Speculations About The Future

Speculations (猜测)

As noted above, while there are few technical barriers to the compatibility layer, it would make cloud computing resemble a commodity. As such, we expect incumbent cloud providers would strongly resist the emergence of a compatibility layer. However, unlike some other arenas (e.g., packet formats in networking) where standardization requires universal agreement, the software for a compatibility layer can be ported to any cloud, even if the cloud does not wish to officially support it. Thus, users of a large incumbent cloud provider which does not officially support the compatibility layer might opt to use the compatibility interface rather than the cloud provider’s lower-level interfaces just to preserve portability.

如上所述，尽管在兼容层上存在的技术障碍较少，但它会使云计算变得更像一种商品。因此，我们预计现有的云服务提供商会强烈抵制兼容层的出现。然而，与某些其他领域（如网络中的数据包格式）需要普遍协议标准化不同，兼容层的软件可以移植到任何云上，即便云服务提供商不愿意正式支持它。因此，某个不正式支持兼容层的大型云服务提供商的用户，可能会选择使用兼容接口，而不是云服务提供商的低级接口，以保持可移植性。

In addition, while large incumbents might not be happy about a compatibility layer, we expect smaller cloud providers will embrace such a layer. For smaller cloud providers, offering proprietary interfaces may not be preferable to adopting a more widely supported standard. By doing so, these smaller providers would have access to a larger market share (i.e., users who have adopted the compatibility layer as their set of APIs) and could compete on price, performance, or various forms of customer service. We are already seeing this in the market, where Google has recently released Anthos [3], an application management platform based on Kubernetes [12] that supports “write once, run anywhere” with a consistent development and operations experience for cloud and onpremise applications. Anthos is already running on Google Compute Cloud (GCP) and AWS, with Azure to follow shortly. Thus, smaller cloud providers may be better off supporting Anthos rather than competing with it with their own proprietary (and presumably less well-known) interfaces.

此外，尽管大型现有企业可能对兼容层不太满意，但我们预计较小的云服务提供商会欢迎这样的层。对于较小的云服务提供商而言，提供专有接口可能不如采用更广泛支持的标准更具优势。通过这样做，这些较小的提供商可以接触到更大的市场份额（即，采用兼容层作为API集合的用户），并可以在价格、性能或各种客户服务形式上展开竞争。我们已经在市场上看到了这种情况，例如，谷歌最近推出了Anthos [3]，一个基于Kubernetes [12] 的应用管理平台，支持“编写一次，随处运行”，为云端和本地应用程序提供一致的开发和操作体验。Anthos已经在谷歌云计算平台（GCP）和AWS上运行，Azure也即将支持。因此，较小的云服务提供商可能更有利于支持Anthos，而不是通过自己专有的（并且可能不太知名的）接口与之竞争。

Once a compatibility layer has gained traction, an intercloud layer can be developed. The question is: who would benefit from its development? Here we must speculate, since we are far from having these two layers in place.

一旦兼容层获得了足够的支持，便可以开发一个Intercloud层。问题是：谁将从它的开发中受益？在这里我们必须进行推测，因为目前我们距离实现这两个层次还很遥远。

We think that once a compatibility layer and an intercloud layer are in place, cloud providers will fall into two categories. There will be stand-alone cloud providers who try to lock customers in with proprietary interfaces and data export fees. These providers will typically be large enough so that they have the resources to offer a variety of proprietary services. However, in contrast to today, we think there will also be commodity cloud providers who directly support the compatibility layer and agree to reciprocal data peering with other commodity cloud providers. These commodity providers, taken together as a whole, form the Sky, which offers a unified interface to a set of heterogeneous and competing cloud providers.

我们认为，一旦兼容层和Intercloud层就位，云服务提供商将分为两类。一类是独立的云服务提供商，他们通过专有接口和数据导出费用来锁定客户。这些提供商通常规模较大，拥有资源来提供各种专有服务。然而，与今天不同的是，我们认为还会出现商品化云服务提供商，他们直接支持兼容层，并同意与其他商品化云提供商进行互惠数据对等。这些商品化云提供商作为一个整体将组成“Sky”，为一组异构且相互竞争的云提供商提供统一的接口。

Why do we believe the Sky will happen? It rests on the nature of innovation in the two classes of providers. In a competitive market, the stand-alone providers compete with each other, and with the Sky. The commodity providers also compete with each other within the Sky, and collectively compete with the stand-alone providers. In terms of tradeoffs, the stand-alone providers have higher margins (because their customers have exit barriers) but must innovate across the board to retain advantages for their proprietary interfaces.

我们相信“Sky”会出现的原因在于这两类提供商的创新性质。在竞争市场中，独立的云服务提供商相互竞争，同时也与“Sky”竞争。而商品化云提供商则在“Sky”内部相互竞争，并集体与独立提供商竞争。从权衡的角度来看，独立提供商的利润率较高（因为它们的客户有退出壁垒），但它们必须在各个方面进行创新，以保持其专有接口的优势。

In contrast, the commodity providers have lower margins, but can innovate more narrowly. That is, a commodity provider might specialize in supporting one or more services; jobs in the Sky that could benefit from these specialized services would migrate there. For example, Oracle could provide a database-optimized cloud, while a company like EMC can provide a storage-optimized cloud. In addition, hardware manufacturers could directly participate in the cloud economy. For example, Samsung might be able to provide the best price-performance cloud storage, while Nvidia can provide hardware-assisted ML services. More excitingly, a company like Cerebras Systems, which builds a wafer-scale accelerator for AI workloads, can offer a service based on its chips. To do so it just needs to host its machines in one or more colocation datacenters like Equinix and port popular ML frameworks like TensorFlow [24], PyTorch [37], MXNet [27] — thus providing a compatibility layer — onto Cerebras-powered servers. Cerebras only needs to provide processing service; all the other services required by customers (such as data storage) can run in existing cloud providers. In contrast, today, a company like Cerebras has only two choices: get one of the big cloud providers like AWS, Azure, or GCP to deploy its hardware, or build its own fully-featured cloud. Both are daunting propositions.

