MultiComputer
Solutions
A Quarterly Publication
from the CSP Inc. MultiComputer Division
August
2002
Contents (click title to jump to story)
Real Time Considerations for Sonar/Radar Applications Running Embedded Linux
2000 SERIES Fault Tolerant Library
CSP
Inc. Reports Third-Quarter Fiscal 2002 Financial Results
Feature Story
Real Time Considerations for
Sonar/Radar Applications Running Embedded Linux
While Linux is slowly penetrating the information technology (IT) infrastructure, it has already established a leading presence in the commercial high performance computing (HPC) and embedded HPC markets. Linux is becoming the OS
of choice for embedded application developers, second only to Wind River's VxWorks. In the HPC market, the combined strengths of Linux and clusters are emerging as true "disruptive technologies" revolutionizing the landscape of high
performance computing. Linux cluster technologies are extremely attractive to DoD programs because of their maturity and broad availability on commodity processors and interconnect fabrics.
In order to perform in real-time, today's sonar/radar signal processors require concurrent processes to execute in parallel. Typical requirements are for concurrent high speed data acquisition, front end signal processing, data reduction, distributed post processing, pre-storage of data for history retrieval and output generation to a human interface or device providing control feedback to the radar or sonar.
Hard real-time synchronization and scheduling is performed outside the Linux kernel. Modern radar systems support multi-modes of operation like surveillance mode, tracking mode or both at the same time, therefore general purpose scheduling by an OS is required. The responsiveness (total latency) of the system must remain consistent with the required feedback.
Overall, the real-time characteristics of the OS should not be a primary concern for multicomputer systems because, for performance reasons, the OS is always placed outside of the "inner loop". Even the most responsive real-time kernel cannot keep up with all real time requirements of these systems that rely on control-flow mechanisms to synchronize themselves.
One primary Linux process per processor maximizes performance and efficiency. A multicomputer has many processors, each one running a primary MPI process as fast it can. If there are no latency issues, a round-robin approach is quite simple and often the most effective way to keep up with the "throughput". Each processor performs the full application with frames distributed sequentially. However, most systems cannot tolerate a latency that is proportional to the number of processes. Therefore parallel and pipelined approaches are combined to match throughput and latency requirements. Using a large number of processors in parallel on a single task reduces the effective efficiency while improving latency. Issues of data exchange, concurrent processing and I/O operations between processors become important and ultimately will limit the number of processors that can be used on any given task without encountering diminishing returns.
Performance, flexibility and reliability of data transfers across the interconnect fabric is central to the global performance of a sonar and radar platform. Rather than attempting to control these micro-tasks on the application DSP processor (requiring a very responsive real time kernel) they could be controlled by an independent network processor on each node. This is the implementation approach CSPI choose with their FastCluster product.
A nano-kernel running off a high speed local SRAM on the network processor (Myrinet LANai), controls all incoming messages at the packet level and recombines packets to deliver larger messages to the application processor while providing hand-checks between network processors across the network. To off load the application processor, the network processor's nano-kernel manages four independent direct memory access channels (DMAs) to perform concurrent in-and-out network operations at peak transfer rates of 250 MBytes/s full duplex, on each processing node.
Linux 2.4 and MPI/GM delivers determinism to real-time Sonar/Radar application. In order to assess the deterministic traits of the OS, CSPI has run different applications under VxWorks real-time kernel and Linux 2.4 kernel on the same node architecture that includes the latest PowerPC with AltiVec and Myrinet-2000 technologies.
Specifically, CSPI tested Mitre Corporation's MPI_RT_STAP benchmark to verify that the variance for each execution of a data frame is acceptable and that both Linux and VxWorks offer similar levels of performance with minor variances. FastCluster gives the Linux developer the benefit of running his application in user-space with the advantage of the full memory management protection and ease of debugging with all the hard synchronization handled inside GM.
The results of running the MPI_RT_STAP benchmark show the response latency of Linux is comparable to a real-time operating system (RTOS) and does not degrade the timing variation performance of the application. Linux based platforms, therefore, offer an open source, standards based, cost effective alternative ready to meet the critical needs of COTS applications.
Product Information
2000 SERIES Fault Tolerant Library
FTLib is an API specifically designed to meet the requirements of the Application Layer for Fault Tolerant platforms. Using leading-edge software technology FTLib allows a network of hundreds of 2000 SERIES I/O and computational nodes to safely communicate with each other. FTLib provides non-blocking, point-to-point message passing that monitors and cleanly reconfigures the network avoiding non-responsive paths and failed nodes. Designed with an application-programming interface (API) modeled on the MPI library, FTLib insulates the application programmers from network details and from node and memory management issues.
Feature Product
FastCluster Desktop 16 Plus
The FastCluster Desktop 16 Plus is a member of the CSP Inc. family of FastCluster computers that combine highly scalable performance and high-density packaging in a Linux cluster computer. As an ultra-dense cluster, FastCluster processing nodes combine the Motorola AltiVec™ technology with the high speed Myrinet-2000™ interconnect to achieve scalable performance for a broad class of High Performance Computing solutions. The FastCluster Desktop 16 Plus can be used in a broad class of environments capitalizing on its very small footprint. It is ideal for use as both a laboratory and a field based computing platform and offers expansion capabilities both within its small package size, or by connecting two or more systems together using its state-of-the-art Myrinet-2000 node to node interconnect. Because each processing node boots its Linux copy directly from its private 32 MB flash memory, requirements for local disks are eliminated, further enhancing its reliability in field environments.
