In the last decade the NWS (National Weather Service) has been especially active in modernizing its field offices and upgrading its observing systems. Advanced Doppler radars have been deployed across the country, an automated surface observation system has been fielded, satellites have been upgraded and all forecast offices have received a new system to integrate and display the many existing and new observational data sets. Although these systems are relatively new they are already feeling the effects of rapidly changing technology and increasing system requirements. Some of the current technology is becoming difficult to maintain and is being stressed to meet the additional demands on the system.
The scope of this discussion is limited to the proposed evolution of the workstations system in the field offices and does not address any planned upgrades of sensors and networks. The design challenge is to arrive at an architecture and hardware complement that will meet the long-term needs of the NWS and one that can be achieved by evolving the existing system.
The Linux operating system has been used by scores of software developers for years. However, recently Linux is also gaining popularity with many other groups for web server, databases, and other applications. The Linux operating system is also being considered for the new AWIPS system architecture. Linux runs on a variety of computer platforms including the Intel-based PCs. Software developed for Linux is more likely to use ANSI compliant compilers and eliminate the developer's temptation to use language extensions provided by the hardware vendor. Using vendor extensions may lead to reliance on one particular vendor's hardware, which can make it difficult to take advantage of technological advances and cost reductions offered by other vendors. Figure 1 provides a performance comparison between a Pentium 3 PC and the current AWIPS hardware. The average performance improvement for various typical AWIPS functions was double. For some CPU-intensive tasks, such as data decoders, the performance improvement was more than ten times better. This cursory test seems to indicate that PCs are capable of meeting the performance requirements for operational forecast systems.
The AWIPS system is broken down into five major components to help describe the key features of the new system architecture. The system at each forecast office consists of the display, data storage, application, data acquisition and local area network components.
3.1 AWIPS Display - The current AWIPS display software takes advantage of the hardware vendor's proprietary graphics card and operating system. To eliminate this dependency, the AWIPS software has been rewritten to work with most 24-bit (true color) graphics cards. The graphics displays are generated using the X11 library and the user interface uses the tcl scripting language and tk toolkit.The proposed display consists of dual processor PCs with dual color monitors running Linux and the AWIPS-Linux software. A single mouse and keyboard will be used to enter data and control the display functions. Because of the need to buffer large amount of images and some raw data, two gigabytes of memory are suggested. Each forecast office will have five or more of these display systems. A third monitor could be used for text display, although a separate PC and monitor are being considered for displaying text.
3.2 Data Storage - With the exception of text data and some hydrological data sets (which are stored in a relational database) hydro-meteorological data is currently stored on redundant data servers in flat file format. The use of a single data repository simplifies data management but may lead to a communications bottleneck as the data demand by workstations and applications increases.
The proposed data storage architecture consists of a Linux data server with dual processors, power supplies, and network cards; and RAID 5 data striping. In addition to storing data on the data server, which stores all meteorological data for a specified period of time, selected data will also be stored on each workstation and application processor. This will provide rapid access to a selected set of data since disk contention is reduced and data does not have to be transmitted over the network.
3.3 Application - AWIPS applications include a wide range of data processing from local forecast models, to storm tracking algorithms and decoders. Currently, AWIPS does not include local forecast models, and other applications are severely restricted by the limited amount of available processing.
Beowulf clusters consist of a number of Linux computers interconnected by a high speed network, such as Myrinet. These clusters compete with more expensive high performance computers and are ideal for executing certain forecast models. These clusters are being evaluated at the Forecast Systems Laboratory for use with mesoscale forecast models.
To meet the high reliability requirements for an operational system several software packages exist to create High Availability Linux configurations. Two or more Linux computers can be configured to allow load leveling between machines or automatic fail-over should one of the computers fail. A dedicated heartbeat LAN between the computers detects when one of the computers fails. High availability Linux configurations are proposed for the applications processors in the new AWIPS architecture.
3.4 Network and Communications Protocol - The current AWIPS system uses FDDI to interconnect all workstations and servers, and a 10 Mb/s ethernet to attach data acquisition processors, printers, and other peripheral devices. Redundant network switches (using the spanning tree) route data between the FDDI and ethernet networks. FDDI technology is nearing obsolescence and the 10 Mb/s ethernet is actively being replaced by faster ethernets. The proposed LAN uses a network switch that supports 100 Mb/s and 1 Gb/s ethernet speeds. These communication speeds are expected to be adequate for the higher resolution national models and base radar data that are expected to be part of AWIPS.
