HotCloud '17 Workshop Program

Over the past decade, we have witnessed exponential growth in the density (petabyte-level) and breadth (across geo-distributed datacenters) of data distribution. It becomes increasingly challenging but imperative to minimize the response times of data analytic queries over multiple geo-distributed datacenters. However, existing scheduling-based solutions have largely been motivated by pre-established mantras (e.g., bandwidth scarcity). Without data-driven insights into performance bottlenecks at runtime, schedulers might blindly assign tasks to workers that are suffering from unidentified bottlenecks.

In this paper, we present Lube, a system framework that minimizes query response times by detecting and mitigating bottlenecks at runtime. Lube monitors geo-distributed data analytic queries in real-time, detects potential bottlenecks, and mitigates them with a bottleneck-aware scheduling policy. Our preliminary experiments on a real-world prototype across Amazon EC2 regions have shown that Lube can detect bottlenecks with over 90% accuracy, and reduce the median query response time by up to 33% compared to Spark’s built-in locality-based scheduler.

Applications have commonly been oblivious to their cloud runtimes. This is primarily because they started their journey in IaaS clouds, running on a guestOS inside VMs. Then to increase performance, many guestOSes have been paravirtualized making them virtualization aware, so that they can bypass some of the virtualization layers, as in virtio. This approach still kept applications unmodified. Recently, we are witnessing a rapid adoption of containers due to their packaging benefits, high density, fast start-up and low overhead. Applications are increasingly being on-boarded to PaaS clouds in the form of application containers or appc, where they are run directly on a cloud substrate like Kubernetes or Docker Swarm. This shift in deployment practices present an opportunity to make applications aware of their cloud. In this paper, we present Paracloud framework for application containers and discuss the Paracloud interface (PaCI) for three cloud operations namely migration, auto-scaling and load-balancing.

Large scale in-memory key-value stores like RAMCloud can perform millions of operations per second per server with a few microseconds of access latency. However, these systems often only provide simple feature sets, and the lack of extensibility is an obstacle for building higher-level services. We evaluate the possibility of using JavaScript for shipping computation to data and for extending database functionality by comparing against other possible approaches. Microbenchmarks are promising; the V8 JavaScript runtime provides near native performance with reduced isolation costs when compared with native code and hardware-based protections. We conclude with initial thoughts on how this technology can be deployed for fast procedures that operate on in-memory data, that maximize gains from JIT, and that exploit the kernel-bypass DMA capabilities of modern network cards.

The performance of mobile phone processors has been steadily increasing, causing the performance gap between server and mobile processors to narrow with mobile processors sporting superior performance per unit energy. Fueled by the slowing of Moore’s Law, the overall performance of single-chip mobile and server processors have likewise plateaued. These trends and the glut of used and partially broken smartphones which become environmental e-waste motivate creating cloud servers out of decommissioned mobile phones. This work proposes creating a compute dense server built out of used and partially broken smartphones (e.g. screen can be broken). This work evaluates the total cost of ownership (TCO) benefit of using servers based on decommissioned mobile devices and analyzes some of the architectural design trade-offs in creating such servers.

Heterogeneity is a growing concern for scheduling on the cloud. Hardware is increasingly heterogeneous (e.g., GPUs, FPGAs, diverse I/O devices), emphasizing the need to build schedulers that identify the internal structure of applications and utilize available hardware resources to their full potential. In this paper we present our initial efforts to build a scheduler that tackles heterogeneity (in hardware and in software) as a primary concern. Our scheduler, HCl (Heterogeneous Cluster), models applications as annotated directed acyclic graphs (DAGs), where each node represents a task. It maps tasks onto hardware nodes, also organized in DAGs. Initial results using application models based on TPC-DS queries running on Apache Spark show that HCl can improve significantly upon approaches that do not consider heterogeneity and generate schedules that approach the critical path in length.

Emerging Cyber-Physical Systems (CPS) such as connected vehicles and smart cities span large geographical areas. These systems are increasingly distributed and interconnected. Hence, a hierarchy of cloudlet and cloud deployments will be key to enable scaling, while simultaneously hosting the intelligence behind these systems. Given that CPS applications are often safety-critical, existing techniques focus on reducing latency to provide real-time performance. While low latency is useful, a shared and precise notion of time is key to enabling coordinated action in distributed CPS. In this position paper, we argue for a global Quality of Time (QoT)-based architecture, centered around a shared virtualized notion of time, based on the timeline abstraction. Our architecture allows applications to specify their QoT requirements, while exposing timing uncertainty to the application. The timeline abstraction with the associated knowledge of QoT enables scalable geo-distributed coordination in CPS, while providing avenues for fault tolerance and graceful degradation in the face of adversity.

Infrastructure-as-a-Service clouds are rapidly evolving into market-like environments that offer a wide range of server contracts. Amazon EC2’s spot market is the clearest example of this trend: it operates over 5000 markets globally where users can rent servers for a variable price. To exploit spot instances, while mitigating the risk of price spikes and revocations, many researchers and startups have developed techniques for modeling and predicting prices to optimize spot server selection. However, prior approaches focus largely on predicting individual server prices, which is akin to predicting the price of a single stock. We argue that researchers should instead focus on “index-based” modeling and prediction that aggregates prices from many markets in each region and availability zone. We show that, for applications flexible enough to select and “trade” servers globally, making decisions based on broader indices lowers costs and improves availability compared to index-agnostic policies.

Latency-sensitive applications like virtualized telecom and industrial IoT systems require a service for ultrafast state externalization to become cloud-native. In this paper we propose a distributed shared memory system, called DAL, which achieves the lowest possible latency by transparently co-locating individual data items with applications working on them. Upon changes in data access patterns, the system automatically adapts data locations to keep the number of remote operations at a minimum. By avoiding the costs of network transport and using shared memory communication, the system can achieve 1 μs data access latency. We envision DAL as a platform component which enables latency-sensitive applications to take advantage of the cloud.