Application-Level Virtual Memory Management in Real-Time Multiprocessor Systems

In such cases, for an image generator, with a given memory space and loading speed form hard disk, the virtual scene must be partitioned such that the memory complexity of all segments which must be present in memory do not exceed the available memory space. This situation requires decreasing the virtual scene granularity: if the spatial dimension of virtual segments is smaller, the memory complexity of their models is smaller, and they can be better fit in the physical memory. In Figure 5, two spatial partitions of the virtual scene are presented. We can assume that the memory complexity of a virtual segment and time required for load it, are proportional with the spatial dimension of the segment.

The dimension of all segments that are at a distance less then the far distance of the viewing frustum is smaller for a smaller (finer) granularity of the virtual scene. In Figure 5, the ratio of spatial dimensions of segments is 24/48 = 0.5.

The lower bound of the virtual scene granularity is limited by the execution time of the mapping function: when the granularity decreases, the number of virtual segments increases and the time required for test and update their status increases.

5. EXPERIMENTAL RESULTS AND CONCLUSIONS
The virtual segments management in large synthetic environments was studied and experimented on a Silicon Graphics multiprocessor, Onyx, with four MIPS R10000 processors and Infinite Reality graphic accelerator, under IRIX 6.4 operating system.

The real-time image generation program is based on OpenGL graphical interface and Performer loaders. The system memory of 512 Mbytes allows the image generation program to be locked in memory, if the program controls the loading of the virtual scene. The text segment of the program has a dimension about 280 Mbytes, so that only a limited memory space (less than 200 Mbytes) is available for scene graph representation, so that the virtual segments of the scene graph management is strongly required.

The synthetic environment consists in the geographical area of our country, over 6 latitude x 10 longitude degrees. This large scene is divided into 240 virtual segments, each over 0.5 x 0.5 degrees. Each segment is a multiresolution model of a terrain region, with features and cultural objects placed inside. The medium complexity of a segment is about 10000 polygons for terrain, features and 3-D objects, and requires a memory space of about 2 Mbyte. The whole database requires 240 x 2 M bytes = 480 Mbytes, which are not available. A virtual segment management function built-in the image generation program, allow to keep in available memory (of about 180 Mbytes) only the segments which are needed for the given observer viewpoint, and replace them when needed.

The virtual segments of the scene graph management replaces the virtual memory mechanism of the operating system. All pages of the processes are locked in the memory, so that, for the image generation program no page fault and swapping can occurs.

The management of memory space is, instead, controlled by the application program. Virtual segments of the synthetic environment are loaded or discarded, depending on the observer viewpoint, and not on page faults, such as in operating system virtual memory management. This large synthetic environment is used for airplane pilots training, using detailed modeling of the terrain, features and cultural area.

6. REFERENCES
[1] Hwang, K., Advanced Computer Architecture: Parallelism, Scalability, Programmability, McGrow-Hill, 1993.
[2] Ionescu, F.,Codesign of a Parallel Architecture for Visual Systems, in Proceedings of CSCS’97 (Bucharest, Romania, May 1997), 242-245.
[3] Ionescu, F., Gheorghiu, C., Englert, C., Popa,C., Suciu, I., Optimization of the execution time in a distributed system implementing a full flight simulator, in Proceedings of ITEC’98, (Lausanne, Switzerland, April 1998), 509-514.
[4] Ionescu, F., Ionescu, M., "Pipeline processing of shared graphs in multiprocessor systems", in Proceedings of PDPTA’99, (Las Vegas, Nevada, June 1999), 1958-1963.
[5] Silicon Graphics, Inc., Topics in IRIX Programming, Mountain View, Calif., 1996.