On the FaaS Track

Abstract

Serverless computing is an emerging paradigm that greatly simplifies the usage of cloud resources and suits well to many tasks. Most notably, Function-as-a-Service (FaaS) enables programmers to develop cloud applications as individual functions that can run and scale independently. Yet, due to the disaggregation of storage and compute resources in FaaS, applications that require fine-grained support for mutable state and synchronization, such as machine learning and scientific computing, are hard to build.In this work, we present Crucial, a system to program highly-concurrent stateful applications with serverless architectures. Its programming model keeps the simplicity of FaaS and allows to port effortlessly multi-threaded algorithms to this new environment. Crucial is built upon the key insight that FaaS resembles to concurrent programming at the scale of a data center. As a consequence, a distributed shared memory layer is the right answer to the need for fine-grained state management and coordination in serverless. We validate our system with the help of micro-benchmarks and various applications. In particular, we implement two common machine learning algorithms: 𝑘-means clustering and logistic regression. For both cases, Crucial obtains superior or comparable performance to an equivalent Spark cluster.

Characterizing serverless platforms with serverlessbench

“…The explicit states are the data explicitly used and managed by serverless functions, while the implicit states are the data used by the language runtime (e.g., cached data and codes, JIT profiles). As many state-of-the-art cloud applications are stateful, many recent works discuss and propose stateful serverless designs [19,24,49,50]. The design of a serverless state management system must highlight two key factors: 1) states being durable beyond function life cycles; 2) state passing being efficient enough.…”

“…The design of a serverless state management system must highlight two key factors: 1) states being durable beyond function life cycles; 2) state passing being efficient enough. Most researches adopt a two-layer design with a high-efficiency local data passing mechanism (using local cache [24,50], local key-value store [19], or shared memory [49]) along with a scalable and durable global object storage.…”

ReactiveFnJ: A choreographed model for Fork-Join Workflow in Serverless Computing

“…These coordination services automatically trigger the execution of each function in the workflow and synchronize their behaviors and states. FaaS orchestration services such as AWS Step Functions or IBM Composer offer limited capabilities to coordinate serverless functions [20]. For instance, even in the AWS Step function, there is no provision for multiple functions to synchronize in parallel when the number of parallel instances is dynamic.…”

Floki

“…Works like Crucial [2], Cloudburst [19], OFC [14], SONIC [12], and SAND [1] focus on inter-function data sharing, disaggregation and physical co-location, or data locality exploitation. However, they need static provisioning of per-node cache resources or focus on small data sizes.…”

“…The slow data transfers between functions make the CloudSort benchmark to be up to 500× slower when executed on AWS Lambda with S3 instead of on a cluster of Virtual Machines (VMs). Recent studies tackle this problem by implementing optimized exchange operators [13,15], using multi-tier storage combining slow with fast storage or solely remote in-memory storage [7,8,16], exploiting per-node caches [2,19], co-locating functions on a single container [1,5,9,18], handling external storage on long-running VMs [3,22], or circumventing the network constraints [21]. However, these methods either use domain-specific optimizations, require two copies of data over the network, are not fully transparent to the user, break the advantage of fine-grained scaling, or use non-serverless components.…”