Almost exactly ten years to the day from publishing of their widely-read, seminal paper on cloud computing, UC Berkeley researchers have issued another ambitious examination of cloud computing – Cloud Programming Simplified: A Berkeley View on Serverless Computing. The new work heralds the rise of ‘serverless computing’ as the next dominant phase of cloud computing.

You may recall the first paper (Above the Clouds: A Berkeley View of Cloud Computing) provided a detailed portrait of cloud computing’s past, present, and future as envisioned by the researchers. Published on February 9, 2009, it has garnered more than 17,000 citations including more than 1,000 in the past year. Not bad.

The new paper[i], dated Feb 10, 2019, takes a few modest bows for past prescience and then plunges into serverless computing, what it is, what the challenges are, and, of course, offers concluding predictions (presented below) without which any such report would be remiss. The paper was issued by authors from the Electrical Engineering and Computer Sciences, UC Berkeley and includes two notable repeats from the original paper: David Patterson (now an emeritus professor) and Ion Stoica.

Here’s a portion of the abstract:

“Serverless cloud computing handles virtually all the system administration operations needed to make it easier for programmers to use the cloud. It provides an interface that greatly simplifies cloud programming, and represents an evolution that parallels the transition from assembly language to high-level programming languages. This paper gives a quick history of cloud computing, including an accounting of the predictions of the 2009 Berkeley View of Cloud Computing paper, explains the motivation for serverless computing, describes applications that stretch the current limits of serverless, and then lists obstacles and research opportunities required for serverless computing to fulfill its full potential.”

Without doubt, the notion of serverless computing is on the rise. In theory, a user “just writes a cloud function in a high-level language, picks the event that should trigger the running of the function—such as loading an image into cloud storage or adding an image thumbnail to a database table—and lets the serverless system handle everything else: instance selection, scaling, deployment, fault tolerance, monitoring, logging, security patches, and so on,” they write.

The researchers bullet out three key distinctions between serverless and serverful clouds:

• Decoupled computation and storage. The storage and computation scale separately and are provisioned and priced independently. In general, the storage is provided by a separate cloud service and the computation is stateless.
• Executing code without managing resource allocation. Instead of requesting resources, the user provides a piece of code and the cloud automatically provisions resources to execute that code.
• Paying in proportion to resources used instead of for resources allocated. Billing is by some dimension associated with the execution, such as execution time, rather than by a dimension of the base cloud platform, such as size and number of VMs allocated.

A few observers say the idea is old hat. The authors counter:

“Some have argued that serverless computing is merely a rebranding of preceding offerings, perhaps a modest generalization of Platform as a Service (PaaS) cloud products such as Heroku, Firebase, or Parse. Others might point out that the shared web hosting environments popular in the 1990s provided much of what serverless computing has to offer. For example, these had a stateless programming model allowing high levels of multi-tenancy, elastic response to variable demand, and a standardized function invocation API, the Common Gateway Interface (CGI), which even allowed direct deployment of source code written in high-level languages such as Perl or PHP. Google’s original App Engine, largely rebuffed by the market just a few years before serverless computing gained in popularity, also allowed developers to deploy code while leaving most aspects of operations to the cloud provider. We believe serverless computing represents significant innovation over PaaS and other previous models.

“Today’s serverless computing with cloud functions differs from its predecessors in several essential ways: better autoscaling, strong isolation, platform flexibility, and service ecosystem support. Among these factors, the autoscaling offered by AWS Lambda marked a striking departure from what came before. It tracked load with much greater fidelity than serverful autoscaling techniques, responding quickly to scale up when needed and scaling all the way down to zero resources, and zero cost, in the absence of demand. It charged in a much more fine-grained way, providing a minimum billing increment of 100 ms at a time when other autoscaling services charged by the hour. In a critical departure, it charged the customer for the time their code was actually executing, not for the resources reserved to execute their program. This distinction ensured the cloud provider had “skin in the game” on autoscaling, and consequently provided incentives to ensure efficient resource allocation.”

There’s a lot to unpack here (apologies to the authors for excerpting so much of their work but it seemed the best way to minimize insertion of error). The paper is best read in full and the authors don’t shrink from laying out challenges (abstraction, system, networking, security, and architecture) and take a stab at identifying “fallacies and pitfalls.” They also examine several applications for serverless computing suitability including, for example, these two typical HPC and machine learning apps:

• “Numpywren: Linear algebra. Large scale linear algebra computations are traditionally deployed on supercomputers or high-performance computing clusters connected by high-speed, low-latency networks. Given this history, serverless computing initially seems a poor fit. Yet there are two reasons why serverless computing might still make sense for linear algebra computations. First, managing clusters is a big barrier for many non-CS scientists [27]. Second, the amount of parallelism can vary dramatically during a computation. Provisioning a cluster with a fixed size will either slow down the job or leave the cluster underutilized
• “Cirrus: Machine learning training. Machine Learning (ML) researchers have traditionally used clusters of VMs for different tasks in ML workflows such as preprocessing, model training, and hyperparameter tuning. One challenge with this approach is that different stages of this pipeline can require significantly different amounts of resources. As with linear algebra algorithms, a fixed cluster size will either lead to severe underutilization or severe slowdown. Serverless computing can address this challenge by enabling every stage to scale to meet its resource demands. Further, it frees developers from managing clusters.”

Time will tell how closely this latest portrait of cloud computing matches what actually unfolds.

The authors contend that by providing a simplified programming environment, serverless computing makes the cloud much easier to use, thereby attracting more people who can and will use it.

“Serverless computing comprises FaaS (function as a service) and BaaS (back end as a service) offerings, and marks an important maturation of cloud programming. It obviates the need for manual resource management and optimization that today’s serverful computing imposes on application developers, a maturation akin to the move from assembly language to high-level languages more than four decades ago,” they write.

They offer the following predictions about serverless computing in the next decade (bold added):

• We expect new BaaS storage services to be created that expand the types of applications that run well on serverless computing. Such storage will match the performance of local block storage and come in ephemeral and durable variants. We will see much more heterogeneity of computer hardware for serverless computing than the conventional x86 microprocessor that powers it today.
• We expect serverless computing to become simpler to program securely than serverful computing, benefiting from the high level of programming abstraction and the fine-grained isolation of cloud functions.
• We see no fundamental reason why the cost of serverless computing should be higher than that of serverful computing, so we predict that billing models will evolve so that almost any application, running at almost any scale, will cost no more and perhaps much less with serverless computing.
• The future of serverful computing will be to facilitate BaaS. Applications that prove to be difficult to write on top of serverless computing, such as OLTP databases or communication primitives such as queues, will likely be offered as part of a richer set of services from all cloud providers.
• While serverful cloud computing won’t disappear, the relative importance of that portion of the cloud will decline as serverless computing overcomes its current limitations.

[i]Cloud Programming Simplified: A Berkeley View on Serverless Computing,

Eric Jonas, Johann Schleier-Smith, Vikram Sreekanti, Chia-Che Tsai, Anurag Khandelwal, Qifan Pu, Vaishaal Shankar, Joao Menezes Carreira, Karl Krauth, Neeraja Yadwadkar, Joseph Gonzalez, Raluca Ada Popa, Ion Stoica and David A. Patterson, https://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-3.pdf

Source: Figure & Table from the paper