GE Reflective Memory PCI-5565 Multimode Fiber Optic Network

I. Introduction to Fiber Optic Networks

In real-time systems such as hardware-in-the-loop simulation systems, high real-time performance is required for data transmission between various parts of the system. Traditional network technologies, such as Ethernet and FDDI, exhibit the following shortcomings in real-time applications: (1) Low data transmission rates; (2) Poor real-time performance in data transmission, with significant and unpredictable transmission delays; (3) Network communication relies on multiple network protocols, resulting in low communication efficiency.

Real-time fiber optic reflective memory network (abbreviated as real-time fiber optic network) is a real-time network based on high-speed fiber optic network shared storage technology. Compared to traditional networking technologies, it not only offers strict transmission determinism and predictability but also features high data transmission speeds, simple communication protocols, low host load, and strong adaptability to both software and hardware platforms.

The real-time fiber optic reflective memory network (abbreviated as real-time fiber optic network) is formed by connecting fiber optic interface boards installed in computers via fiber optic cables, creating a ring network (as shown in Figure 1). The onboard memory of each node's fiber optic interface board contains a copy of shared data from other nodes. Logically, all nodes in the network share the same memory. Data written at one point is simultaneously updated at multiple points, enabling high-speed data transmission and sharing.

II. Real-Time Network Requirements

To enhance computing power, the natural approach is to develop more powerful computers, such as the Tianhe series of supercomputers. However, supercomputers have long development cycles, high costs, and limited application areas. To address this, the concept of cluster systems was proposed. A computer cluster system connects PCs or workstations via a network to form a high-performance computing system. Cluster systems execute tasks in parallel across multiple computers, forming a real-time system.

In real-time systems, the correctness of the final result depends not only on the logical outcome of each computation step but also on the timing of the result. The completion time of tasks is a decisive characteristic of real-time systems. Based on the degree of real-time performance requirements, real-time systems can be classified into two types: soft real-time and hard real-time systems. For soft real-time systems, event responses are required to be real-time but are not strictly enforced. However, for hard real-time systems, each task has a processing deadline and must be completed within the specified time; otherwise, it may affect the completion of global tasks, causing undesirable damage or irreversible catastrophic consequences. Many real-time systems currently adopt hard real-time systems due to their stronger real-time performance. The application of real-time systems requires real-time interconnection, building real-time networks to achieve real-time data transmission between network nodes.

Real-time networks must possess three characteristics: high speed, reliability, and predictability. The most important of these is communication predictability, which means that the data transmission time between nodes in a real-time network is deterministic. With the expanding application areas of real-time networks, they are no longer limited to interconnecting computer cluster systems but are widely used in various interconnected systems with real-time requirements, such as hardware-in-the-loop simulation and high-speed data acquisition.

During the development of real-time networks, two design approaches have emerged: multi-CPU shared global memory based on a single bus and distributed memory based on networks.

Compared to multi-CPU shared global memory systems based on a single bus, in distributed memory systems based on networks, each node only accesses its own local memory, eliminating memory access conflicts. This approach avoids significant bus arbitration control time and has become the primary design method for real-time networks.

Real-time networks must ensure high speed, reliability, and predictability. However, technologies such as Gigabit Ethernet and Fiber Channel (FC) protocols perform poorly in terms of predictability. This is primarily because the working mode of device drivers and software network protocols introduces additional uncertain overhead. Secondly, issues such as shared conflicts and collisions (e.g., CSMA/CD carrier sense multiple access with collision detection in Gigabit Ethernet GBE and token ring arbitration in FC protocols) further increase the uncertainty of transmission time. Fiber optic reflective memory networks effectively address these issues.

Reflective memory networks are a design concept for hard real-time networks based on distributed memory over networks. Their characteristics include: 1) Reflective memory networks hardware-implement data transmission and sharing between distributed RAMs, resulting in very low implementation latency; 2) Reflective memory networks rely on hardware implementation and do not require complex network protocol control, achieving higher effective rates under the same transmission bandwidth; 3) Various data processing in reflective memory networks is implemented through hardware circuits, operating periodically at fixed clock frequencies without uncertain time overhead, ensuring predictability in data processing.

III. Comparison Between Fiber Optic Reflective Memory Networks and Ethernet

1) Real-Time Performance

Reflective Memory (RFM) is a high-speed, replicated shared memory network based on ring/star topology. It supports multi-computer systems with different bus structures and allows the sharing of high-speed, stable-rate real-time data across different operating systems.

Real-time networks built on reflective memory are a strong real-time, high-bandwidth local area network technology that provides efficient data transmission between interconnected computers. Reflective memory networks virtualize a segment of globally shared network memory across all interconnected nodes, enabling memory-to-memory communication in distributed systems without software overhead for applications. Each node computer is equipped with a reflective memory card featuring dual-port memory. Application software at various layers of each node computer can directly read and write the memory on the reflective memory card. When data