## Data Centers and "The Cloud" > [!quote] > _There is no cloud; it's someone else's computer._ ### Edge Computing Finding a definition of edge computing is a tricky task, as it is one of those terms that have been manipulated extensively in recent years, for marketing purposes. By definition, edge computing is a distributed computing paradigm that brings computation and data storage closer to the sources of data. This approach is in contrast to more traditional cloud computing architectures, where data processing happens in centralized data centers, which can be located far from the data source. The essence of edge computing lies in its ability to process data near the edge of the network, closer to where the data is generated, rather than sending vast amounts of data across the network to a central data center for further processing. This proximity to data sources offers several key advantages. Firstly, working at the edges reduces latency or the time it takes for a data packet to travel between its source and destination. This is particularly important for applications requiring real-time processing and decision-making, such as autonomous vehicles, industrial automation, and certain healthcare technologies. By processing data locally, these systems can react more quickly to input, which is of relevance for performance and safety. Secondly, edge computing can reduce bandwidth needs. When data is processed locally, only relevant, processed information needs to be sent over the network to central servers or cloud services. This is the equivalent of sending a raw file or a compressed file over a network. If, for instance, you captured an image with a camera device in bitmap (BMP) format, that image will be composed of every pixel as captured by the sensor. If the image is color, it will have a certain number of bits assigned per color (red, green, and blue). A single bitmap file for high-resolution cameras can be several gigabytes of data. If instead of sending the BMP file over the network we send a JPEG, the amount of resources (storage, bandwidth) will be significantly smaller due to the fact the JPEG is a compressed format. If we located the BMP-to-JPEG image processor as close to the sensor as possible, that would count as edge processing. An extreme case of this example would be a processor that can detect features in the image. For instance, imagine that the processor is coded to detect the presence of a horse in an image; if there is a horse in the field of view, the processor would output a "true" flag. Otherwise, the processor would return false. In this case, the resources needed to manipulate the information out of the edge processor would dramatically reduce: from a full-blown bitmap file of several gigabytes down to 2 bits. As silly as the example may be, it paints the picture of the advantage of processing things closer to the sources. Additionally, edge computing enhances privacy and security. By processing sensitive data locally, instead of sending it across a network to a central server, there's less risk of interception or unauthorized access. This is particularly relevant for industries that handle sensitive data, like healthcare or finance. However, the lunch price again: edge computing also presents challenges. It requires more complex management and coordination of computing resources, as processing tasks are distributed across numerous edge locations. This can involve significant logistical and technical complexities in deploying and maintaining a large number of edge computing nodes.  >[!Figure] >Edge concept (source: #ref/Geng ) Most companies will utilize both edge computing and cloud computing environments for their business requirements, as edge computing should be seen as complementary to cloud computing—real‐time processing of data locally at the device while sending select data to the cloud for analysis and storage. The adoption of edge computing still is not a smooth path. IoT device adoption has shown to have longer implementation durations with higher costs than expected. Especially the integration into legacy infrastructure has seen significant challenges, requiring heavy customizations. #### Edge Data Center Architecture The data center industry, including makers of computer cabinets, is coming to market with a wide variety of edge data center designs. There is already an impressive number of options. The selection will likely grow as time goes on. The following is a rundown of today's leading approaches to edge data center design: - Purposely built structure: Organizations might choose to build edge data centers as purpose-built structures. They are constructing a building of suitable size to hold whatever edge computing capacity they require and installing server racks, cooling equipment, and power backup equipment. The advantage of this approach is that the owner can design it for their exact requirements. The downside is cost and repeatability. This is an expensive route to take if an organization needs to build, say 1,000 such sites. - Existing structure: In this approach, the owner can place a data center in a building that wasn't designed to be a data center. For example, if you have a vacant floor in an office building, you could bring in your server racks and cooling systems and set up a small edge data center. This is not too different from the traditional server room that predominated in the pre-cloud era. For financial services companies, such setups have long been the norm. They want computing to be as near to the trading floor as possible. The limitations in this model have to do with structural issues and the availability of electrical power. Most buildings were not built to support fully loaded server racks, which can place up to 5,000 pounds of weight in six square feet of floor space. That could cause the floor to collapse. Most buildings are not designed for the kind of power and cooling required for even a small data center. With the edge, what may be the most likely use of this approach will be the conversion of small utility buildings near cell towers into micro data centers. Or a cell tower company might adapt a building close to the tower to serve as a data center. - Self-contained pod or shelter: In the last few years, several companies have developed edge data centers in self-contained pods. These vary in size and specification, but two examples offer a glimpse of what's available on the market. Modular data centers come