There is an increasing need to bring machine learning to a wide diversity of hardware devices. Current frameworks rely on vendor-specific operator libraries and optimize for a narrow range of server-class GPUs. Deploying workloads to new platforms - such as mobile phones, embedded devices, and accelerators (e.g., FPGAs, ASICs) - requires significant manual effort. We propose TVM, a compiler that exposes graph-level and operator-level optimizations to provide performance portability to deep learning workloads across diverse hardware back-ends. TVM solves optimization challenges specific to deep learning, such as high-level operator fusion, mapping to arbitrary hardware primitives, and memory latency hiding. It also automates optimization of low-level programs to hardware characteristics by employing a novel, learning-based cost modeling method for rapid exploration of code optimizations. Experimental results show that TVM delivers performance across hardware back-ends that are competitive with state-of-the-art, hand-tuned libraries for low-power CPU, mobile GPU, and server-class GPUs. We also demonstrate TVM's ability to target new accelerator back-ends, such as the FPGA-based generic deep learning accelerator. The system is open sourced and in production use inside several major companies.

/pdf/tvm-an-automated-end-to-end-optimizing-compiler-for-deep-33mpz38jk4.pdf

TVM: an automated end-to-end optimizing compiler for deep learning

This is a thought piece on data-intensive science requirements for databases and science centers. It argues that peta-scale datasets will be housed by science centers that provide substantial storage and processing for scientists who access the data via smart notebooks. Next-generation science instruments and simulations will generate these peta-scale datasets. The need to publish and share data and the need for generic analysis and visualization tools will finally create a convergence on common metadata standards. Database systems will be judged by their support of these metadata standards and by their ability to manage and access peta-scale datasets. The procedural stream-of-bytes-file-centric approach to data analysis is both too cumbersome and too serial for such large datasets. Non-procedural query and analysis of schematized self-describing data is both easier to use and allows much more parallelism.

Scientific Data Management in the Coming Decade

As both CPUs and GPUs become employed in a wide range of applications, it has been acknowledged that both of these Processing Units (PUs) have their unique features and strengths and hence, CPU-GPU collaboration is inevitable to achieve high-performance computing. This has motivated a significant amount of research on heterogeneous computing techniques, along with the design of CPU-GPU fused chips and petascale heterogeneous supercomputers. In this article, we survey Heterogeneous Computing Techniques (HCTs) such as workload partitioning that enable utilizing both CPUs and GPUs to improve performance and/or energy efficiency. We review heterogeneous computing approaches at runtime, algorithm, programming, compiler, and application levels. Further, we review both discrete and fused CPU-GPU systems and discuss benchmark suites designed for evaluating Heterogeneous Computing Systems (HCSs). We believe that this article will provide insights into the workings and scope of applications of HCTs to researchers and motivate them to further harness the computational powers of CPUs and GPUs to achieve the goal of exascale performance.

https://dl.acm.org/doi/pdf/10.1145/2788396

A Survey of CPU-GPU Heterogeneous Computing Techniques

Growing main memory capacity has fueled the development of in-memory big data management and processing. By eliminating disk I/O bottleneck, it is now possible to support interactive data analytics. However, in-memory systems are much more sensitive to other sources of overhead that do not matter in traditional I/O-bounded disk-based systems. Some issues such as fault-tolerance and consistency are also more challenging to handle in in-memory environment. We are witnessing a revolution in the design of database systems that exploits main memory as its data storage layer. Many of these researches have focused along several dimensions: modern CPU and memory hierarchy utilization, time/space efficiency, parallelism, and concurrency control. In this survey, we aim to provide a thorough review of a wide range of in-memory data management and processing proposals and systems, including both data storage systems and data processing frameworks. We also give a comprehensive presentation of important technology in memory management, and some key factors that need to be considered in order to achieve efficient in-memory data management and processing.

/pdf/in-memory-big-data-management-and-processing-a-survey-1k2b0h2jut.pdf

In-Memory Big Data Management and Processing: A Survey

With modern computer architecture evolving, two problems conspire against the state-of-the-art approaches in parallel query execution: (i) to take advantage of many-cores, all query work must be distributed evenly among (soon) hundreds of threads in order to achieve good speedup, yet (ii) dividing the work evenly is difficult even with accurate data statistics due to the complexity of modern out-of-order cores. As a result, the existing approaches for plan-driven parallelism run into load balancing and context-switching bottlenecks, and therefore no longer scale. A third problem faced by many-core architectures is the decentralization of memory controllers, which leads to Non-Uniform Memory Access (NUMA). In response, we present the morsel-driven query execution framework, where scheduling becomes a fine-grained run-time task that is NUMA-aware. Morsel-driven query processing takes small fragments of input data (morsels) and schedules these to worker threads that run entire operator pipelines until the next pipeline breaker. The degree of parallelism is not baked into the plan but can elastically change during query execution, so the dispatcher can react to execution speed of different morsels but also adjust resources dynamically in response to newly arriving queries in the workload. Further, the dispatcher is aware of data locality of the NUMA-local morsels and operator state, such that the great majority of executions takes place on NUMA-local memory. Our evaluation on the TPC-H and SSB benchmarks shows extremely high absolute performance and an average speedup of over 30 with 32 cores.

