PERFORMANCE AND PARALLEL PROCESSING TEAM

PP1

RESEARCH TASK: SWEB: A scalable WWW server on multi-computers for the ADL

PLANNED ACTIVITY

The team is investigating issues involved in developing a scalable World Wide Web (WWW) server on a cluster of workstations and parallel machines, using the Hypertext Transport Protocol (HTTP). The main objective is to augment the performance of the ADL server by utilizing the power of multicomputers to match large demands in simultaneous access requests from the Internet. The team has implemented a prototype of a multi-processor WWW server, called SWEB, on a distributed memory machine, the Meiko CS-2, and networked SUN and DEC workstations. The method can be adapted to dynamic load changes of system resources. Experimental results indicate that the system delivers scalable performance on multi-computers with a small overhead for resource monitoring and scheduling.

The current research and development issues for the Team include extending the functionalities of the high-performance system for incorporation into ADL and generalizing the scheduling and load balancing scheme for different architectures. More specifically, the Team plans to:

1. modify the system code for the WWW Naviserver currently adopted by the ADL WWW prototype;
2. incorporate the new NCSA single-workstation httpd code to improve system performance and upgrade the ADL multi-processor web server to prepare for the anticipated increase in load;
3. characterize ADL-specific operations to assist the SWEB scheduler in making load balancing decisions;
4. study and tune the performance of the ADL server in a real user environment;
5. extend the SWEB system on a shared memory machine using lightly weighted multi-threads to improve system performance;
6. develop the global scheduling and load balancing scheme for a Java-based, client-server computing style in the ADL system.

ACTUAL ACTIVITY

All of the subtasks (a)-(f) are either completed or close to completion. For (a) we have almost completed the porting of the SWEB system to the Naviserver (AOL server). The current version of the system works, but with a number of bugs that we are currently correcting. For (b) we have used the new NCSA single-workstation httpd code with pre-forking capability and have improved the multiprocessor system performance. For (c) (d), we also have studied wavelet-based image browsing operations and examined the system performance for a set of image browsing operations. We plan to continue to study other type of ADL operations. For (e), we have received support from the U.S. Navy (NRaD, San Diego) to port the system on their shared memory machine. We have built load monitoring and balancing schemes in their multi-partition Convex machine. For (f), we have conducted research on global scheduling for adaptive client-server computing. For example, we have divided computation adaptively between client and server to reduce response times. We have again looked at the application of adaptive client-server computing in wavelet-based image browsing.

PP2

RESEARCH TASK: A compact image storage scheme for wavelet-based multi-resolution subregion retrieval.

PLANNED ACTIVITY

The images in the ADL collections are usually too large to be displayed on a single screen. Fast retrieval of subregions from compressed large image data is therefore necessary for a practical image browsing system. The Team is working to support the following two functions for browsing wavelet-based image data in ADL: retrieval of a subregion and zooming of a subregion to any desired resolution. Both fast access and compact storage of image data are important, since the number and sizes of images in the ADL collection are both large. Hybrid coding technique based on quad-tree and Huffman coding methods that the Team has developed support decompression/retrieval of image subregions in multi-resolution image data. Our method achieves effective image data compression to save disk space but also minimizes the time spent for retrieving subregions. We have conducted preliminary experiments with sample satellite images from the ADL collections and our results show that 70-90% space reduction ratio is achieved for quantized image coefficient data and the time for accessing subregion is less than few seconds on SPARC 5 with a SCSI-2 disk.

The current work of the Team involves extending the results of its approach for application in the ADL system and as well as investigating the performance of its methods for a variety of ADL images. In particular, the Team plans to:

1. modify the subregion code for the progressive transmission of image data used in the ADL WWW prototype;
2. implement an out-of-core image decomposition system to transform large-scale ADL images (e.g. 40MB or more) into a hierarchical wavelet data format since the current implementation deals only with small images due to the memory constraints;
3. extend the subregion/zoom code for the progressive transmission of image data used in the ADL WWW prototype and provide a mechanism for handling color images;
4. study the performance of the code for varieties of ADL image data;
5. investigate the data placement scheme on disk for browsing large images, with the goal of minimizing the data access times, with the possible application of parallel I/O methods.

ACTUAL ACTIVITY

All of the subtasks (a)-(e) are either completed or close to completion. For (a), the client-server version of subimage accessing is complete and we have also built a Java interface, which is now being used by the ADL testbed. For (b), we have completed an out-of-core image decomposition for handling large images. The code is yet to be fully tested because the ADL project has only just acquired a DEC server with 2GB memory that is sufficient for this task. For (c), we have started to examine subregion retrieval for color images but still have a number of difficulties to overcome. One issue is storing multi-band data in a compact manner. A research assistant is currently investigating this problem and Java implementation issues. For (d), we have examined the compression of ADL images and have found that the compression ratio is good. We will continue to conduct further experiments. For (e), we have just begun to examine the data placement and parallel I/O issue. We feel it is important for accessing large images and this will be a major task next year.