High Availability High availability features are implemented in all members of the FastCluster product family to meet mission critical needs of scientific and commercial applications as well as the requirements of DOD applications. These features are specially designed to increase product availability, maintainability and to provide very fast recovery from component faults.
Self-booting from FLASH memory provides instant start-up. Live Insertion allows the replacement of a processing node without shutting down the other nodes or the network.
Modularity Up to 16 nodes can be contained in the FastCluster Desktop 16 Plus (Four MultiComputer boards plus a host processor board). This provides up to 56 GFLOPS peak performance. Additional nodes can be added with a Myrinet connection to another chassis.
Compact System Packaging The FastCluster Desktop 16 Plus provides a 5-slot chassis with VME Extensions. The enclosure has been designed with adherence to standard connections and rack-mountable dimensions to make it easy to integrate standard components in the chassis. Removable front panels, rear I/O Interface panels, and optional slides facilitate the servicing of components. The enclosure is designed to optimize system airflow. Cabling wireways, and peripheral device mounting capability are designed to make the most space efficient use of the small footprint provided by the chassis. From one to four 2841 MultiComputer Modules can be contained within the FastCluster Desktop 16 Plus, providing a desktop computer with 14 to 56 GFLOPS peak performance.
Open Source Software The Yellow Dog Linux™ operating system from Terra Soft Solutions® Inc is the foundation for the software environment for all CSPI FastCluster products. Yellow Dog Linux is a full featured software suite specifically designed for multiprocessor PowerPC systems with emphasis on robustness, high performance, low power consumption and low cost.
Linux provides an open source UNIX like operating system with a POSIX implementation including true multitasking, virtual memory, shared libraries, demand loading, work load balancing, and support for TCP/IP networking. Yellow Dog Linux is easily integrated with clustering software such as MPI and includes a full suite of GNU compiler tools to facilitate development.
CSP Inc. Reports
Third-Quarter Fiscal 2002 Financial Results
BILLERICA, Mass., July 31, 2002 -- CSP Inc. (Nasdaq: CSPI), a provider of e-business solutions, image processing software, network management and storage systems integration services and dense cluster computing systems, today reported financial results for the third quarter of fiscal 2002 ended June 30, 2002.
The Company reported sales of $7.8 million for the third quarter of fiscal 2002, compared with sales of $9.5 million in the year-earlier quarter.
For the third quarter of fiscal 2002, CSP Inc. reported a net loss of $84,000, or $(0.02) per share, including a restructuring charge of $130,000 related to severance resulting from the closure of the company's Karlsruhe, Germany office. This compares with a net loss of $2.7 million, or $0.77 per share, in the third quarter of fiscal 2001. CSP's third-quarter fiscal 2001 results reflect a $1.2 million impairment charge and a $74,000 charge for corporate restructuring. Excluding the charges, the Company reported a third-quarter, fiscal 2002 net loss of $28,000, or $0.01 per share, compared with a third-quarter, fiscal 2001 net loss of $517,000, or $0.15 per share.
CSP's closure of MODCOMP's Karlsruhe, Germany office reduced the Company's workforce by six employees and is expected to result in annualized savings of approximately $600,000. CSP continues to operate its MODCOMP subsidiaries in Cologne and Bad Hamburg, Germany.
"Protracted delays in defense appropriations and the worldwide weakness in technology spending, continue to negatively affect our financial results," stated Chairman, President and Chief Executive Officer, Alex Lupinetti.
"In our Systems segment, significant procurement delays and extended program deployment incubation periods continue to impact sales of our MultiComputer products," said Lupinetti. "On the positive side, while the current defense budget does not contain increased appropriations for programs with digital signal processing (DSP) as a key component, there appears to be interest building in the military for Linux-based DSP systems. We are well poised to capitalize on this trend with our Linux-based FastCluster MultiComputer systems, which can be used in a wide variety of military applications.".
Continued Lupinetti, "We also are encouraged by the long-term prospects for increased military spending, particularly on programs that need DSP systems for detection and tracking, stimulation/simulation, and command, control, communication and computing applications. We will continue to invest in the development of our MultiComputer product line, for both the high- and low-end of the market, in order to capitalize on future opportunities."
"During the past year, our Systems and Service Integration segment has been significantly affected by the telecommunications slowdown, primarily in Europe. However, we believe that European budgets for IT outsourcing appear to be loosening and we are hopeful that our service integration business will begin to rebound in coming quarters," Lupinetti continued.
MODCOMP recently complemented its enterprise application integration technology with the purchase of UK-based Sipher Software, a developer of customized communications solutions using message-oriented middleware," said Lupinetti. "With Sipher's technology, MODCOMP is now able to deliver large volumes of business critical data while maintaining high performance and reliability standards. Sipher also brings to MODCOMP an established presence in the financial services market with clients such as Barclays, Citibank and JP Morgan Chase."
Although we see some reason for optimism with our service integration business in Europe, we expect that the global slowdown in technology spending will continue to hamper a full recovery of both our System and Service Integration and E-Business Software segments for the near-term. Likewise, current military spending appropriations and deployment issues will continue to affect MultiComputer sales. During this difficult period, we will invest carefully and strategically in each of our business segments, while remaining resolute in our vigilance to control costs and maintain our strong balance sheet with $14.8 million in cash and no debt," concluded Lupinetti.