary currents in influencing large scale ocean movements [Singh and Hennessy 1992] To the best of our knowledge, six people have ....

Nieh, J. and Levoy, M. 1992. Volume rendering on scalable shared-memory MIMD architectures. Tech. Rep. CSL-TR-92-537, Computer Systems Laboratory, Stanford Univ., Stanford, Calif. Aug.

....because of the way memory is laid out, it reads only a single word on each page. Thus, there is signicant fragmentation in communicating remote data in pages at column boundaries in shared memory systems. Volrend renders three dimensional volume data into an image using a ray casting method [16]. The volume data are read only. Its inherent data referencing pattern on data that are written (task queues and image data) is migratory, while its induced pattern at page granularity involves multiple producers with multiple consumers. Both the read accesses to the read only volume and the write ....

J. Nieh and M. Levoy, Volume Rendering on Scalable Shared Memory MIMD Architectures, in Proceeding of the 1992 Workshop on Volume Visualization, pp 17-24, October 1992.

....computations, leads to severe load imbalances. Traditional parallelization techniques therefore require explicitly balancing processor workloads, either by intelligent partitioning or dynamic work stealing (such as the SPLASH 2 implementations for volume rendering, radiosity and a ray tracing [NL92], SGL94] An XMT implementation would take one of the following approaches: 1) A synchronous approach, with a spawn at each stage of the computation. The first spawn block creates a thread for each preliminary unit of work (a primary ray, a node with in degree 0, etc. after a join, threads ....

J. Nieh, M. Levoy, "Volume Rendering on Scalable Shared-Memory MIMD Architectures," I Proceedings of the Boston Workshop on Volume Visualization, Boston, MA, October 1992.

....provide the ability to compare, contrast, and show potential for future research results. Keywords: volume visualization, splatting, ray casting, taxonomy 1 References [98] 49] 16, 48, 113] 44] 52] 107] 111, 81] 94] 16, 15] 34] 101] 40] 66] 73] 90] 116] 26] 20] [68] [6] 54, 52, 63, 28, 67] 7, 1, 42, 4, 62] 31, 75, 30, 38, 86] 41, 21, 119, 112, 118, 45, 58] 55, 65, 14, 29, 79] 115, 53, 70, 74, 87, 19, 117, 32] 72, 57, 92, 36, 7] 111, 12, 90, 97, 78, 5] 15, 27, 35, 43, 87, 28, 51, 22, 23, 59, 60, 64, 63, 66, 68, 71, 80, 81, 91, 93, 99, 47, 108] 103, ....

.... 113] 44] 52] 107] 111, 81] 94] 16, 15] 34] 101] 40] 66] 73] 90] 116] 26] 20] 68] 6] 54, 52, 63, 28, 67] 7, 1, 42, 4, 62] 31, 75, 30, 38, 86] 41, 21, 119, 112, 118, 45, 58] 55, 65, 14, 29, 79] 115, 53, 70, 74, 87, 19, 117, 32] 72, 57, 92, 36, 7] 111, 12, 90, 97, 78, 5] [15, 27, 35, 43, 87, 28, 51, 22, 23, 59, 60, 64, 63, 66, 68, 71, 80, 81, 91, 93, 99, 47, 108] [103, 116, 83, 106, 107, 105, 2, 76, 110, 77] 11, 104, 101, 96, 89, 26, 73, 6, 17, 52, 54, 85, 114, 9, 82, 37, 24] 8, 61, 3, 56, 67, 88, 44, 10, 69, 48, 16, 84, 113, 95, 13, 50, 94, 40, 46] 100, 109, 102] 25, 18, 33, 39] ....

J. Nieh and M. Levoy. Volume rendering on scalable shared-memory MIMD architectures. In Proceedings of 1992 Workshop on Volume Visualization, pages 17--24, October 1992.

....storage overhead. Linear speedup is parallel run time on processors that is a fraction of the best sequential run time , or . Storage efficiency is storage with sample points, equivalent to the sequential algorithm s storage. Previous parallel volume rendering algorithms have restricted platforms [7], data set sizes [17] 6] filter quality [9] 12] 2] 10] 3] view angle freedom, or correctness [13] Users shouldn t have to sacrifice functionality to achieve higher performance with parallel computers. We have developed a parallel algorithm that uses unrestricted viewpoints and filters, with ....

