Proceedings Article | 13 July 1998
KEYWORDS: Asynchronous transfer mode, Picture Archiving and Communication System, Local area networks, Medical imaging, Sun, Switches, Image acquisition, Computed tomography, Image segmentation, Receivers
We have replaced our shared-media Ethernet and FDDI network with a multi-tiered, switched network using OC-12 (622 Mbps) ATM for the network backbone, OC3 (155 Mbps) connections to high-end servers and display workstations, and switched 100/10 Mbps Ethernet for workstations and desktop computers. The purpose of this research was to help PACS designers and implementers understand key performance factors in a high- speed switched network by characterizing and evaluating its image delivery performance, specifically, the performance of socket-based TCP (Transmission Control Protocol) and DICOM 3.0 communications. A test network within the UCLA Clinical RIS/PACS was constructed using Sun UltraSPARC-II machines with ATM, Fast Ethernet, and Ethernet network interfaces. To identify performance bottlenecks, we evaluated network throughput for memory to memory, memory to disk, disk to memory, and disk to disk transfers. To evaluate the effect of file size, tests involving disks were further divided using sizes of small (514 KB), medium (8 MB), and large (16 MB) files. The observed maximum throughput for various network configurations using the TCP protocol was 117 Mbps for memory to memory and 88 MBPS for memory to disk. For disk to memory, the peak throughput was 98 Mbps using small files, 114 Mbps using medium files, and 116 Mbps using large files. The peak throughput for disk to disk became 64 Mbps using small files and 96 Mbps using medium and large files. The peak throughput using the DICOM 3.0 protocol was substantially lower in all categories. The measured throughput varied significantly among the tests when TCP socket buffer was raised above the default value. The optimal buffer size was approximately 16 KB or the TCP protocol and around 256 KB for the DICOM protocol. The application message size also displayed distinctive effects on network throughput when the TCP socket buffer size was varied. The throughput results for Fast Ethernet and Ethernet were expectedly lower but the patterns were interestingly different from those for ATM. To achieve the optimum throughput in a TCP-based high-speed switched medial imaging network, the size of the TCP socket buffer is the most important parameter to optimize. If the DICOM 3.0 protocol is used, however, the performance gain by tuning system parameters is minimal, particularly if small files are used. Compared to socket-based TCP, the decrease in throughput caused by DICOM 3.0 protocol overhead is significantly larger in a high-speed switched network. This suggests that the protocol itself is the bottleneck in high-speed networks and that the protocol should be fine-tuned to take advantage of the services provided by such networks and not to duplicate them. To design a successful high-speed PACS network, it is important that bandwidth-demanding workstations and servers be on the same subnet and use the same technology so that no routing and data conversions are required.