A higher-level communications protocol that supports data broadcast, using a single transmission, will transmit real-time data to the local data caches on the various processors. The broadcast protocol allows for retransmission of data if one of the nodes did not receive the complete transmission.
3.5 Data Acquisition - The AWIPS national network consists of a Satellite Broadcast Network that broadcasts synoptic observations, satellite imagery, guidance products, and other information. A WAN (Wide area network) is used to forward information to the network control facility and exchange information between sites. Local data, such as radar and mesonet observations, are received with special communications processors. These processors will also be replaced with Linux-based machines.
The following paragraphs describe the evolutionary change in the system architecture using these components.
4.1 Old Architecture - Figure 2 shows the system components as they are configured for the current AWIPS system architecture. Hewlett Packard (HP) is the manufacturer of the majority of the computer hardware. The HP data and applications servers are configured such that a failure of any one of these machines automatically results in another machine assuming the load of the failed machine. In this architecture, the FDDI ring interconnects the data and applications servers, and several HP workstations. The ring interfaces to a lower speed 10 Mb/s ethernet that interconnects the "front end" data acquisition and communications processors.
Real-time data flows from the acquisition and communications processor, through the network switch to the data servers, where the data is then decoded and stored. The workstations and other applications access the data on the servers as needed.
4.2 Transition Architecture - The transition architecture (Figure 3) includes a fast ethernet port on the existing network switch. This ethernet port is connected to another high-speed network switch that will be part of the advanced system architecture. All new hardware will be connected to this switch and will be able to communicate with the older machines on the other part of the network. As new machines are added to the network, older machines are removed and the older networks are slowly decommissioned.
With this enhanced architecture, the data flows from the Linux communications processors on the new high-speed network, to another Linux machine where the data is decoded and stored on its disks. Initially, only some of the data will be decoded on the Linux machine and the rest will be decoded on the HP server. A broadcast protocol will send some of the decoded data from the server to all of the workstations on the new high-speed network
4.3 Proposed Architecture - Figure 4 depicts the proposed final AWIPS system architecture. All of the original computers will be replaced with Linux computers connected to the new high-speed network. The old hardware, as well as the FDDI and low speed ethernet will be decommissioned. Additional application and acquisition processors will be added to accommodate specific new requirements.
It is envisioned that data will flow from the various ingest and communications processors on the high speed network to redundant Linux machines for decoding. The decoded data will then be stored on a network file server and also broadcast to all machines on the local network. The workstations and application processors will have most of the data stored on their local disks.
The large majority of the display and data acquisition code has been ported to a Linux platform. The data decoders have been ported and are being tested for reliability and data accuracy after decoding. Also, the AWIPS contractor has converted the SBN communications processor (CP) code to run on a Linux platform and has provided initial systems to the NWS and FSL for evaluation.
Porting the hydrological applications (written mostly in Fortran) is a significant effort that is currently in progress and expected to be completed before the end of this year. In order to support the existing AWIPS configuration in the field and the transitional hardware, all code is being compiled each day for the HP-UX and the Linux platforms.
Several AWIPS/Linux workstations have been deployed to NWS field offices for evaluation. These systems are currently connected to the 10 Mb/s ethernet and therefore see significant performance improvement only for processes that do not require transfer of large amounts of data. However, tests performed at FSL using the transitional configuration with a 100 Mb/s ethernet show significant overall improvements in performance. The CP, 100 Mb/s ethernet switch, and decoder processor for the AWIPS transitional architecture will be installed at a larger number of field offices in the near future.
Although Linux has been widely used for web servers, data servers, and software development, it has not been used by government or industry for mission critical applications. There are a number of issues that concern potential Linux customers. Among these are maintenance and support, system configuration management, and security. Although it will take time to satisfactorily answer these questions, the risk appears manageable. A number of companies, including HP, IBM, and Red Hat, provide Linux support. Red Hat and other Linux distributors package Linux with other software, such as X-window software and drivers for a variety of equipment. Linus Torvalds, the founder of Linux, controls the evolution of the Linux kernel. Rapidly changing technology and hardware can be a challenge for managing a large number of systems. This will require additional work from the software developers and integrators to insure that all software continues to work with new releases of hardware and software. Making large purchases, rather than small interactive purchases can help to reduce the variety of hardware that has to be maintained.