Infrastructure as a Service (IaaS) is a delivery model of cloud computing, which provides the ability to users to acquire and release resources according to their demand and pay according to their usage. Resources are provisioned from the cloud as Virtual Machines (VMs), many of which can be deployed on a single computing node, realizing a multi-tenancy model. While virtualization and multitenancy are two sources of workload-execution overhead that have been studied in the past, we still need a thorough, empirical investigation of the joint impact of these overheads, on workload execution. Additionally, commercial and private IaaS providers offer mechanisms that facilitate the lease and use of single infrastructure resources, but to execute multi-job workloads IaaS users still need to select adequate provisioning and allocation policies to instantiate resources and map computational jobs to them. Even though some studies on the policies employed in cloud environments already exist, current and potential IaaS users need deeper insight on the achieved performance and incurred cost of the used policies, derived through empirical investigation.

Storage for cluster applications is typically provisioned based on rough, qualitative characterizations of applications. Moreover, configurations are often selected based on rules of thumb and are usually homogeneous across a deployment; to handle increased load, the application is simply scaled out across additional machines and storage of the same type. As deployments grow larger and storage options (e.g., disks, SSDs, DRAM) diversify, however, current practices are becoming increasingly inefficient in trading off cost versus performance. To enable more cost-effective deployment of cluster applications, we develop scc—a storage configuration compiler for cluster applications. scc automates cluster configuration decisions based on formal specifications of application behavior and hardware properties. We study a range of storage configurations and identify specifications that succinctly capture the trade-offs offered by different types of hardware, as well as the varying demands of application components. We apply scc to three representative applications and find that scc is expressive enough to meet application Service Level Agreements (SLAs) while delivering 2–4.5 × savings in cost on average compared to simple scale-out options. scc’s advantage stems mainly from its ability to configure heterogeneous—rather than conventional, homogeneous—cluster architectures to optimize cost. 1

Memory is rapidly becoming a precious resource in many data processing environments. This paper introduces a new data structure called a Compressed Buffer Tree (CBT). Using a combination of buffering, compression, and lazy aggregation, CBTs can improve the memory efficiency of the GroupBy-Aggregate abstraction which forms the basis of many data processing models like MapReduce and databases. We evaluate CBTs in the context of MapReduce aggregation, and show that CBTs can provide significant advantages over existing hashbased aggregation techniques: up to 2 × less memory and 1.5 × the throughput, at the cost of 2.5 × CPU. 1

Abstract—Coprocessors such as GPUs are increasingly being deployed in clusters to process scientific and compute-intensive jobs. In this work, we study if GPU-based heterogeneous clusters can benefit client-server applications. Specifically, we consider the practical situation where multiple client-server applications share a heterogeneous cluster (multi-tenancy), and experience unpredictable variations in incoming client request rates, including steep load spikes. Even for “compute-intensive ” client-server applications, it is unclear if a GPU-based cluster can seamlessly deliver acceptable response times in the presence of multi-tenancy and load spikes. We argue that a cluster-level scheduler that is aware of application load, request deadlines and the heterogeneity is necessary in this situation. We propose a novel scheduler called Symphony that enables efficient, dynamic sharing of a GPU-based heterogeneous cluster across multiple concurrentlyexecuting client-server applications, each with arbitrary load spikes. Symphony performs three key tasks: it (i) monitors the load on each application, (ii) collects past performance data and dynamically builds simple performance models of available processing resources and (iii) computes a priority for pending requests based on the above parameters and the requests ’ slack. Based on this, it reorders client requests across different applications to achieve acceptable response times. We also define how client-server applications should interact with a scheduler such as Symphony, and develop an API to this end. We deploy Symphony as user-space middleware on a high-end heterogeneous cluster with dual quad-core Xeon CPUs and dual NVIDIA Fermi GPUs. An evaluation using representative applications shows that in the presenceofload spikes(i)Symphonyincurs2−20×fewer requests that do not meet response time constraints compared with other schedulers, and (ii) in order to achieve the same performance as Symphony, other schedulers need 2 × more cluster nodes. I.

Memory is rapidly becoming a precious resource in many data processing environments. This paper introduces a new data structure called a Compressed Buffer Tree (CBT). Using a combination of buffering, compression, and lazy aggregation, CBTs can improve the memory efficiency of the GroupBy-Aggregate abstraction which forms the basis of many data processing models like MapReduce and databases. We evaluate CBTs in the context of MapReduce aggregation, and show that CBTs can provide significant advantages over existing hashbased aggregation techniques: up to 2 × less memory and 1.5 × the throughput, at the cost of 2.5 × CPU. 1

Shared storage services enjoy wide adoption in commercial clouds. But most systems today provide weak performance isolation and fairness between tenants, if at all. Misbehaving or high-demand tenants can overload the shared service and disrupt other well-behaved tenants, leading to unpredictable performance and violating SLAs. This paper presents Pisces, a system for achieving datacenter-wide per-tenant performance isolation and fairness in shared key-value storage. Today’s approaches for multi-tenant resource allocation are based either on per-VM allocations or hard rate limits that assume uniform workloads to achieve high utilization. Pisces achieves per-tenant weighted fair shares (or minimal rates) of the aggregate resources of the shared service, even when different tenants ’ partitions are co-located and when demand for different partitions is skewed, time-varying, or bottlenecked by different server resources. Pisces does so by decomposing the fair sharing problem into a combination of four complementary mechanisms—partition placement, weight allocation, replica selection, and weighted fair queuing—that operate on different time-scales and combine to provide system-wide max-min fairness. An evaluation of our Pisces storage prototype achieves nearly ideal (0.99 Min-Max Ratio) weighted fair sharing, strong performance isolation, and robustness to skew and shifts in tenant demand. These properties are achieved with minimal overhead (<3%), even when running at high utilization (more than 400,000 requests/second/server for 10B requests). 1.

An increasing number and variety of enterprises are moving workloads to cloud platforms. Whether serving external customers or internal business units, cloud platforms typically allow multiple users, or tenants, to share the same physical server and network infrastructure, as well as use common platform services. Examples of these shared, multi-tenant services include key-value stores, block storage volumes, message queues, and notification services. These leverage the expertise of the cloud provider in building, managing, and improving common services, and enable the statistical multiplexing of resources between tenants for higher utilization. Because they rely on shared infrastructure, however, these services face two key, related issues: • Multi-tenant interference and unfairness: Tenants

not be interpreted as representing the official policies, either expressed or implied, of any sponsoring institution, the U.S. government or any other entity. Further, the authors would like to thank Alexey Tumanov, Samantha Gottlieb, and the members of Google’s cluster management team for their technical feedback. The authors are also thankful to Google for providing the traces used for the experimental evaluation presented in this paper. Last but not least, we also thank the members and companies of