Remote rendering and graphics for network systems Applications

Author: Manyu Tang

Prepared for Prof. Javed I. Khan
Department of Computer Science, Kent State University
Date: November 2003

Abstract: Knowledge developed from the network has made two big contributions to computer graphics:  Providing a huge computation power by using a cluster of computers as a computing resource and Enable remote rendering by using network as a communication infrastructure.  In this survey we first introduce the basics about computer rendering and followed by usage of network in the computer rendering. Finally, we conclude survey by presenting existing problem of combining network and rendering and current solutions to them. 

Other Survey's on Internetwork-based Applications
Back to Javed I. Khan's Home Page

Table of Contents:



What is Computer rendering

Computer rendering can arguably be defined as the process of creating pictures using computer by extracting useful information from a dataset.  Following the fast development of computer technology, the applications of computer rendering has moved from dealing with 2D images to  3D images.  The main applications categories include:  Computer Aided Design, Scientific Visualization, and Computer Animation and Computer Games.

The common question to the computer rendering is the difference between displaying a digitized picture and rendering a dataset.  The digitized picture, for instance, from a digital camera,  is actually a file which contains a dataset for a 2D image whose size is about 1 or 2 MB.  The operations we can perform on are simple such as rotating, enlarge and reducing image.  Computer rendering usually copes with a large dataset whose size of 100 MB is not unusual.  The operations on the dataset is very complicated.  They usually require advanced mathematic theory, fast processors and a long time.   

Additionally, the job of computer rendering is not only to display a  image of the a dataset but also manipulate the dataset by applying different algorithms to solve a various problem.  For example, in a building 3D dataset we can see the outside of building.  We can also let computer take off the roof and see the inside of the building such as the arrangement of the rooms.  We can even transparent the wall of 

the building to see how many cables buried inside.  Overall, computer rendering is a very advanced topic and it has been widely used as a tool in many different scientific research areas.

The general rendering methods

Computer rendering is a very complicated task which involves nearly every aspects of computer software and hardware back up by the high level mathematic theory.  Generally, there are following methods to accomplish the rendering process.  But, most of the time, those methods need to be integrated together in order to achieve the best rendering result.

The rendering pipeline

The goal of computer rendering is to produce a useful scene from a dataset.  Such a task is often inherently interactive and iterative: one can not expect to accomplish it by simply pushing a large dataset into a black box.  Suppose that we have a set of vertices that defines a set of primitives.  Because the usual representation is given in terms of locations in space, we can refer to the set of primitive types and vertices as the geometry of the data.  In a complex scene, there may be thousands or even millions of vertices that define the objects.  We must process all these vertices in a similar manner to form and image in the frame buffer.  If we  think in terms of processing the geometry of our objects to obtain an image, we employ the following block diagram which shows the four major steps in the rendering process.


Many of the steps in the image process can be viewed as transformations between representations of objects in different coordinate system.  Some vertex data (for example, spatial coordinates) are transformed by 4 x 4 floating-point matrices. Spatial coordinates are projected from a position in the 3D world to a position on a screen.  If advanced features are enabled, this stage is even busier. If texturing is used, texture coordinates may be generated and transformed here. If lighting is enabled, the lighting calculations are performed using the transformed vertex, surface normal, light source position, material properties, and other lighting information to produce a color value.


Clipping, a major part of primitive assembly, is the elimination of portions of geometry which fall outside a half-space, defined by a plane. Point clipping simply passes or rejects vertices; line or polygon clipping can add additional vertices depending upon how the line or polygon is clipped.  The results of this stage are complete geometric primitives, which are the transformed and clipped vertices with related color, depth, and sometimes texture-coordinate values and guidelines for the rasterization step.


In general, three-dimension objects are kept in three dimensions as long as possible, as they pas through the pipeline.  Eventually, after multiple stages of transformation and clipping, the geometry of the remaining primitives (those that have not been clipped out and will appear in the image) must be projected into two-dimensional space.