Nieh, Jason and Marc Levoy. Volume Rendering on Scalable Shared-Memory MIMD Architectures. In Proceedings of 1992 Workshop on Volume Visualization (Oct. 1992), 1724.

....follow a ray s path through the voxel space. No attempt is made at smart data distribution, as it is very difficult due to page granularity n 2 n DRAFT VERSION: PLEASE DO NOT DISTRIBUTE 5 restrictions and because the viewpoint changes across frames. This algorithm is described more fully in [NiL92]. Raytrace: This application renders a three dimensional scene onto a two dimensional image plane using optimized ray tracing. A hierarchical uniform grid (similar to an octree) is used to represent the scene for efficient access, and early ray termination and antialiasing are implemented. Rays ....

Jason Nieh and Marc Levoy, "Volume Rendering on Scalable Shared-Memory MIMD Architectures", In Proceedings of the Boston Workshop on Volume Visualization, October 1992.

....a shared bus. These two machines are capable of rendering the 256 3 data set at 0.25 and 0.36 seconds per frame respectively. Using the same Dash Multiprocessor system, a data set of 256 Theta 256 Theta 226 size is rendered at 0. 34 seconds per frame using an adaptive 2 rendering technique [Nieh92] Using the custom hardware in a four Raster Manager RealityEngine [Akeley93] Onyx with a 150 MHz R4400 Central Processing Unit (CPU) a data set of 512 Theta 512 Theta 64 8 bit values 2 An adaptive rendering technique starts by first rendering a rough image from a smaller version of the ....

Jason Nieh and Marc Levoy. Volume rendering on scalable sharedmemory MIMD architectures. In Workshop on Volume Visualization, pages 17--24. ACM, October 1992.

....of the voxel representation 37 and the pixel raster projection lend themselves to parallel organization and parallel computation. In the past year several parallel implementations of volume rendering have been reported using either a shared memory (e.g. ray casting on the DASH architecture [NL92] a ray casting and splatting on a BBN TC2000 [Cha91] a distributed memory architecture (e.g. ray casting on an nCUBE 6410 [CM92] or a SIMD machine (e.g. ray casting on the Connection Machine CM2) The Cube project at Stony Brook, directed by Arie Kaufman, has been leading in the field of ....

J. Nieh and M. Levoy. Volume rendering on scalable shared-memory mimd architectures. Workshop on Volume Visualization, pages 17--24, Oct 1992.

....processor which initiated this ray. Proper partitioning of the volume and 5 proper assignment of these partitions to the processors are essential for exploiting spatial and temporal coherency. Some examples of such an approach can be found in [1] 3] 4] In the object dataflow approach [5] 8] 9][13], the volume is partitioned similar to the partitioning in the ray dataflow approach. In addition, the screen is also partitioned and assigned to the processors. During the rendering stage, the processors cast rays for the assigned screen segments. For each ray, a processor fetches the data from ....

Nieh, J., Levoy, M.,"Volume Rendering on Scalable Shared-Memory MIMD Architecture", Proceedings of 1992 Workshop on Volume Visualization, Boston, MA, pp. 17-24.

....of small or moderate scale multiprocessors makes them viable platforms for speeding up volume rendering to achieve the desired frame rates. Successful parallel implementations of ray casting renderers have been developed for both centralized and distributed shared memory multiprocessors [5]. Parallel implementations of the faster shear warp method have also been developed [9, 4] but these have not scaled well beyond 8 to 12 processor systems. As programmers trying to improve parallel performances, we first attempt to understand why, studying the performance, scalability and ....

....Since the goal of real time volume rendering of reasonably large data sets has not been achieved, it is important to exploit parallelism to obtain higher performance. Based on the serial ray casting algorithm, a parallel ray caster for coherent shared address space systems was developed in [5]. The reasons for its good parallel performance high temporal locality as well as its spatial and temporal locality properties have been carefully analyzed [3, 10] More recently, parallel versions of the shear warp algorithm have also been implemented on similar systems [9, 4] The parallel ....

Nieh J. and Levoy M. Volume rendering on scalable sharedmemory MIMD architectures. In Proceedings of the 1992 Workshop on Volume Visualization, pages 17--24, 1992.

