arXiv subjects

Shivnath Babu

Publications and source records attributed to Shivnath Babu.

6 recordsLinked to original sources

arXiv2020

Black or White? How to Develop an AutoTuner for Memory-based Analytics [Extended Version]

There is a lot of interest today in building autonomous (or, self-driving) data processing systems. An emerging school of thought is to leverage AI-driven "black box" algorithms for this purpose. In this paper, we present a contrarian view. We study the problem of autotuning the memory allocation for applications running on modern distributed data processing systems. For this problem, we show that an empirically-driven "white-box" algorithm, called RelM, that we have developed provides a close-to-optimal tuning at a fraction of the overheads compared to state-of-the-art AI-driven "black box" algorithms, namely, Bayesian Optimization (BO) and Deep Distributed Policy Gradient (DDPG). The main reason for RelM's superior performance is that the memory management in modern memory-based data analytics systems is an interplay of algorithms at multiple levels: (i) at the resource-management level across various containers allocated by resource managers like Kubernetes and YARN, (ii) at the container level among the OS, pods, and processes such as the Java Virtual Machine (JVM), (iii) at the application level for caching, aggregation, data shuffles, and application data structures, and (iv) at the JVM level across various pools such as the Young and Old Generation. RelM understands these interactions and uses them in building an analytical solution to autotune the memory management knobs. In another contribution, called GBO, we use the RelM's analytical models to speed up Bayesian Optimization. Through an evaluation based on Apache Spark, we showcase that RelM's recommendations are significantly better than what commonly-used Spark deployments provide, and are close to the ones obtained by brute-force exploration; while GBO provides optimality guarantees for a higher, but still significantly lower compared to the state-of-the-art AI-driven policies, cost overhead.

cs.DC↗

arXiv2018

Analyzing Query Performance and Attributing Blame for Contentions in a Cluster Computing Framework

There are many approaches is use today to either prevent or minimize the impact of inter-query interactions on a shared cluster. Despite these measures, performance issues due to concurrent executions of mixed workloads still prevail causing undue waiting times for queries. Analyzing these resource interferences is thus critical in order to answer time sensitive questions like 'who is causing my query to slowdown' in a multi-tenant environment. More importantly, dignosing whether the slowdown of a query is a result of resource contentions caused by other queries or some other external factor can help an admin narrow down the many possibilities of performance degradation. This process of investigating the symptoms of resource contentions and attributing blame to concurrent queries is non-trivial and tedious, and involves hours of manually debugging through a cycle of query interactions. In this paper, we present ProtoXplore - a Proto or first system to eXplore contentions, that helps administrators determine whether the blame for resource bottlenecks can be attributed to concurrent queries, and uses a methodology called Resource Acquire Time Penalty (RATP) to quantify this blame towards contentious sources accurately. Further, ProtoXplore builds on the theory of explanations and enables a step-wise deep exploration of various levels of performance bottlenecks faced by a query during its execution using a multi-level directed acyclic graph called ProtoGraph. Our experimental evaluation uses ProtoXplore to analyze the interactions between TPC-DS queries on Apache Spark to show how ProtoXplore provides explanations that help in diagnosing contention related issues and better managing a changing mixed workload in a shared cluster.

cs.DC↗

arXiv2015

Tempo: Robust and Self-Tuning Resource Management in Multi-tenant Parallel Databases

Multi-tenant database systems have a component called the Resource Manager, or RM that is responsible for allocating resources to tenants. RMs today do not provide direct support for performance objectives such as: "Average job response time of tenant A must be less than two minutes", or "No more than 5% of tenant B's jobs can miss the deadline of 1 hour." Thus, DBAs have to tinker with the RM's low-level configuration settings to meet such objectives. We propose a framework called Tempo that brings simplicity, self-tuning, and robustness to existing RMs. Tempo provides a simple interface for DBAs to specify performance objectives declaratively, and optimizes the RM configuration settings to meet these objectives. Tempo has a solid theoretical foundation which gives key robustness guarantees. We report experiments done on Tempo using production traces of data-processing workloads from companies such as Facebook and Cloudera. These experiments demonstrate significant improvements in meeting desired performance objectives over RM configuration settings specified by human experts.

