Workshop Program

We will present a survey of resource allocation and networking problems that arise in cloud computing clusters and data centers. Examples of problems that will be presented include cloud infrastructure provisioning, stream processing, green computing, data center networking, and data locality-based load balancing. The common theme in solving these problems is the use of queue length feedback to make optimal decisions, without the knowledge of traffic statistics. An optimal resource allocation scheme will be presented for one of the problems, the so-called Infrastructure-as-a-Service or IaaS problem, by exploiting a relationship between convex optimization and stochastic Lyapunov techniques.

R. Srikant is the Fredric G. and Elizabeth H. Nearing Endowed Professor of Electrical and Computer Engineering and a Professor in the Coordinated Science Lab, both at the University of Illinois at Urbana-Champaign. He is the author or coauthor of two monographs, The Mathematics of Internet Congestion Control and Network Optimization and Control, and a coauthor of the book Communication Networks: An Optimization, Control and Stochastic Networks Perspective. His research interests include communication networks, queueing theory, machine learning, and optimization. He is currently the Editor-in-Chief of the IEEE/ACM Transactions on Networking and is a Fellow of the IEEE.

Leadership class systems are heavily shared resource environments with users contending for shared system resources. This results in users experiencing huge performance variations, and also affects the overall throughput of the system. To alleviate the problem, system software tools must be built taking into consideration user requirements and resource availability, a feedback driven approach. Realizing a feedback-based compute environment for peta-scale systems have two challenging tasks. First, collecting discreet, coarse-grained system statistics from multiple systems using minimum system resources and without affecting the user jobs is a hard problem. Second, with discreet data collected from disparate sources the challenge is in associating the data for meaningful interpretations to drive feedback-based decision systems in real-time. In this paper, we elaborate on a feedback-based computing framework with respect to the peta-scale compute and storage system at the Oak Ridge Leadership Computing Facility. We describe our feedback-based approach for dynamic resource allocation, context-aware scheduling and application checkpointing.

Web services, large and small, use in-memory caches like memcached to lower database loads and quickly respond to user requests. These cache clusters are typically provisioned to support peak load, both in terms of request processing capabilities and cache storage size. This kind of worst-case provisioning can be very expensive (e.g., Facebook reportedly uses more than 10,000 servers for its cache cluster) and does not take advantage of the dynamic resource allocation and virtual machine provisioning capabilities found in modern public and private clouds. Further, there can be great diversity in both the workloads running on a cache cluster and the types of nodes that compose the cluster, making manual management difficult. This paper identifies the challenges in designing large-scale self-managing caches. Rather than requiring all cache clients to know the key to server mapping, we propose an automated load balancer that can perform line-rate request redirection in a far more dynamic manner. We describe how stream analytic techniques can be used to efficiently detect key hotspots. A controller then guides the load balancer’s key mapping and replication level to prevent overload, and automatically starts additional servers when needed.

The next decade of computing workloads is expected to be dominated by soft-real time applications such as multimedia and machine vision. Such workloads are characterized by transient spikes requiring over provisioning of compute servers, adversely affecting the cost, energy usage, and environmental impact of data centers. In many of these applications, although deadlines need to be met to provide QoS guarantees, other quality parameters of the application (for example, visual quality in video processing) can be tuned in conjunction with hardware parameters (for example, DVFS) to give acceptable performance under overload conditions. In this paper, we experimentally demonstrate a predictive control approach for improving overload capacity and energy efficiency by incorporating control variables from both the hardware and the application layer. Further, we illustrate the impact of the choice of multiprocessor real-time scheduling algorithms on the performance of the controller for heterogeneous workloads.

Virtual desktop infrastructure (VDI) deployments are a rapidly growing segment in the Mobile/Cloud Era. Compared to traditional enterprise desktop deployments, VDI can reduce the total cost of ownership by as much as 50%. However, the cost of powering a VDI deployment is still a significant IT expense. Typically these deployments consist of a complex system of interconnected server, storage and networking components. Thus, it is challenging to minimize energy consumption of the entire system while at the same time satisfying performance requirements. In this work, we first derive an accurate VDI performance model and then propose a hierarchical heuristic to minimize energy consumption without violating performance constraints. We also demonstrate that such an approach reduces algorithmic complexity significantly. Results from hardware experiments show that in scenarios with low consolidation ratios energy savings range from 3% to 6%, while for high consolidation ratios they range from 11% to 25%. In all cases, the measured end-to-end performance penalty is minimal.

Supervisory Control and Data Acquisition (SCADA) systems, which are widely used in monitoring and controlling critical infrastructure sectors, are highly vulnerable to cyber attacks. Current security solutions can protect SCADA systems from known cyber assaults, but most solutions require human intervention. This paper applies autonomic computing technology to monitor SCADA system performance, and proactively estimate upcoming attacks for a given system model of a physical infrastructure. We also present the feasibility of intrusion detection systems for known and unknown attack detection. A dynamic intrusion response system is designed to evaluate recommended responses, and appropriate responses are executed to influence attack impacts. We used a case study of a water storage tank to develop an attack that modifies Modbus messages transmitted between slaves and masters. Experimental results show that, with little or no human intervention, the proposed approach enhances the security of the SCADA system, reduces protection time delays, and maintains water storage tank performance.

Cloud applications depend on third party services for features ranging from networked storage to maps. Web-based application programming interfaces (web APIs) make it easy to use these third party services but hide details about their structure and resource needs. However, due to the lack of implementation-level knowledge, cloud applications have little information when these third party services break or even unproperly implemented. This paper outlines research to extract workload details from data collected by probing web APIs. The resulting workload profiles will provide early warning signs when web APIs have broken component. Such information could be used to build feedback loops to deal with possible high response times of web APIs. It will also help developers choose between competing web APIs. The challenge is to extract profiles by assuming that the systems underlying web APIs use common cloud computing practices, e.g., auto scaling. In early results, we have used blind source separation to extract per-tier delays in multi-tier storage services using response times collected from API probes. We modeled median and 95th percentile delay within 10% error at each tier. Finally, we set up two competing storage services, one of which used a slow key-value store. We probed their APIs and used our profiles to choose between the two. We showed that looking at response times alone could lead to the wrong choice and that detailed workload profiles provided helpful data.