Name: 英偉達™（NVIDIA®）網絡技術：數據中心設計注意事項 (NVIDIA Networking: Data Center Design Considerations)
Uploaded: 2025-10-13T23:58:10.000Z
Duration: 17 min 59 s

Welcome to the Data Center Design Considerations course.

歡迎來到資料中心設計考量課程。

The purpose of this unit is to equip you with the basic knowledge to understand the main data center requirements and how they can be fulfilled.

這個單元的目的是讓大傢俱備基本知識，瞭解主要的資料中心需求以及如何滿足這些需求。

We'll start with defining what is a network topology and what are the considerations when choosing the right topology for one's needs.

我們會先定義什麼是網路拓撲，以及選擇適合自己需求的拓撲時需要考量的因素。

Then we'll present two design the hierarchical network design that is considered the traditional design and the leaf spine used in modern data centers.

接著，我們會介紹兩種設計：被認為是傳統設計的分層網路設計，以及現代資料中心使用的葉脊架構。

Later, we'll compare layer two versus Layer three designs and understand how their behavior and protocols impact network performance.

之後，我們會比較第二層和第三層的設計，並瞭解它們的行為和協定如何影響網路效能。

Let's start with an overview of general data center design considerations.

我們先來概覽一下資料中心的一般設計考量。

A data center is a facility used to house computer systems and associated components such as compute, network and storage systems.

資料中心是一個用於存放電腦系統及其相關元件的設施，例如計算、網路和儲存系統。

A data center can be the size of a small server room to a few buildings.

資料中心的大小可以從一個小型伺服器機房到幾棟建築物不等。

Geographically distributed, the data center is a critical business asset where companies run their workloads.

地理分佈式的資料中心是企業運行工作負載的關鍵商業資產。

The modern data center has evolved from a facility containing an on premises infrastructure to one that connects on premises systems with cloud infrastructures where networks, applications and workloads are virtualized in multiple private and public clouds.

現代資料中心已從一個包含本地基礎設施的設施演變成一個連接在地系統與雲端基礎設施的設施，其中網路、應用程式和工作負載在多個私有雲和公有雲中進行虛擬化。

Data centers require scalability, resiliency and security.

資料中心需要可擴展性、彈性和安全性。

The modern data center is also becoming more technologically advanced to provide better efficiency.

現代資料中心也變得越來越先進，以提供更好的效率。

Data center design is the process of planning all of a data center's essential computational and non computational parameters.

資料中心設計是規劃資料中心所有重要的計算和非計算參數的過程。

Some are listed number and type of required servers, network layout, power, cooling and ventilation systems, physical data center security, disaster recovery and business continuity planning.

其中一些包括所需的伺服器數量和類型、網路佈局、電力、冷卻和通風系統、實體資料中心安全、災難恢復和業務連續性規劃。

Modern data center design will incorporate many different aspects that will make it easy to manage and more efficient.

現代資料中心設計將融入許多不同的方面，使其易於管理且更有效率。

We won't cover in this unit all data center design aspects, but we will focus on networking considerations.

在這個單元中，我們不會涵蓋所有資料中心設計方面，但我們會專注於網路考量。

One important design aspect is the network layout or topology.

一個重要的設計方面是網路佈局或網路拓撲。

A network topology is the schematic arrangement of elements such as links and nodes, which ultimately constitutes the structure of the network.

網路拓撲是節點和連結等元件的示意性佈局，最終構成網路的結構。

It describes not just how devices on the network are connected, but

它不僅描述了網路上的裝置如何連接，

also how data moves from one node to another.

還描述了資料如何從一個節點移動到另一個節點。

The following are key considerations when choosing a network.

以下是選擇網路時的關鍵考量。

Availability Enterprises that run critical applications need maximum availability of their network resources.

可用性 運行關鍵應用程式的企業需要其網路資源的最大可用性。

High availability features must be considered throughout the network.

高可用性功能必須在整個網路中加以考量。

Hardware and software, network protocols, as well as environmental and power features all contribute to the overall availability of the network.

硬體和軟體、網路協定，以及環境和電源功能都對網路的整體可用性有所貢獻。

Reliability in many of the industries, even brief downtime and delays are unacceptable.

可靠性 在許多產業中，即使是短暫的停機和延遲也是無法接受的。

For this reason, network reliability is a fundamental consideration.

是以，網路可靠性是一項基本考量。

Performance the topology of a network is key to determining its performance.

效能 網路的拓撲是決定其效能的關鍵。

Choosing the right topology can increase performance while making it easier to locate faults, troubleshoot errors, and

選擇正確的拓撲可以提高效能，同時更容易定位故障、排除錯誤，並

more effectively allocate resources across the network.

Future Growth if the network is expected to grow in size in the mid to long term.

未來成長 如果預期網路在中長期會擴大規模。

Choose a network topology that's easy to add new nodes without negatively affecting performance or the user experience.

選擇一個容易新增節點，且不會影響效能或使用者體驗的網路拓撲。

You can choose a topology that is perfectly suited to your needs, but it may not be affordable.