in a range of sizes. They can hold up to 10 52U [[Units, Chassis and Racks#Rack Units|racks]] and support up to 350 kW of power in their IT load. It functions as a complete data center, with a secure entrance, server racks, and built-in cooling. - Portable / Mobile: Some edge cases require the temporary presence of an edge data center. Examples include oil and gas production, seasonal agriculture, and military deployments. For these uses, a portable edge data center might be the best option. Some of these designs are designed to be transported on a road trailer. The trailer can carry electrical generators to keep the equipment and cooling systems running. - Industrial container: Indoor and industrial edge computing scenarios may make existing computer containers serve as edge data centers. For example, the Chatsworth Products RMR Modular can hold a rack up to 47U inside. For an industrial facility, that could support more than enough computer hardware for their edge computing needs. - All-weather cabinet: These are specially designed protective enclosures that house the equipment, ensuring their operational integrity and security in various outdoor environmental conditions. These cabinets are engineered to withstand extreme temperatures, whether scorching heat or freezing cold, and are often equipped with climate control systems such as heating, ventilation, and air conditioning (HVAC) to maintain optimal internal conditions. In addition to temperature control, all-weather cabinets protect against rain, snow, dust, and other particulate matter, thus preventing damage to the sensitive electronic components inside. These cabinets are essential for data centers that require outdoor installations or have components of their infrastructure located outside, such as telecommunications companies, airports, internet service providers, and organizations with remote or distributed IT systems. The robust design also includes security features to deter unauthorized access and vandalism, ensuring that the housed technology remains safe and operational around the clock. # Cloud Computing There are a few typical characteristics of cloud computing that are important to understand: - Available over the network: Cloud computing capabilities are available over the network by a wide range of devices including mobile phones, tablets, and PC workstations. While this seems obvious, it is an often overlooked characteristic of cloud computing. - Rapid elasticity: Cloud computing capabilities can scale rapidly outward and inward with demand (elastically), sometimes providing the customer with a sense of unlimited capacity. The elasticity is needed to enable the system to provision and clear resources for shared use, including components like memory, processing, and storage. Elasticity requires the pooling of resources. - Resource pooling: In a cloud computing model, computing resources are pooled to serve multiple customers in a multi‐tenant model. Virtual and physical resources get dynamically assigned based on customer demand. The multi‐tenant model creates a sense of location independence, as the customer does not influence the exact location of the provided resources other than some higher‐level specification like a data center or geographical area. - Measured service: Cloud systems use metering capabilities to provide usage reporting and transparency to both user and provider of the service. The metering is needed for the cloud provider to analyze consumption and optimize usage of the resources. As elasticity and resource pooling only work if cloud users are incentivized to release resources to the pool, metering by the concept of billing acts as a financial motivator, creating a resource return response. - On‐demand self‐service: The consumer can provision the needed capabilities without requiring human interaction with the cloud provider. This can typically be done by a user interface (UI), using a web browser, enabling the customer to control the needed provisioning, or by an application programming interface (API). APIs allow software components to interact with one another without any human involvement, enabling easier sharing of services. Without the ability to consume cloud computing over the network, using rapid elasticity and resource pooling, on‐demand self‐service would not be possible. ## Cloud Computing Service and Deployment Models Cloud computing helps companies focus on what matters most to them, with the ability to avoid non‐differentiating work such as procurement, maintenance, and infrastructure capacity planning. As cloud computing evolved, different service and deployment models emerged to meet the needs of different types of end users. Each model provides different levels of control, flexibility, and management to the customer, allowing the customer to choose the right solution for a given business problem.  > [!Figure] > _Ownership levels in the IT stack_ (source: #ref/Geng ) The different service models are: - Infrastructure as a Service: (IaaS) allows the customer to rent basic IT infrastructure including storage, network, OS, and computers (virtual or dedicated hardware), on a pay‐as‐you‐go basis. The customer can deploy and run its own software on the provided infrastructure and has control over OS, storage, and limited control of select networking components. In this model, the cloud provider manages the facility to the virtualization layers of the IT stack, while the customer is responsible for the management of all layers above virtualization. - Platform as a service: (PaaS) provides the customer with an on‐demand environment for developing, testing, and managing software applications, without the need to set up and manage the underlying infrastructure of servers, storage, and network. In this model, the cloud provider operates the facility to the runtime layers of the IT stack, while the customer is responsible for the management of all layers above the runtime. - Software as a service: (SaaS) refers to the capability to provide software applications over the Internet, managed by the cloud provider. The provider is responsible for the setup, management, and upgrades of the application, including all the supporting infrastructure. The application is typically accessible using a web browser or other thin client interface (e.g. smartphone apps). The customer only has control over a limited set of application‐specific configuration settings. In this model, the cloud provider