Virtual execution environments allow for consolidation of multiple applications onto the same physical server, thereby enabling more efficient use of server resources. However, users often statically configure the resources of virtual machines through guesswork, resulting in either insufficient resource allocations that hinder VM performance, or excessive allocations that waste precious data center resources. In this paper, we first characterize real-world resource allocation and utilization of VMs through the analysis of an extensive dataset, consisting of more than 250k VMs from over 3.6k private enterprise clusters. Our large-scale analysis confirms that VMs are often misconfigured, either overprovisioned or underprovisioned, and that this problem is pervasive across a wide range of private clusters. We then propose ADARES, an adaptive system that dynamically adjusts VM resources using machine learning techniques. In particular, ADARES leverages the contextual bandits framework to effectively manage the adaptations. Our system exploits easily collectible data, at the cluster, node, and VM levels, to make more sensible allocation decisions, and uses transfer learning to safely explore the configurations space and speed up training. Our empirical evaluation shows that ADARES can significantly improve system utilization without sacrificing performance. For instance, when compared to threshold and prediction-based baselines, it achieves more predictable VM-level performance and also reduces the amount of virtual CPUs and memory provisioned by up to 35% and 60% respectively for synthetic workloads on real clusters.

Modern cluster storage systems perform a variety of background tasks to improve the performance, availability, durability, and cost-efficiency of stored data. For example, cleaners compact fragmented data to generate long sequential runs, tiering services automatically migrate data between solid-state and hard disk drives based on usage, recovery mechanisms replicate data to improve availability and durability in the face of failures, cost saving techniques perform data transformations to reduce the storage costs, and so on. In this work, we present Curator, a background MapReduce-style execution framework for cluster management tasks, in the context of a distributed storage system used in enterprise clusters. We describe Curator’s design and implementation, and evaluate its performance using a handful of relevant metrics. We further report experiences and lessons learned from its five-year construction period, as well as thousands of customer deployments. Finally, we propose a machine learning-based model to identify an efficient execution policy for Curator’s management tasks that can adapt to varying workload characteristics.

There is an increasing trend in the use of on-premise clusters within companies. Security, regulatory constraints, and enhanced service quality push organizations to work in these so called private cloud environments. On the other hand, the deployment of private enterprise clusters requires careful consideration of what will be necessary or may happen in the future, both in terms of compute demands and failures, as they lack the public cloud's flexibility to immediately provision new nodes in case of demand spikes or node failures. In order to better understand the challenges and tradeoffs of operating in private settings, we perform, to the best of our knowledge, the first extensive characterization of on-premise clusters. Specifically, we analyze data ranging from hardware failures to typical compute/storage requirements and workload profiles, from a large number of Nutanix clusters deployed at various companies. We show that private cloud hardware failure rates are lower, and that load/demand needs are more predictable than in other settings. Finally, we demonstrate the value of the measurements by using them to provide an analytical model for computing durability in private clouds, as well as a machine learning-driven approach for characterizing private clouds' growth.

Latency to end-users and regulatory requirements push large companies to build data centers all around the world. The resulting data is "born" geographically distributed. On the other hand, many machine learning applications require a global view of such data in order to achieve the best results. These types of applications form a new class of learning problems, which we call Geo-Distributed Machine Learning (GDML). Such applications need to cope with: 1) scarce and expensive cross-data center bandwidth, and 2) growing privacy concerns that are pushing for stricter data sovereignty regulations. Current solutions to learning from geo-distributed data sources revolve around the idea of first centralizing the data in one data center, and then training locally. As machine learning algorithms are communication-intensive, the cost of centralizing the data is thought to be offset by the lower cost of intra-data center communication during training. In this work, we show that the current centralized practice can be far from optimal, and propose a system for doing geo-distributed training. Furthermore, we argue that the geo-distributed approach is structurally more amenable to dealing with regulatory constraints, as raw data never leaves the source data center. Our empirical evaluation on three real datasets confirms the general validity of our approach, and shows that GDML is not only possible but also advisable in many scenarios.

Allreduce is an abstraction commonly used for solving machine learning problems. It is an operation where every node starts with a local value and ends up with an aggregate global result. MPI provides an Allreduce implementation. Though it has been widely adopted, it is somewhat limited; it lacks fault tolerance and cannot run easily on existing systems. In this work, we propose RABIT1, an Allreduce library suitable for distributed machine learning algorithms that overcomes the aforementioned drawbacks; it is faulttolerant and can easily run on top of existing systems. We compare RABIT with existing solutions and show that it performs competitively.

: Identifying documents that contain timely and vital information for an entity of interest, a task known as vital filtering, has become increasingly important with the availability of large document collections. To efficiently filter such large text corpora in a streaming manner, we need to compactly represent previously observed entity contexts and quickly estimate whether a new document contains novel information. Existing approaches to modeling contexts, such as bag of words, latent semantic indexing, and topic models are limited in several respects: they are unable to handle streaming data, do not model the underlying topic of each document, suffer from lexical sparsity, and/or do not accurately estimate temporal vitalness. In this paper, we introduce a word embedding-based non-parametric representation of entities that addresses the above limitations. The word embeddings provide accurate and compact summaries of observed entity contexts further described by topic clusters that are estimated in a non-parametric manner. Additionally we associate a staleness measure with each entity and topic cluster, dynamically estimating their temporal relevance. This approach of using word embeddings, non-parametric clustering, and staleness provides an efficient yet appropriate representation of entity contexts for the streaming setting, enabling accurate vital filtering.

Online learning algorithms operate on a single instance at a time, allowing for updates that are fast, simple and perform well in a wide range of practical settings [1]. In this paper, we focus on online learning algorithms under delayed feedback, where the true labels arrive minutes, hours or even days later. Current solutions to this problem assume that all the instances are kept until the corresponding labels arrive [2]. In this project, we aim to examine this problem in a scenario where limited resources are available, so that it becomes infeasible to keep all of the instances in memory until the correct labels arrive. Herein we analyze different strategies for keeping only the data that matters: either by dropping certain training instances, features or both. We borrow techniques from the Active Learning literature, modified to fit this scenario. We also propose new techniques and baselines.