cs.DB↗

arXiv2015

ROBUS: Fair Cache Allocation for Multi-tenant Data-parallel Workloads

Systems for processing big data---e.g., Hadoop, Spark, and massively parallel databases---need to run workloads on behalf of multiple tenants simultaneously. The abundant disk-based storage in these systems is usually complemented by a smaller, but much faster, {\em cache}. Cache is a precious resource: Tenants who get to use cache can see two orders of magnitude performance improvement. Cache is also a limited and hence shared resource: Unlike a resource like a CPU core which can be used by only one tenant at a time, a cached data item can be accessed by multiple tenants at the same time. Cache, therefore, has to be shared by a multi-tenancy-aware policy across tenants, each having a unique set of priorities and workload characteristics. In this paper, we develop cache allocation strategies that speed up the overall workload while being {\em fair} to each tenant. We build a novel fairness model targeted at the shared resource setting that incorporates not only the more standard concepts of Pareto-efficiency and sharing incentive, but also define envy freeness via the notion of {\em core} from cooperative game theory. Our cache management platform, ROBUS, uses randomization over small time batches, and we develop a proportionally fair allocation mechanism that satisfies the core property in expectation. We show that this algorithm and related fair algorithms can be approximated to arbitrary precision in polynomial time. We evaluate these algorithms on a ROBUS prototype implemented on Spark with RDD store used as cache. Our evaluation on a synthetically generated industry-standard workload shows that our algorithms provide a speedup close to performance optimal algorithms while guaranteeing fairness across tenants.

cs.DC↗

arXiv2012

Stubby: A Transformation-based Optimizer for MapReduce Workflows

There is a growing trend of performing analysis on large datasets using workflows composed of MapReduce jobs connected through producer-consumer relationships based on data. This trend has spurred the development of a number of interfaces--ranging from program-based to query-based interfaces--for generating MapReduce workflows. Studies have shown that the gap in performance can be quite large between optimized and unoptimized workflows. However, automatic cost-based optimization of MapReduce workflows remains a challenge due to the multitude of interfaces, large size of the execution plan space, and the frequent unavailability of all types of information needed for optimization. We introduce a comprehensive plan space for MapReduce workflows generated by popular workflow generators. We then propose Stubby, a cost-based optimizer that searches selectively through the subspace of the full plan space that can be enumerated correctly and costed based on the information available in any given setting. Stubby enumerates the plan space based on plan-to-plan transformations and an efficient search algorithm. Stubby is designed to be extensible to new interfaces and new types of optimizations, which is a desirable feature given how rapidly MapReduce systems are evolving. Stubby's efficiency and effectiveness have been evaluated using representative workflows from many domains.

cs.DB↗

arXiv2011

Why Did My Query Slow Down?

Many enterprise environments have databases running on network-attached server-storage infrastructure (referred to as Storage Area Networks or SANs). Both the database and the SAN are complex systems that need their own separate administrative teams. This paper puts forth the vision of an innovative management framework to simplify administrative tasks that require an in-depth understanding of both the database and the SAN. As a concrete instance, we consider the task of diagnosing the slowdown in performance of a database query that is executed multiple times (e.g., in a periodic report-generation setting). This task is very challenging because the space of possible causes includes problems specific to the database, problems specific to the SAN, and problems that arise due to interactions between the two systems. In addition, the monitoring data available from these systems can be noisy. We describe the design of DIADS which is an integrated diagnosis tool for database and SAN administrators. DIADS generates and uses a powerful abstraction called Annotated Plan Graphs (APGs) that ties together the execution path of queries in the database and the SAN. Using an innovative workflow that combines domain-specific knowledge with machine-learning techniques, DIADS was applied successfully to diagnose query slowdowns caused by complex combinations of events across a PostgreSQL database and a production SAN.

cs.DB↗