DCI Architecture Promoted by the IOWN Global Forum

2.1 Challenges in scalability

ICT infrastructures need to support various types of data processing with different requirements. For example, online transaction processing needs to respond to a large number of inquiries while data batch processing must handle vast amounts of data. The data-processing requirements of ICT infrastructures can also change rapidly over time. For example, an event, such as a sale, generates a large number of access attempts to an Internet shopping site, requiring ICT infrastructures to respond to these changes within minutes or hours. Therefore, ICT infrastructures must have sufficient scalability to flexibly respond to such changes in demand.

2.2 Challenges in performance

Some types of data processing, such as communicating the movements of individual participants in a virtual space to the people around them or processing high-speed transactions in financial applications, must meet stringent response-time requirements. Current ICT infrastructures cannot adequately satisfy such exacting requirements, especially in data transfer.

2.3 Challenges in energy consumption

Today’s ICT infrastructures have various bottlenecks. For example, if data transfer is a bottleneck, the central processing unit (CPU) consumes most of its processing power simply waiting for data transfer. It has become common to allocate processing to accelerators, which are adept in areas that are different from those of the CPU, such as graphics processing units (GPUs), which specialize in image processing to achieve high-speed processing, and data processing units (DPUs), which specialize in specific types of computation such as network processing and security processing. Efficient use and appropriate role sharing of computing resources can optimize the power efficiency of the entire ICT infrastructure.

3.1 What is the Open APN?

The Open APN is a high-speed network that consists entirely of optical communication. It is characterized by its ability to provide a high-capacity network with low latency, particularly with predictable latency. Therefore, it can transfer data between various devices connected to the network as if these devices were close to each other, regardless of their locations and the distances between them. More information on the Open APN and NTT’s related activities are given in the article, “Activities for Detailing the Architecture of the Open APN and Promoting Its Practical Application,” in this issue [3].

3.2 What is a data pipeline?

A data pipeline is a uniform stream that executes functions, such as acquisition, processing, conversion, and presentation of the data required by various applications. Figure 2 shows an example of a data pipeline for a system that aggregates and analyzes surveillance-camera footage on the basis of metadata [4]. In this example, the footage from a surveillance camera is sent to an analysis server, which analyzes information contained in the metadata, such as information about what is captured in the video. The metadata are sent to the analysis server for the purpose of data aggregation, and raw video data are stored in storage units. Using a terminal, analysts can analyze data on the basis of the metadata and refer to the actual footage data in the storage units. In Fig. 2, each oval represents data processing, i.e., a set of autonomous actions, such as data collection and analysis. A process is executed by a functional node, represented with a rectangle, which exists outside the process. A database/storage, represented with a drum, is typically a file system or database. It is a function for storing data. Connecting these are data flows, which are functions to transfer data to data processing or storage. The IOWN Global Forum has defined this representation of a data pipeline as a data-pipeline diagram [4]. Therefore, a data pipeline is an abstraction of resources used to enable an application to function.

Fig. 2. Example of a data pipeline of a system that aggregates and analyzes surveillance-camera footage on the basis of the metadata.

4.1 Functional cards, physical nodes, and DCI clusters

A schematic diagram of the various resources managed by the DCI is shown in Fig. 3.

As mentioned above, data processing refers to the function of collecting or analyzing data in the data pipeline and is typically implemented using CPUs. GPUs, which specialize in image processing to achieve high-speed processing, and DPUs, which specialize in specific types of computation, such as network processing and security processing, have been attracting attention. Storage is implemented using solid-state drives (SSDs) and hard disk drives.

Fig. 3. Schematic of various resources managed by the DCI.

In the DCI, data processing and storage are classified into two categories: host board and functional card. A host board typically consists of a CPU and memory unit. It corresponds to a motherboard in a typical server. Functional cards are other components, such as GPUs, DPUs, network interfaces, and storage units. A functional card corresponds to an expansion board in a general server, such as a network interface controller (NIC) or a graphics card, or to a storage unit such as an SSD. A DCI physical node consists of a host board and multiple functional cards. An intra-node interconnect connects the host board and functional cards within a DCI physical node. Peripheral Component Interconnect Express (PCIe), which is widely used in servers, is an example of an intra-node interconnect. However, more flexible standards, such as Compute Express Link (CXL), are being studied.

A DCI cluster is a computing infrastructure consisting of multiple physical nodes. It also includes the characteristic mechanisms of an inter-node interconnect and DCI gateway.

The inter-node interconnect is a network that connects DCI physical nodes. Although general clusters also have a network that connects servers that make up the cluster, the inter-node interconnect differs from such a network in that some functional cards can directly access inter-node interconnects. Functional cards that can directly access inter-node interconnects are called “network-capable” functional cards. Such a card can directly transfer data to and from functional cards and host boards residing in other physical nodes without going through the intra-node interconnect of the physical node to which it belongs. An NIC is of course a network-capable functional card. It is assumed that some GPUs and storage units and some functional cards are also network-capable.

The DCI gateway interconnects DCI clusters. It is connected to other DCI clusters via the Open APN. Network-capable functional cards can also directly access the DCI gateway and transfer data to and from functional cards belonging to other DCI clusters via the Open APN at high speed.

4.2 Logical service node and data-pipeline implementation

How a DCI cluster can create a data pipeline for an application is explained. To create a data pipeline, a DCI cluster selects functional cards and host boards within itself and builds a logical service node (LSN) from them.

If an LSN consists of elements within a single physical node, all data flows in a data pipeline can be transferred via the intra-node interconnect in that physical node. However, if the LSN spans multiple physical nodes, an inter-node interconnect is required to transfer the data flow. A data pipeline can be created using multiple LSNs that belong to different DCI clusters. In this case, data flows across LSNs must go through a DCI gateway.

Figure 4 shows an example of a data-pipeline configuration. This DCI cluster has selected the functional cards shown in yellow, i.e., Functional cards A, B, and C, to configure the LSN. Not being network-capable, Functional Card A cannot directly connect to the inter-node interconnect. Consequently, it connects to the inter-node interconnect via the intra-node interconnect and Functional Card X.

Fig. 4. Example of data-pipeline configuration.

Functional Card B of DCI physical node #2, however, is network-capable. Therefore, the inter-node interconnect can be directly connected to Functional Card B. Functional Card C is also network-capable and is connected to the DCI gateway. This means Functional Card C can connect to another LSN in another DCI cluster via the DCI gateway.

As described above, a feature of the DCI is the ability to configure flexible data pipelines across DCI physical nodes and DCI clusters.

5.1 Addressing challenges in scalability

In conventional ICT infrastructures, computing resources and wide-area networking resources for data transfer have evolved independently, making flexible scaling across physical nodes and clusters difficult. With the DCI, however, network-capable functional cards are directly interconnected via inter-node interconnects, thus eliminating data-transfer bottlenecks. LSNs in different DCI clusters can also be easily integrated via DCI gateways. The integration of resources across physical nodes and clusters in this way makes for high scalability.

5.2 Addressing challenges in performance

With the DCI, network-capable functional cards are directly interconnected via inter-node interconnects. Thus, data can be transferred between them at extremely high speed without going through a CPU. Because data can also be transferred across DCI clusters at high speed via the Open APN, applications can have sufficient response performance with stringent response-time requirements.

5.3 Addressing challenges in energy consumption

With the DCI, because there are few bottlenecks in data transfer, CPUs are not kept waiting needlessly for data transfer. Accelerators, such as GPUs and DPUs, can also be selected flexibly depending on the computing requirements. This improves computing-resource utilization and makes it possible to optimize overall system energy consumption.