....it often becomes infeasible to employ a uniprocessor for the task of rendering numerous animation frames. Several parallel rendering algorithms have been proposed on shared memory architectures, where the complete scene is statically distributed to, and is shared among, all the processors [4][17]. Objects are fetched using fast hardware based communication protocols, and the communication process is transparent to the user. Algorithms designed using this scheme take advantage of coherency between rays to reduce communication overheads by efficiently caching the data not available locally ....

....initiated this ray. Proper partitioning of the volume and proper assignment of these partitions to the processors are essential for exploiting spatial and temporal coherency. Some examples of such an approach can be found in [1] 3] 6] In the object dataflow approach (for example [9] 13] 14] 15][17]) the volume is partitioned similar to the partitioning in the ray dataflow approach. In addition, the screen is also partitioned and assigned to 6 the processors. During the rendering stage, the processors cast rays for the screen segments assigned to them. For each ray, a processor fetches the ....

J. Nieh, Levoy, M.,"Volume Rendering on Scalable Shared-Memory MIMD Architecture", Proceedings of 1992 Workshop on Volume Visualization, Boston, MA, pp. 17-24.

....the way memory is laid out, it reads only a single word on each page. Thus, there is significant fragmentation in communicating remote data in pages at column boundaries in SVM systems. Irregular Applications Volrend renders three dimensional volume data into an image using a ray casting method [58]. The volume data are read only. Its inherent data referencing pattern on data that are written (task queues and image data) is migratory, while its induced pattern at page granularity involves multiple producers with multiple consumers. CHAPTER 1. INTRODUCTION 9 Both the read accesses to the ....

J. Nieh and M. Levoy. Volume rendering on scalable shared-memory MIMD architectures. In Proceedings of the 1992 Workshop on Volume Visualization, pages 17--24, 1992.

....structures evolve over time. For visualizing large data sets, parallel processing is often used to speed up the expensive volumetric rendering process. Although the subject of rendering a single volumetric data set using a parallel computer has been studied extensively by numerous researchers [17, 16, 14, 22, 10], parallel animation of TVVD, in contrast, has received relatively little attention. Compared to parallel volume rendering of a single data set, rendering TVVD in parallel poses a di#erent set of design tradeo#s. First, because TVVD typically consists of a sequence of data volumes, the I O ....

J. Nieh and M. Levoy, Volume Rendering on Scalable Shared-Memory MIMD Architectures, in 1992 Workshop on Volume Visualization, 1992, pp. 17--24. Boston, October 19-20.

....SPLASH 2 version [97] to run more e#ciently on SVM systems. A global lock that was not necessary was removed, and task queues are implemented better for SVM and SMPs [47] Inherent communication is small. We present results only for the SMP protocols due to simulation cycle limitations. Volrend [65]: This application renders a three dimensional volume using a ray casting technique. The volume is represented as a cube of voxels (volume elements) and an octree data structure is used to traverse the volume quickly. The program renders several frames from changing viewpoints, and early ray ....

J. Nieh and M. Levoy. Volume rendering on scalable shared-memory MIMD architectures. In Proceedings of the Boston Workshop on Volume Visualization, Oct. 1992.

....0.00 Ocean rowwise [14, 45, 28] 514x514 ocean 248.3 18.40 21.76 (19.26) 9.21 Water nsquared [48] 4096 molecules 360.6 21.00 15.26 (46.17) 62.76 Water spatial [48] 15625 molecules 157.2 6.80 41.60 (41.80) 9.69 Radix local [11, 23, 49] 4M keys 5. 9 90.79 26.76 (27.00) 53.21 Volrend stealing [37, 48, 28] 256x256x256 cst head 13.2 45.30 43.81 (42.44) 50.44 Raytrace [44, 48] 256x256 car 29.8 39.89 2.52 (50.03) 59.01 Barnes original [3, 22, 43] 32K particles 47.7 117.65 41.07 (68.25) 1.98 Barnes spatial [28] 128K particles 219.2 20.16 40.99 (37.70) 33.84 Table 1: Application statistics. The ....

J. Nieh and M. Levoy. Volume rendering on scalable sharedmemory MIMD architectures. In Proceedings of the Boston Workshop on Volume Visualization, Oct. 1992.