相比之下，商品化云提供商的利润率较低，但可以更为专注地进行创新。也就是说，一个商品化提供商可能专注于支持一个或多个特定服务；在“Sky”中受益于这些专业服务的作业会迁移到这里。例如，Oracle可以提供一个数据库优化的云，而像EMC这样的公司则可以提供一个存储优化的云。此外，硬件制造商也可以直接参与云经济。例如，三星可能提供最佳性价比的云存储，而Nvidia则可以提供硬件加速的机器学习服务。更令人兴奋的是，像Cerebras Systems这样的公司，专门为AI工作负载构建晶圆级加速器，可以基于其芯片提供服务。为此，它只需要在一个或多个像Equinix这样的联合数据中心中托管其机器，并将流行的机器学习框架（如TensorFlow [24]、PyTorch [37]、MXNet [27]）移植到Cerebras驱动的服务器上，从而提供兼容层。Cerebras只需提供处理服务；客户所需的所有其他服务（如数据存储）可以在现有的云服务提供商中运行。

相比之下，像Cerebras这样的公司目前却只有两种选择：让AWS、Azure或GCP等大型云服务提供商部署其硬件，或者构建自己的全功能云。这两种选择都令人望而生畏。

However, such a dynamic will only be effective if the intercloud layer can find these accelerated services, since most individual users won’t be up-to-date on the relative performance of various services on the Sky. Thus, to make Sky computing work, we need all three layers to be effective: the compatibility layer to hide any differences in implementation between clouds; the intercloud layer to automatically find the best price/performance/properties for various services; and reciprocal peering to make data movements free and fast.

然而，这种动态只有在Intercloud层能够找到这些加速服务的情况下才会有效，因为大多数个体用户并不会了解Sky中各种服务的相对性能。因此，为了使Sky计算有效，我们需要这三个层次都发挥作用：兼容层以隐藏云之间实现的差异；Intercloud层自动找到各种服务的最佳价格/性能/特性；以及互惠对等以实现数据传输的免费和快速。

How they work

• Compatibility Layer: hide differences between clouds
• Intercloud Layer: find the best choice (service) for a specific task
• Reciprocal Peering: make data-migration free and swift

To be clear, we are not predicting the demise of proprietary clouds; we think that in the long-term we will continue to have both kinds of providers. The stand-alone providers will cater to those customers who need more assistance and where price and performance are not the critical factors. This might mean servicing the needs of smaller users who have less expertise in managing the cloud. However, we think the bulk of computation-intense workloads, particularly for sophisticated users, will migrate to the Sky, because of the greater access to innovation.

需要明确的是，我们并不是在预测专有云的消亡；我们认为，从长远来看，我们将继续拥有这两种类型的提供商。独立提供商将满足那些需要更多帮助的客户需求，而在这些客户中，价格和性能并不是关键因素。这可能意味着服务于那些在管理云方面经验较少的小型用户。然而，我们认为大部分计算密集型工作负载，特别是对于复杂用户，将会迁移到Sky，因为那里能够更好地接触到创新。

One might doubt that the commodity providers would provide more innovation than the stand-alone providers, but the PC market offers such an example. In 1981, IBM released its Personal Computer which pushed the computing industry in overdrive. The IBM PC was based on an open architecture. The only proprietary component was the BIOS (Basic Input/Output System) firmware. However, one year later, the BIOS was reverse engineered and this opened the floodgates for PC clones. As expected, this led to much lower prices; however, it also led to more innovation. The clone manufacturers like Compaq, not IBM, were responsible for creating the first portable PC, and for leading the adoption of the new Intel processors (e.g., Intel’s x386). Furthermore, this open architecture led to an industry of add-ons such as hardware accelerators and storage devices. By 1986, IBM PC compatible computers reached half of the entire PC market with the PC clones outselling IBM PCs by significant numbers.

有人可能会怀疑商品化提供商是否能提供比独立提供商更多的创新，但个人计算机市场提供了一个这样的例子。1981年，IBM推出了个人计算机，这推动了计算行业的快速发展。IBM PC基于开放架构，唯一的专有组件是BIOS（基本输入/输出系统）固件。然而，一年后，BIOS被逆向工程，这为PC克隆机打开了大门。正如预期的那样，这导致了价格的大幅下降；然而，它也带来了更多的创新。克隆制造商如康柏（Compaq）而不是IBM，负责创造第一台便携式个人计算机，并引领了新英特尔处理器（如英特尔的x386）的采用。此外，这种开放架构催生了硬件加速器和存储设备等附加产品的行业。到1986年，IBM PC兼容计算机在整个PC市场的份额达到了50%，而PC克隆的销量大幅超过了IBM PC。

Our line of reasoning is, of course, purely speculative. However, the lack of adoption of a compatibility layer in even the smaller cloud providers is because they do not yet see such a layer increasing their revenues (because they are each competing individually against the larger clouds). However, if such providers simultaneously agree on reciprocal peering, then collectively the Sky becomes a viable competitive counterbalance to the large proprietary clouds, and allows the commodity clouds to focus their innovation efforts more narrowly.

我们推理的过程当然是纯粹的推测。然而，甚至在较小的云服务提供商中，缺乏兼容层的采用，是因为他们尚未看到这样的层能增加他们的收入（因为他们各自单独与大型云服务提供商竞争）。然而，如果这些提供商同时同意进行互惠对等，那么Sky就可以成为对大型专有云的有效竞争平衡，并允许商品化云更专注于其创新工作。