你可以選擇一個完全符合你需求的拓撲，但價格可能不親民。

There will probably be a lower priced alternative that's nearly as effective.

可能會有一個價格較低、效果也差不多的替代方案。

The hierarchical design separates a network into distinct layers where each layer has its role in the network.

分層設計將網路劃分為不同的層級，每一層級在網路中都有其角色。

This hierarchical network design is most effective for north south traffic patterns where a majority of the data goes in and out of the data center, such as client server applications.

這種分層網路設計最適合南北向的流量模式，也就是大部分資料進出資料中心，例如客戶端伺服器應用程式。

The model scales somewhat well because of its modularity, but

這個模型因為模組化，所以有一定程度的可擴展性，但

may be subject to bottlenecks if uplinks between layers are oversubscribed.

如果層級之間的上行鏈路超載，可能會出現瓶頸。

Cloud computing, virtualization, Big data and related trends have come together to transform data centers.

雲端計算、虛擬化、大數據及相關趨勢共同改變了資料中心。

Modern data center environments are becoming much more dependent on east west traffic as more data moves between servers and storage nodes in order to meet end user demands.

現代資料中心環境越來越依賴東西向流量，因為為了滿足終端使用者需求，資料在伺服器和儲存節點之間的移動越來越多。

The network design needs to serve the modern data center's increasing demand for server to server communication, scalability and resiliency in the past few years, many data centers have transitioned to a leaf spine design to overcome the limitations of the hierarchical design.

網路設計需要滿足現代資料中心日益增長的伺服器對伺服器通訊、可擴展性和彈性的需求。過去幾年，許多資料中心已轉向葉脊架構，以克服分層設計的限制。

A leaf spine architecture is a two layer full mesh topology composed of a leaf layer and a spine layer.

葉脊架構是一種兩層的全網狀拓撲，由葉層和脊層組成。

The spine switches connect the leaf switches with one another, whereas the leaf switches connect servers to the network.

脊交換器將葉交換器彼此連接，而葉交換器則將伺服器連接到網路。

For that reason, the leaf switches are also referred to as the top of the rack switches or TORs.

是以，葉交換器也被稱為機架頂部交換器或 TOR。

Every leaf switch is connected to every spine switch, and obviously vice versa.

每個葉交換器都連接到每個脊交換器，反之亦然。

The leaf spine architecture has many advantages, as listed here.

葉脊架構有許多優點，如下所示。

If the network reaches a certain scale, a third level of switches should be considered.

當網路達到一定規模時，應考慮第三層交換器。

In this case, the upmost level, called the superspine or core, is used to interconnect the spine switches.

在這種情況下，最上層稱為超脊或核心層，用於互連脊交換器。

Now that we know that the leaf spine design is the best choice for modern data centers, let's see whether the network should be configured to operate at Layer two or Layer three.

既然我們知道葉脊設計是現代資料中心的最佳選擇，讓我們看看網路應該配置為在第二層還是第三層運行。

In a Layer two network, networking devices are configured for switching, meaning they make Layer two forwarding decisions.

在第二層網路中，網路設備會被設定為交換，也就是說它們會做出第二層的轉送決策。

They build and maintain a Mac address table that maps Mac addresses to exit ports.

它們會建立並維護一個 MAC 位址表，將 MAC 位址對應到出口埠。

The switches process ethernet frames and forward them based on the destination Mac address that is matched against the Mac address.

交換器會處理乙太網路訊框，並根據與 MAC 位址表比對到的目標 MAC 位址來進行轉送。

Table Overall, a layer 2 network does not scale well and provides lower bandwidth and performance than a Layer three network.

總體來說，第二層網路的可擴展性不佳，並且與第三層網路相比，頻寬和效能較低。

Also, it's harder to achieve redundancy and multipath support.

此外，實現備援和多路徑支援也比較困難。

We'll explore the reasons for that in the following slides.

我們會在接下來的投影片中探討原因。

Let's focus on redundancy and multipath support A redundant network design allows for network availability by duplicating elements such as switches or links, so that if an element fails, critical applications are not disrupted.

讓我們來看看備援和多路徑支援。備援的網路設計透過複製交換器或連線等元素來確保網路的可用性，這樣一來，如果某個元素發生故障，關鍵應用程式也不會中斷。

Redundancy attempts to eliminate any single point of failure in the network.

備援試圖消除網路中任何單點故障。

In this example, each leaf switch has a link to each of the spine switches so that there are two available paths for the host to communicate.

在這個例子中，每個葉交換器都與每個脊交換器都有一個連線，是以主機有兩條可用的路徑進行通訊。

Let's assume that when host A communicates with host C, the path is via Spine 1.

我們假設當主機 A 與主機 C 通訊時，路徑是透過脊 1。

But what happens if the link between Leaf 1 and Spine 1 fails?

但如果葉 1 和脊 1 之間的連線發生故障，會發生什麼事？

There is another available path via Spine 2 that can now be used for traffic forwarding.