/pdf/morsel-driven-parallelism-a-numa-aware-query-evaluation-25l29dmzgg.pdf

Morsel-driven parallelism: a NUMA-aware query evaluation framework for the many-core age

The multi-core architectures of today's computer systems make parallelism a necessity for performance critical applications. Writing such applications in a generic, hardware-oblivious manner is a challenging problem: Current database systems thus rely on labor-intensive and error-prone manual tuning to exploit the full potential of modern parallel hardware architectures like multi-core CPUs and graphics cards. We propose an alternative design for a parallel database engine, based on a single set of hardware-oblivious operators, which are compiled down to the actual hardware at runtime. This design reduces the development overhead for parallel database engines, while achieving competitive performance to hand-tuned systems.We provide a proof-of-concept for this design by integrating operators written using the parallel programming framework OpenCL into the open-source database MonetDB. Following this approach, we achieve efficient, yet highly portable parallel code without the need for optimization by hand. We evaluated our implementation against MonetDB using TPC-H derived queries and observed a performance that rivals that of MonetDB's query execution on the CPU and surpasses it on the GPU. In addition, we show that the same set of operators runs nearly unchanged on a GPU, demonstrating the feasibility of our approach.

/pdf/hardware-oblivious-parallelism-for-in-memory-column-stores-2ecva5k5kx.pdf

Hardware-oblivious parallelism for in-memory column-stores

Methods, systems, and computer readable medium for displaying results of a search query. In one implementation, the method includes receiving a query, obtaining documents that satisfy the query, constructing a facet hierarchy based on documents that satisfy the query, creating a cube structure based on the facet hierarchy, and displaying a multi-dimensional search interface based on the cube structure.

Cube faceted data analysis

In-memory databases require careful tuning and many engineering tricks to achieve good performance. Such database performance engineering is hard: a plethora of data and hardware-dependent optimization techniques form a design space that is difficult to navigate for a skilled engineer --- even more so for a query compiler. To facilitate performance-oriented design exploration and query plan compilation, we present Voodoo, a declarative intermediate algebra that abstracts the detailed architectural properties of the hardware, such as multi- or many-core architectures, caches and SIMD registers, without losing the ability to generate highly tuned code. Because it consists of a collection of declarative, vector-oriented operations, Voodoo is easier to reason about and tune than low-level C and related hardware-focused extensions (Intrinsics, OpenCL, CUDA, etc.). This enables our Voodoo compiler to produce (OpenCL) code that rivals and even outperforms the fastest state-of-the-art in memory databases for both GPUs and CPUs. In addition, Voodoo makes it possible to express techniques as diverse as cache-conscious processing, predication and vectorization (again on both GPUs and CPUs) with just a few lines of code. Central to our approach is a novel idea we termed control vectors, which allows a code generating frontend to expose parallelism to the Voodoo compiler in a abstract manner, enabling portable performance across hardware platforms.We used Voodoo to build an alternative backend for MonetDB, a popular open-source in-memory database. Our backend allows MonetDB to perform at the same level as highly tuned in-memory databases, including HyPeR and Ocelot. We also demonstrate Voodoo's usefulness when investigating hardware conscious tuning techniques, assessing their performance on different queries, devices and data.

/pdf/voodoo-a-vector-algebra-for-portable-database-performance-on-43krfagdgv.pdf

Voodoo - a vector algebra for portable database performance on modern hardware

Modern analytics applications use a diverse mix of libraries and functions. Unfortunately, there is no optimization across these libraries, resulting in performance penalties as high as an order of magnitude in many applications. To address this problem, we proposed Weld, a common runtime for existing data analytics libraries that performs key physical optimizations such as pipelining under existing, imperative library APIs. In this work, we further develop the Weld vision by designing an automatic adaptive optimizer for Weld applications, and evaluating its impact on realistic data science workloads. Our optimizer eliminates multiple forms of overhead that arise when composing imperative libraries like Pandas and NumPy, and uses lightweight measurements to make data-dependent decisions at run-time in ad-hoc workloads where no statistics are available, with sub-second overhead. We also evaluate which optimizations have the largest impact in practice and whether Weld can be integrated into libraries incrementally. Our results are promising: using our optimizer, Weld accelerates data science workloads by up to 23X on one thread and 80X on eight threads, and its adaptive optimizations provide up to a 3.75X speedup over rule-based optimization. Moreover, Weld provides benefits if even just 4--5 operators in a library are ported to use it. Our results show that common runtime designs like Weld may be a viable approach to accelerate analytics.

/pdf/evaluating-end-to-end-optimization-for-data-analytics-r2r8wzgv53.pdf

Evaluating end-to-end optimization for data analytics applications in weld

Database Cracking is an appealing approach to adaptive indexing: on every range-selection query, the data is partitioned using the supplied predicates as pivots. The core of database cracking is, thus, pivoted partitioning. While pivoted partitioning, like scanning, requires a single pass through the data it tends to have much higher costs due to lower CPU efficiency. In this paper, we conduct an in-depth study of the reasons for the low CPU efficiency of pivoted partitioning. Based on the findings, we develop an optimized version with significantly higher (single-threaded) CPU efficiency. We also develop a number of multi-threaded implementations that are effectively bound by memory bandwidth. Combining all of these optimizations we achieve an implementation that has costs close to or better than an ordinary scan on a variety of systems ranging from low-end (cheaper than $300) desktop machines to high-end (above $60,000) servers.

/pdf/database-cracking-fancy-scan-not-poor-man-s-sort-2e5z2l7izn.pdf

Holger Pirk

Papers

Hardware-oblivious parallelism for in-memory column-stores

Cube faceted data analysis

Voodoo - a vector algebra for portable database performance on modern hardware

Evaluating end-to-end optimization for data analytics applications in weld

Database cracking: fancy scan, not poor man's sort!