....run length encoding, projection template, multiresolution volumes, interactive classification, parallel processing. 1 Introduction An effective approach to achieve high frame rates for volume rendering is to parallelize a fast rendering algorithm that relies on some algorithmic optimizations [1, 2, 3, 4]. Two requirements must be met for this approach to achieve interactive rendering. First, the serial volume rendering algorithm must be fast enough. Second, the parallel version of the serial algorithm must scale well as the number of processors increases. Many parallel volume rendering ....

....volume renderers. Among the most efficient ones is Lacroute s [3] a real time parallel volume rendering algorithm on a multiprocessor SGI Challenge using the shear warp factorization [5] which could render a 256 3 volume data set at over 10 Hz. A dynamic task stealing scheme was borrowed from [1] for load balancing. Parker et al. 4] proposed another interactive parallel ray casting algorithm on SGI workstations. Using 128 processors, their algorithm rendered a 1GByte full resolution Visible Woman data set at over 10 Hz. One of their optimizations for ray casting was using a multi level ....

[Article contains additional citation context not shown here]

J. Nieh and M. Levoy. "Volume Rendering on Scalable Shared-Memory MIMD Architectures". Proc. 1992 Workshop on Volume Visualization, 1992, 17-24.

....node separately. According to the actual view direction, a frontto back sorting of the subblocks has to be performed, along with a final global communication step to merge the resulting subpictures. A high speed implementation on a DASH shared memory architecture was presented by Nieh et al. [12]. Due to the fast memory access, coupled with first and second order level data caches and the utilization of spatial data coherences, impressive results were achieved. In terms of the overall rendering times, a comparison with other parallel implementations appears rather difficult. This is due ....

Nieh, Jason and Marc Levoy. "Volume Rendering on Scalable Shared-Memory MIMD Architectures". Workshop on Volume Visualization, Boston, October 19-20, 1992, pp. 17-24.

....image q network injection bandwidth at a node t time consumed per frame routing patterns that approximate the random distribution used to characterize network performance. A fine grain randomly interleaved block data distribution achieves this and makes the redistribution size view independent [Nieh92]. This data distribution is the context for the remainder of section 4. 4.1.1 Redistribution Costs Redistribution size is affected by replication of the data se t . Def i ne a da t a s i ze d and a rep l i ca t i on fac t or r (1 r n) where r is the number of copies of the data stored in ....

....data, and object partitions. 4.1. Image Partition with Static Data Distribution In this class of algorithms, nodes are assigned one or more subsets of image lattice points to compute. Often shafts subsets are used which equates to assigning screen regions to nodes [Chal91] Corr92] Mont92] [Nieh92] [V zi92] Yoo91] Data subsets are distributed among the nodes in a static distribution a specified data point is always stored in the same node s local memory. To render their region(s) nodes access remote or local data as necessary (Fig. 2) based on the current view transformation. ....

Jason Nieh and Marc Levoy. "Volume Rendering on Scalable Shared-Memory MIMD Architectures." 1992 Workshop on Volume Visualization, 17-24, October 1992. Workshop Proceedings.

....7. They show a marked improvement, with a much smoother curve and an equivalent speed up of 76.9 on 256 processors. It is interesting to note that the modified processor farm algorithm which we are proposing is very similar to one used by Nieh and Levoy for volume rendering on the DASH machine [Nieh92]. In fact, our approach was devised independently, and for different reasons. Nieh and Levoy used it as a way to minimise dynamic data communication; we are using it to avoid shared variable contention. In order to set these results in context, we show in Table 1 the actual rendering times in ....

Nieh J. and Levoy M., Volume Rendering on Scalable Shared-Memory MIMD Architectures, Proc. ACM Workshop on Volume Visualization, 1992, pp. 17--24.

No context found.

J. Nieh and M. Levoy. Volume rendering on scalable shared-memory MIMD architectures. In

No context found.

J. Nieh and M. Levoy. Volume rendering on scalable shared-memory MIMD architectures. In

No context found.

J. Nieh and M. Levoy. Volume rendering on scalable shared-memory MIMD architectures. In Proceedings of the Boston Workshop on Volume Visualization, October 1992.

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