Rasterization is the conversion of both geometric and pixel data into fragments. Each fragment square corresponds to a pixel in the framebuffer. Line and polygon stipples, line width, point size, shading model, and coverage calculations to support antialiasing are taken into consideration as vertices are connected into lines or the interior pixels are calculated for a filled polygon. Color and depth values are assigned for each fragment square.


Networks and rendering

Network based parallel computing backend has been applied to the computer rendering since long time ago.  One difficulty of computer rendering is that it requires a tremendous computing resource. Usually, a standalone computer is not powerful to absorb all the computation required within a reasonable time.  With the advanced network technology and reducing cost of personal computer, using a group of computers as a computing resource becomes a more and more attractive resolution.  In addition, the multiple computers have also been used as the final rendering device for the large scene.

Only lately, the remote rendering became an interesting topic.  As we mentioned early, computer rendering requires high computing power to result a huge final dataset for the rendering.  Plus, the rendering  application is usually interactive oriented and every interaction requires a new computation and result a new dataset.  As remote rendering stands that the dataset and rendering device do not exist at the same place.  If we rendering an object via the Internet, physical distance between the rendering device and computed dataset could be ten thousands away from each other. Thus, remote rendering is groundless without high bandwidth network's support.  Only recently, the development of the network technology has fulfilled the such requirement.  However, huge dataset and frequent data transmission are still the bottle neck for the remote network performance.

In this survey we will discuss the above two aspects about computer rendering on the network systems, mainly focusing on the research activities in these two fields and their progress.  The next section will be given quite details on the current development of the network based computer rendering.  Then, we will talk about the problems in the network based rendering and how to overcome.  Finally, a set of useful web reference will be given for the interesting reader for the further research.


Network Aided Rendering

Cluster based rendering

Computing farm

As we mentioned it early, the computer rendering is an intensive computing process which often exceeds the single computer's computation power.  Using a cluster or a group of computers as a single computing source can be attractive approach to solve this problem.  Over here, we define a cluster as a collection of interconnected computers used as a single, unified computing resource via some form of network.  Though clusters which are generally expected to be connected by a local area network, they could also be connected by a wide area network such as Internet.  Clusters are usually inexpensive to build since they can be assembled using conventional personal computer technology and freely available software such as a Beowulf cluster. 

With computation is actually performed on a cluster, the original rendering dataset need to be decomposed to each single computation node in the cluster. The decomposition task can be realized by either algorithmic decomposition (different parts of the program are run on the same data) or by domain decomposition (the same program/algorithm is run on different parts of the data).  Under the cluster computing environment, the processors usually divided into two groups: 

  1. Geometry processor: it determines which geometric objects appear on the display and assigns shades or colors to these objects.
  2. Rasterization processor: it takes the output from Geometry processor and determines which pixels should be used to approximate a line segment between the projected vertices.

According to when to decide which object gets projected eventually or when to sort those geometric primitives following their depth, the algorithms of how to distribute the original dataset are usually divided as followings,

  1. sort middle:   sort procedure is done between geometry processors and rasterization processors.  This configuration was popular with high-end graphics workstations a few years ago, when special hardware was available for each task and there were fast internal buses to convey information through the sorting step.
  2. sort last:   sort procedure is done after rasterization processors.  This configuration is popular with commodity computers connected by a standard network.
  3. sort first:   sort procedure is done before the geometry processor. This configuration is ideally suited for generating high-resolution display.  Suppose that we want to display our output at a resolution much greater than we can get with typical CRT or LCD displays that have a resolution in the range of 1-2 million pixels.  Such displays are needed when we wish to examine high-resolution data that might contain more than 100 million geometric primitives.


