Extracting structured information from templatic documents is an important problem with the potential to automate many real-world business workflows such as payment, procurement, and payroll. The core challenge is that such documents can be laid out in virtually infinitely different ways. A good solution to this problem is one that generalizes well not only to known templates such as invoices from a known vendor, but also to unseen ones. We developed a system called Glean to tackle this problem. Given a target schema for a document type and some labeled documents of that type, Glean uses machine learning to automatically extract structured information from other documents of that type. In this paper, we describe the overall architecture of Glean, and discuss three key data management challenges : 1) managing the quality of ground truth data, 2) generating training data for the machine learning model using labeled documents, and 3) building tools that help a developer rapidly build and improve a model for a given document type. Through empirical studies on a real-world dataset, we show that these data management techniques allow us to train a model that is over 5 F1 points better than the exact same model architecture without the techniques we describe. We argue that for such information-extraction problems, designing abstractions that carefully manage the training data is at least as important as choosing a good model architecture.

System provisioning, resource allocation, and configuration decisions for I/O-intensive workflow applications are complex even for expert users. Users face choices at multiple levels: allocating resources to individual sub-systems (e.g., the application layer, the storage layer) as well as configuring each of these optimally (e.g., replication level, chunk size, caching policies in case of storage) all having a large impact on the overall application performance. This paper presents a solution to address the problem of supporting these provisioning, allocation and configuration decisions for workflow applications. To enable selecting a good choice in a reasonable time, we propose an approach that accelerates the exploration of the configuration space based on a low-cost performance predictor that estimates total execution time of a workflow application in a given setup. We evaluate the predictor in a number of different scenarios including the Montage application: a workflow composed of over 7,500 tasks structured in 10 different stages with varying characteristics. Our evaluation shows that: (i) the predictor is effective in identifying the desired system configuration, (ii) it can scale to model a complex workflow application run on a 100-node cluster, while (iii) using orders of magnitude less resources than running the actual application. Additionally, we extend the predictor to estimate the energy usage of the system, and we present our experience with incorporating it in the development process of a distributed storage system.

System provisioning, resource allocation, and system configuration decisions for I/O-intensive workflow applications are complex even for expert users. Users face choices at multiple levels: allocating resources to individual sub-systems (e.g., the application layer, the storage layer) and configuring each of these optimally (e.g., replication level, chunk size, caching policies in case of storage) all having a large impact on overall application performance. This paper presents our progress on addressing the problem of supporting these provisioning, allocation and configuration decisions for workflow applications. To enable selecting a good choice in a reasonable time, we propose an approach that accelerates the exploration of the configuration space based on a low-cost performance predictor that estimates total execution time of a workflow application in a given setup. Our evaluation shows that: (i) the predictor is effective in identifying the desired system configuration, (ii) it can scale to model a workflow application run on an entire cluster, while (iii) using over 2000x less resources (machines x time) than running the actual application.

Graph processing has gained renewed attention. The increasing large scale and wealth of connected data, such as those accrued by social network applications, demand the design of new techniques and platforms to efficiently derive actionable information from large scale graphs. Hybrid systems that host processing units optimized for both fast sequential processing and bulk processing (e.g., GPUaccelerated systems) have the potential to cope with the heterogeneous structure of real graphs and enable high performance graph processing. Reaching this point, however, poses multiple challenges. The heterogeneity of the processing elements (e.g., GPUs implement a different parallel processing model than CPUs and have much less memory) and the inherent irregularity of graph workloads require careful graph partitioning and load assignment. In particular, the workload generated by a partitioning scheme should match the strength of the processing element the partition is allocated to. This work explores the feasibility and quantifies the performance gains of such low-cost partitioning schemes. We propose to partition the workload between the two types of processing elements based on vertex connectivity. We show that such partitioning schemes offer a simple, yet efficient way to boost the overall performance of the hybrid system. Our evaluation illustrates that processing a 4-billion edges graph on a system with one CPU socket and one GPU, while offloading as little as 25% of the edges to the GPU, achieves 2x performance improvement over state-of-the-art implementations running on a dual-socket symmetric system. Moreover, for the same graph, a hybrid system with dualsocket and dual-GPU is capable of 1.13 Billion breadth-first search traversed edge per second, a performance rate that is competitive with the latest entries in the Graph500 list, yet at a much lower price point.