manages all layers of the IT stack. The typical deployment models are: - Public Cloud: Public cloud is owned and operated by a cloud service provider. In this model, all hardware, software, and supporting infrastructure are owned and managed by the cloud provider, and it is operated out of the provider’s data center(s). The resources provided are made available to anyone for use, based on pay-as-you-go or for free. Examples of public cloud providers include AWS, Microsoft Azure, Google Cloud, and Salesforce.com - Private Cloud: Private cloud refers to cloud computing resources provisioned exclusively by a single business or organization. It can be operated and managed by the organization, by a third party, or a combination of them. The deployment can be located at the customer’s own data center (on-premises) or in a third‐party data center. The deployment of computing resources on-premises, using virtualization and resource management tools, is sometimes called “private cloud.” This type of deployment provides dedicated resources, but it does not provide all of the typical cloud characteristics. While traditional IT infrastructure can benefit from modern virtualization and application management technologies to optimize utilization and increase flexibility, there is a very thin line between this type of deployment and true private cloud. - Hybrid Cloud: A hybrid cloud is a combination of public and private cloud deployments using technology that allows infrastructure and application sharing between them. The most common hybrid cloud use case is the extension of on‐premises infrastructure into the cloud for growth, allowing it to utilize the benefits of the cloud while optimizing existing on‐premises infrastructure. Most enterprise companies today are using a form of hybrid cloud. Typically they will use a collection of public SaaS‐based applications like Salesforce, Office 365, and Google Apps, combined with public or private IaaS deployments for their other business applications. - Multi‐cloud: As more cloud providers entered the market within the same cloud service model, companies started to deploy their workloads across these different provider offerings. A company may have compute workloads running on AWS and Google Cloud at the same time to ensure a best-of-breed solution for their different workloads. Companies are also using a multi-cloud approach to continually evaluate various providers in the market or hedge their workload risk across multiple providers. Multi‐cloud, therefore, is the deployment of workloads across different cloud providers within the same service model (IaaS/PaaS/SaaS). > [!warning] > To be #expanded ## Architecture Technological development in the field of data centers and infrastructure relating to cloud computing is split between two areas: on‐premises deployments and public cloud provider deployments. The on‐premises deployments, often referred to as private cloud, have either been moving to standard [[Units, Chassis and Racks|rackmount]] server and storage hardware combined with new software technology like OpenStack or Microsoft Azure Stack or more packaged solutions. As traditional hardware deployments have not always provided customers with the cloud benefits needed due to management overhead, converged infrastructure solutions have been getting traction in the market. A converged infrastructure solution packages networking, servers, storage, and virtualization tools on a turnkey appliance for easy deployment and management. As more and more compute consumption moved into public cloud computing, a lot of technical innovation in the data center and IT infrastructure has been driven by the larger public cloud providers in recent years. Due to the unprecedented scale at which these providers have to operate, their large data centers are very different from traditional hosting facilities. Individual “pizza box” servers or single server applications no longer work in these warehouses full of computers. By treating these extensive collections of systems as one massive warehouse‐scale computer (WSC), these providers can achieve levels of reliability and service performance that businesses and customers nowadays expect. To support thousands of physical servers, in these hyperscale data centers, cloud providers had to develop new ways to deploy and maintain their infrastructure, maximizing the compute density while minimizing the cost of power, cooling, and human labor. If one were running a cluster of 10,000 physical servers, that would require stellar staff [[Reliability Assessment Methods|reliability]] for the hardware components used; it would still mean that in a given year on average, one server would fail every day. To manage hardware failures in WSCs, cloud providers devised alternative rack server designs to enable more straightforward swap out of failed servers and generally lower operational costs. As part of a larger interconnected system, WSC servers are low‐end server-based, built-in tray or blade enclosure format. Racks that hold together tens of servers and supporting infrastructure like power conversion and cluster servers into a single rack compute unit. The physical racks can be completely custom-designed by the cloud provider enabling specific applications for computing, storage, or machine learning applications. ## The Open Compute Project The Open Compute Project^[https://www.opencompute.org/] contains detailed specifications of the racks and hardware components used by companies like Facebook, Google, and Microsoft to build their WSCs. Though not a standards body, the OCP Community forges new technology norms, helping thus grow an ecosystem where industry players collaborate in a safe framework, shaping a diverse supply chain. The OCP is an industry nonprofit focused on establishing open-source hardware, software, and Systems Management solutions for the Data Center market. The OCP community actively works on a set of projects and sub-projects in different areas and domains: - Networking - Rack & Power - Storage - Server - Telecommunications - DC Facility - Hardware management - System Firmware - Security > [!warning] > Section under #development # The Energy Conundrum See this: https://semianalysis.com/2024/10/14/datacenter-anatomy-part-1-electrical/ > [!warning] > To be #expanded # The Thermal (Cooling) Conundrum See this: https://semianalysis.com/2025/02/13/datacenter-anatomy-part-2-cooling-systems/ > [!warning] > To be #expanded