透過脊 2 還有另一條可用的路徑，現在可以用於流量轉發。

This is the idea of a redundant network design.

There is an active path and a backup path.

有一條活躍路徑和一條備用路徑。

If the active path fails, the backup path takes over.

如果活躍路徑發生故障，備用路徑就會接管。

But if there are two available paths, why not use them both at the same time?

但如果兩條路徑都可用，為什麼不同時使用它們呢？

This will allow for load balancing the traffic, hence providing twice the bandwidth and better resource utilization.

這將允許流量負載平衡，從而提供兩倍的頻寬和更好的資源利用率。

This is the advantage of multipath support.

Multipath is a strategy where traffic forwarding to a single destination can occur over multiple paths.

多路徑是一種策略，其中到單一目的地的流量轉發可以透過多條路徑進行。

It can substantially increase bandwidth by load balancing traffic over multiple paths.

透過在多條路徑上進行流量負載平衡，它可以大幅增加頻寬。

However, it may be difficult to deploy it, especially in a layer 2 network.

不過，部署起來可能會有點困難，尤其是在第二層網路中。

We already know that a redundant network design Requires additional redundant layer 2 links to provide a backup path in case of a link or switch failure.

我們已經知道，一個備援的網路設計需要額外的備援第二層連結，以便在連結或交換器發生故障時提供備用路徑。

Those redundant links allow multiple paths between a pair of nodes.

這些備援連結允許節點對之間有多條路徑。

Well, that sounds terrific, because that's exactly what we're trying to achieve.

嗯，聽起來很棒，因為這正是我們想要達成的目標。

But because of the behavior of layer two networks, this creates a layer two loop that eventually results in traffic denial.

但由於第二層網路的行為，這會產生一個第二層迴路，最終導致流量被拒絕。

You probably recall from the previous unit how broadcast frames are handled by the switches.

你可能還記得上一單元中交換器如何處理廣播訊框。

The broadcast frames circulate endlessly in the network, consuming all available bandwidth.

廣播訊框會在網路上無止盡地循環，耗盡所有可用頻寬。

This phenomenon is called a broadcast storm and it must be avoided.

這種現象稱為廣播風暴，必須避免。

Otherwise bandwidth is unavailable for normal network traffic.

否則，正常網路流量將無法使用頻寬。

So the simple conclusion is that a redundant layer 2 network design causes broadcast storms.

所以簡單的結論是，冗餘的第二層網路設計會導致廣播風暴。

Does that mean we can't have a redundant layer two network?

這是不是意味著我們不能有冗餘的第二層網路？

We can and should have a redundant physical design.

我們可以，也應該有冗餘的實體設計。

But logically there can only be one active path at a time between any pair of end nodes.

但在邏輯上，任何一對終端節點之間一次只能有一個作用中的路徑。

We'll now present the protocol that provides this capability.

我們現在要介紹提供這項功能的協定。

Spanning Tree Protocol, or stp, ensures a loop free logical topology for Ethernet networks.

生成樹協定，也就是 STP，為乙太網路確保了無迴圈的邏輯拓撲。

It allows a network design that includes redundant links and provides automatic backup paths if an active link fails.

它讓網路設計包含冗餘的連結，並在作用中的連結失效時提供自動備份路徑。

STP achieves that by identifying redundant links and putting the redundant ports in blocking state.

STP 透過識別冗餘連結並將冗餘連接埠設為封鎖狀態來達成此目的。

Then, when a topology change occurs, SDP reacts and moves block ports to forwarding state.

然後，當拓撲發生變更時，STP 會做出反應，將封鎖的連接埠移至轉送狀態。

It is important to mention that the convergence time of SDP is between 30 to 50 seconds, during which data traffic is not served.

值得一提的是，STP 的收斂時間約在 30 到 50 秒之間，在這段期間資料流量無法傳輸。

Maybe this convergence time was acceptable in the 1980s when STP was initially developed, but

或許在 1980 年代 STP 最初開發時，這個收斂時間是可以接受的，但

no modern data center can tolerate such a long convergence time.

沒有任何現代資料中心能容忍這麼長的收斂時間。

Rapid Spanning Tree Protocol, or RSTP was introduced in the 1990s to provide rapid recovery after a topology change.

快速生成樹協定，也就是 RSTP，在 1990 年代被引入，以在拓撲變更後提供快速恢復。

There are different enhancements to support faster convergence, and we won't discuss the details here.

有許多不同的增強功能來支援更快的收斂，我們在此不討論細節。

In general, RSTP allows the switches to almost immediately react to a topology change and start negotiating port states.

總體來說，RSTP 讓交換器幾乎能立即對拓撲變更做出反應並開始協商連接埠狀態。

In addition, RSTP provides alternate ports that immediately move to forwarding state if a port becomes unavailable.

此外，RSTP 提供備用連接埠，若連接埠變得無法使用，這些備用連接埠會立即移至轉送狀態。