Large, real-world graphs are famously difficult to process efficiently. Not only they have a large memory footprint but most graph processing algorithms entail memory access patterns with poor locality, data-dependent parallelism, and a low compute-to- memory access ratio. Additionally, most real-world graphs have a low diameter and a highly heterogeneous node degree distribution. Partitioning these graphs and simultaneously achieve access locality and load-balancing is difficult if not impossible.

This paper evaluates the potential gains a workflow-aware storage system can bring. Two observations make us believe such storage system is crucial to efficiently support workflow-based applications: First, workflows generate irregular and application-dependent data access patterns. These patterns render existing storage systems unable to harness all optimization opportunities as this often requires conflicting optimization options or even conflicting design decision at the level of the storage system. Second, when scheduling, workflow runtime engines make suboptimal decisions as they lack detailed data location information. This paper discusses the feasibility, and evaluates the potential performance benefits brought by, building a workflow-aware storage system that supports per-file access optimizations and exposes data location. To this end, this paper presents approaches to determine the application-specific data access patterns, and evaluates experimentally the performance gains of a workflow-aware storage approach. Our evaluation using synthetic benchmarks shows that a workflow-aware storage system can bring significant performance gains: up to 7× performance gain compared to the distributed storage system - MosaStore and up to 16× compared to a central, well provisioned, NFS server.

This paper explores the ability to use graphics processing units (GPUs) as co-processors to harness the inherent parallelism of batch operations in systems that require high performance. To this end we have chosen bloom filters (space-efficient data structures that support the probabilistic representation of set membership) as the queries these data structures support are often performed in batches. Bloom filters exhibit low computational cost per amount of data, providing a baseline for more complex batch operations. We implemented BloomGPU a library that supports offloading bloom filter support to the GPU and evaluate this library under realistic usage scenarios. By completely offloading Bloom filter operations to the GPU, BloomGPU outperforms an optimized CPU implementation of the bloom filter as the workload becomes larger.

Grid computing is turning from promise into reality. Naturally, however, this is not happening for all applications at once. Bag-of-tasks (BoT) applications (those parallel applications whose tasks are independent) are, due to their simplicity, the first class of applications to be massively executed on grids (e.g. SETI@home). BoT applications are especially amendable for grid execution because they can run on intermittent resources (i.e. resources with no guarantees on availability or reliability). In this scenario, good performance and reliable results are provided by eager schedulers, which rely on replication to overcome unfortunate task-to-processor assignments. A little surprisingly, however, eager schedulers are not prepared to use space-shared resources (e.g. parallel supercomputers). This happens because using a space-shared resource involves submitting a detailed request to the resource scheduler, specifying the number of processors needed and the amount of time these processors are to be allocated - information that eager schedulers are not prepared to provide. This is unfortunate because space-shared computers are the most powerful computing resources available today and thus could greatly improve the execution time of BoT applications. This work proposes an automatic way to craft such requests in order to convert space-shared resources into intermittent ones, therefore rendering them naturally usable by eager schedulers. Such conversion is based on adaptive heuristics and allows eager schedulers use such powerful computing resources without modifications.

In this paper we discuss the difficulties involved in the scheduling of applications on computational grids. We highlight two main sources of difficulties: 1) the size of the grid rules out the possibility of using a centralized scheduler; 2) since resources are managed by different parties, the scheduler must consider several different policies. Thus, we argue that scheduling applications on a grid require the orchestration of several schedulers, with possibly conflicting goals. We discuss how we have addressed this issue in the context of PAUA, a grid for Bag-of-Tasks applications (i.e. parallel applications whose tasks are independent) that we are currently deploying throughout Brazil.

We here discuss how to run Bag-of-Tasks applications on computational grids. Bag-of-Tasks applications (those parallel applications whose tasks are independent) are both relevant and amendable for execution on grids. However, few users currently execute their Bag-of-Tasks applications on grids. We investigate the reason for this state of affairs and introduce MyGrid, a system designed to overcome the identified difficulties. MyGrid provides a simple, complete and secure way for a user to run Bag-of-Tasks applications on all resources she has access to. Besides putting together a complete solution useful for real users, MyGrid embeds two important research contributions to grid computing. First, we introduce some simple working environment abstractions that hide machine configuration heterogeneity from the user. Second, we introduce work queue with replication (WQR), a scheduling heuristics that attains good performance without relying on information about the grid or the application, although consuming a few more cycles. Note that not depending on information makes WQR much easier to deploy in practice.