Vincent Messerli

Parallelizing I/O-intensive image access and processing applications

This article presents methods and tools for building parallel applications based on commodity components: PCs, SCSI disks, Fast Ethernet, Windows NT. Chief among these tools is CAP, our computer-aided parallelization tool. CAP generates highly pipelined applications that run communication and I/O operations in parallel with processing operations. One of CAPs successes is the Visible Human Slice Server, a 3D tomographic image server that allows clients to choose and view any cross section of the human body.

Synthesizing parallel imaging applications using the CAP (computer-aided parallelization) tool

Imaging applications such as filtering, image transforms and compression/decompression require vast amounts of computing power when applied to large data sets. These applications would potentially benefit from the use of parallel processing. However, dedicated parallel computers are expensive and their processing power per node lags behind that of the most recent commodity components. Furthermore, developing parallel applications remains a difficult task: writing and debugging the application is difficult (deadlocks), programs may not be portable from one parallel architecture to the other, and performance often comes short of expectations. In order to facilitate the development of parallel applications, we propose the CAP computer-aided parallelization tool which enables application programmers to specify at a high-level of abstraction the flow of data between pipelined-parallel operations. In addition, the CAP tool supports the programmer in developing parallel imaging and storage operations. CAP enables combining efficiently parallel storage access routines and image processing sequential operations. This paper shows how processing and I/O intensive imaging applications must be implemented to take advantage of parallelism and pipelining between data access and processing. This papers contribution is (1) to show how such implementations can be compactly specified in CAP, and (2) to demonstrate that CAP specified applications achieve the performance of custom parallel code. The paper analyzes theoretically the performance of CAP specified applications and demonstrates the accuracy of the theoretical analysis through experimental measurements.

Performances of the PS/sup 2/ parallel storage and processing system for tomographic image visualization

We propose a new approach for developing parallel I/O- and compute-intensive applications. At a high level of abstraction, a macro data flow description describes how processing and disk access operations are combined. This high-level description (CAP) is precompiled into compilable and executable C++ source language. Parallel file system components specified by CAP are offered as reusable CAP operations. Low-level parallel file system components can, thanks to the CAP formalism, be combined with processing operations in order to yield efficient pipelined parallel I/O and compute intensive programs. The underlying parallel system is based on commodity components (PentiumPro processors, Fast Ethernet) and runs on top of WindowsNT. The CAP-based parallel program development approach is applied to the development of an I/O and processing intensive tomographic 3D image visualization application. Configurations range from a single PentiumPro I-disk system to a four PentiumPro 27-disk system. We show that performances scale well when increasing the number of processors and disks. With the largest configuration, the system is able to extract in parallel and project into the display space between three and four 512/spl times/512 images per second. The images may have any orientation and are extracted from a 100 MByte 3D tomographic image striped over the available set of disks.

Performance of CAP-Specified Linear Algebra Algorithms

The traditional approach to the parallelization of linear algebra algorithms such as matrix multiplication and LU factorization calls for static allocation of matrix blocks to processing elements (PEs). Such algorithms suffer from two drawbacks : they are very sensitive to load imbalances between PEs and they make it difficult to take advantage of pipelining opportunities. This paper describes dynamic versions of linear algebra algorithms, where subtasks (matrix block multiplication, matrix block LU factorization) are dynamically allocated to PEs. It analyses theoretically the performance of the dynamic algorithms. This papers contribution is to show that the dynamic-pipelined linear-algebra algorithms can be specified compactly in CAP and yet achieve good performance. CAP is a C++ language extension for the specification of parallel applications based on macro-dataflow graphs. The CAP model, based on macro-dataflow graphs, is general and supports pipelining.

Visible Human Slice Web Server: a first assessment

The Visible Human Slice Server started offering its slicing services at the end of June 1998. From that date until the end of May, more than 280,000 slices were extracted from the Visible Man, by layman interested in anatomy, by students and by specialists. The Slice Server is based one Bi-Pentium PC and 16 disks. It is a scaled down version of a powerful parallel server comprising 5 Bi-Pentium Pro PCs and 60 disks. The parallel server program was created thanks to a computer-aided parallelization framework, which takes over the task of creating a multi-threaded pipelined parallel program from a high-level parallel program description. On the full blown architecture, the parallel program enables the extraction and resampling of up to 5 color slices per second. Extracting 5 slice/s requires to access the disks and extract subvolumes of the Visible Human at an aggregate throughput of 105 MB/s. The publicly accessible server enables to extract slices having any orientation. The slice position and orientation can either be specified for each slice separately or as a position and orientation offered by a Java applet and possible future improvements. In the very near future, the Web Slice Server will offer additional services, such as the possibility to extract ruled surfaces and to extract animations incorporating slices perpendicular to a user defined trajectory.

Computer-aided parallelization of continuous media applications: the 4D beating heart slice server

Parallel servers for I/O and compute intensive continuous media applications are difficult to develop. A server application comprises many threads located in different address spaces as well as files striped over multiple disks located on different computers. The present contribution describes the construction of a continuous media server, the 4D beating heart slice server, based on a computer-aided parallelization tool (CAP) and on a library of parallel file system components enabling the combination of pipelined parallel disk access and processing operations. Thanks to CAP, the presented architecture is concisely described as a set of threads, operations located within the threads and flow of data and parameters (tokens) between operations. Continuous media applications are supported by allowing tokens to be suspended during a period of time specified by a user-defined function. Our target application, the 4D beating heart server supports the extraction of freely oriented slices from a 4D beating heart volume (one 3D volume per time sample). This server application requires both a high I/O throughput for accessing from disks the set of 4D sub-volumes (extents) intersecting the desired slices and a large amount of processing power to extract these slices and to resample them into the display grid. With a server configuration of 3 PCs and 24 disks, up to 7.3 slices can be delivered per second, i.e. 43 MB/s are continuously read from disks and 4.1 MB/s of slice parts are extracted, transfered to the client, merged, buffered and displayed. This performance is close to the maximal performance deliverable by the underlying hardware. The observed single stream server delay jitter varies between 0.6s (52% of maximal display rate) and 1.4s (92% of the maximal display rate). For the same resource utilization, the jitter is proportional to the number of streams that are accessed synchronously. The presented 4D beating heart application suggests that powerful continuous media server applications can be built on top of a set of simple PCs connected to SCSI disks.

Performances of multiprocessor multidisk architectures for continuous media storage

Multimedia interfaces increase the need for large image databases, capable of storing and reading streams of data with strict synchronicity and isochronicity requirements. In order to fulfill these requirements, we consider a parallel image server architecture which relies on arrays of intelligent disk nodes, each disk node being composed of one processor and one or more disks. This contribution analyzes through bottleneck performance evaluation and simulation the behavior of two multi-processor multi-disk architectures: a point-to-point architecture and a shared-bus architecture similar to current multiprocessor workstation architectures. We compare the two architectures on the basis of two multimedia algorithms: the compute-bound frame resizing by resampling and the data-bound disk-to-client stream transfer. The results suggest that the shared bus is a potential bottleneck despite its very high hardware throughput (400Mbytes/s) and that an architecture with addressable local memories located closely to their respective processors could partially remove this bottleneck. The point- to-point architecture is scalable and able to sustain high throughputs for simultaneous compute- bound and data-bound operations.