Multi-display rendering

Over the last few years, computer have become faster and more powerful by almost all measurement.  Computer processor speeds have increased, memory has become cheaper, and hard drive sizes have grown.  However, there is one computer component that has remained largely unchanged.  Displays, such as standard computer monitors, have remained at a resolution of around one million pixels.  In order to build a higher resolution display, current approach is to combine a group of computers' monitor to make a big display.  And, each monitor in the group is backed by a computer for rendering only a part of the big display.  Most of these display fall into one of following three broad classes,

  1. abutted displays: They require that all displays be carefully aligned so that no pixels overlap.  Abutted systems are fairly common and are used in everything from sports stadium scoreboard to trade show exhibits. Some examples of abutted displays are CAVE [2], Office of Real Soon Now" [1], and the display wall system at Lawrence Livermore National Laboratory [12].
  2. regular overlap displays: they require displays to be carefully aligned so that there is some controlled overlap between displays. the displays are required to have precise geometric relationships that ensure regularity between overlap regions.  The overlap regions are used to blend imagery across display boundaries.  This is done to help hide both photometric and geometric discontinuities at the boundaries.  Princeton's Scalable Display Wall [6] and Standford's Interactive Mural [5] are both examples of regular overlap displays.
  3. rough overlap displays:  it is most complex display system because it allows rough overlap regions between displays.  The only requirement is that displays actually overlap.  This means that overlap regions can be of arbitrary shape and size.  At the cost of its complex design, rough overlap display gives most flexibilities. UNC's PixelFlex system is an example of this type of this type of display [3].

Client and Server based rendering

As we have discussed so far, the rendering device and rendering dataset are at the same location.  Nevertheless, we might have a situation where the rendering dataset is at one place and we need to view the scene from the other site, especially under the situation of networked multi-user virtual environments which require that users share a common scene over a network.  Example include networked walkthroughs of large information spaces (buildings, databases, ultimately a 3-D Internet) and interactive applications such as immersive cinema, networked games and computer supported cooperative work.

A virtual environment requires efficient rendering of the three-dimensional objects forming the simulated world.  In a distributed virtual environment, the work is divided between processes.  One  process will maintain the actor database and run the simulation, whereas another is responsible for rendering.  These processes will often run on separate CPUs or workstations, which creates the need for rendering.  In multi-user systems, this is always required, independent of network topology.

Given today's typical hardware setup with high speed CPUs, fast system buses and comparatively slow network transmission, it is very reasonable to assume that the network is the most constrained resource of the whole system.  Currently, three distinct models for distributed graphics have been developed for the visualization of distributed geometry databases, with overall goal of minimizing bandwidth consumption on the network.

  1. Image-based:  Rendering is performed by the sender, and the resulting stream of pixels is sent over the net (e.g. digital TV, X pixmaps).
  2. Immediate-mode drawing:  The low-level drawing commands used by drawing APIs are issued by the application performing the rendering, not immediately executed, but sent over the network as a kind of remote procedure call.  The actual rendering is then performed by the remote CPU (e.g. distributed GL[11], PEX in immediate mode [12]).
  3. Geometry replication:  A copy of  the geometric database is stored locally for access by the rendering process.  The database can either be available before application start (kept on local hard disk, such as seen in computer games like DOOM [13] and networked simulations such as NPSNET [14], or downloaded just before usage, such as current VRML browsers do [15]. 

[other parts of your paper]

Complications with the network:

Load balance on cluster

As the cluster does the computation, the node in the cluster needs to communicate with each other to update required dataset which creates a busy network traffic and consequently spends a lot of unnecessary CPU time.  In addition, some nodes might have light computation load while others have heavy load to do.  Then, nodes will have to wait on the nodes with heavy load  to finish for moving down to next computing stage.  Thus, how to reasonably distribute the workload on each node is essential process to get best cluster performance.  Currently, there are following algorithms developed to assign a proper workload to each processor,

  1. Data driven - Tasks are subdivided for each processing element in a preprocessing step.  Each processing element then works on it assigned dataset.  There is no communication between processing elements.
  2. Demand driven - Tasks are finely subdivided.  When a processing element is finished processing one task, it is assigned another.  This continues until all tasks have been completed.
  3. Data parallel - The data itself is subdivided with each processing element performing a task on a piece of the data.  Processing elements will have to communicate with each other in order to access data not available to a particular processing element.



Data transmission

Variations of data replication are commonly used for networked rendering applications.  However, low network throughput and large database sizes are  the main problems which constrain the usability of every possible method.  As the data has to be shipped to the user at some point, this problem is always present.  Making the user wait for more than a couple minutes destroys immersion and makes many interactive applications completely useless.  Furthermore, extended waiting periods mean that a download process cannot be invoked frequently, so exploratory behavior of 3-D data spaces becomes impossible.  Currently, there are following approaches to attack this problem,

Software Approach


Hardware Approach



Research Papers for More Information on This Topic

  1. Ian foster, Joseph Insley, Gregor von Laszewski, Carl Kesslman and Marcus Thiebaux, "Distance Visualization: Data Exploration on the Grid", "Computer, 1999", pp36
  2. Kengel, P. Hastreiter, B. Tomandl, K. Eberhardt and T. Ertl, "Combining Local and Remote Visualization Techniques for Interactive Volume Rendering in Medical Application", "Proceedings of the conference on Visualization, 00"
  3. Jian Yang, Jiaoying Shi, Zhefan Jin and Hui Zhang, "Design and Implementation of A Large-scale Hybrid Distribute Graphics System", "Proceedings of the Fourth Eurographis Workshop on Parallel Graphics and Visualization , 2002"
  4. Mark Coates, Rui Castro and Robert Nowak, "Maximum Likelihood Network Topology Identification from Edge-based Unicast Measurements", "ACM SIGMTRICS Performance Evaluation Review, Proceeding of the 2002 ACM SIGMRYTICS international conference on Measurement and modeling of computer systems"
  5. Wim Lamotte, Eddy Flerackers, Frank VanReeth, Rae Earnshaw and Joao Mena De Matos, "Visinet: Collaborative 3D Visualization and VR over ATM Networks", "IEEE Computer Graphics and Applications", pp 66-75
  6. Greg Humphreys, Ian Buck, Matthew Eldridge and Pat Hanrahan, "Distributed Rendering for Scalable Displays", "Proceedings of the 2000 ACM/IEEE conference on Supercomputing"
  7. Kwan-Liu Ma David M. Camp, "High Performance Visualization of Time-Varying Volume Data over a Wide-Area Network", "Proceedings of the IEEE/ACM SC2000 Conference, 2000, pp.29
  8. Wes Bethel, Brian Tiernery, Jason Lee, Dan Gunter and Stephen Lau, "Using High-Speed WANs and Network Data Caches to Enable Remote and Distributed Visualization", "Proceedings of the IEEE/ACM SC2000 Conference, 2000", pp. 28
  9. Gerd Hesina and Dieter Schmalstieg, "A Network Architechture for Remote Rendering", "Second International Workshop on Distributed Interactive Simulation and Real-Time Application 1998, pp88"
  10. K.Engel, P. Hastreiter, B. Tomandl, K. Eberhardt and T. Ertl, "Combining Local and Remote Visualization Techinques for Interactive Volume Rendering in Medical Applications", "Proceeding of the conference on Visualization '00"
  11. Neider J., Davis T., WOO M., OpenGL Programming Guide- The Officicial Guide to Learning OpenGL, Adison-Wesley Publishing Company (1993)
  12. Rost, J.Freiedberg, P. Nishimoto, PEX: A Network-Transparent 3D Graphics System, Computer Graphics & Applications Vol. 9, No. 4, pp. 14-26 (1989)
  13. De Floriani L., P. Marzano, E. Puppo, Multiresolution Models for Topographic Surface Description, The Visual Computer, Vol.12, No.7, Springer International, pp. 317-345 (August 1996)
  14. Macedonia M., M. Zyda, D. Partt, P. Barham, S.Zeswitz, NPSNET: A Network Software Architecture for Large-Scale Virtual Environment, Presence, Vol.3, No.4 pp. 265-287 (1994)
  15. Hardenberg J., G.Bell, M. Pesce, VRML: Using 3D to surf the Web, SIGGRAPH'95 Course Notes, No.12 (1995)

Research Groups

Other Relavant Links

  1. Parallel Storage System:
  2. Abilene:
  3. Supernet:
  4. ESnet:
  5. National Transparent Optical Network:
  6. The Message Passing Interface (MPI) Standard:



This survey is based on electronic search in ACM's digital library, IEEE Intracom proceeding, and their citations