arXiv subjects

Daniel Lemire

Publications and source records attributed to Daniel Lemire.

At least 37 records · Page 2Linked to original sources

arXiv2022

Number Parsing at a Gigabyte per Second

With disks and networks providing gigabytes per second, parsing decimal numbers from strings becomes a bottleneck. We consider the problem of parsing decimal numbers to the nearest binary floating-point value. The general problem requires variable-precision arithmetic. However, we need at most 17 digits to represent 64-bit standard floating-point numbers (IEEE 754). Thus we can represent the decimal significand with a single 64-bit word. By combining the significand and precomputed tables, we can compute the nearest floating-point number using as few as one or two 64-bit multiplications. Our implementation can be several times faster than conventional functions present in standard C libraries on modern 64-bit systems (Intel, AMD, ARM and POWER9). Our work is available as open source software used by major systems such as Apache Arrow and Yandex ClickHouse. The Go standard library has adopted a version of our approach.

cs.DS↗

arXiv2022

Roaring Bitmaps: Implementation of an Optimized Software Library

Compressed bitmap indexes are used in systems such as Git or Oracle to accelerate queries. They represent sets and often support operations such as unions, intersections, differences, and symmetric differences. Several important systems such as Elasticsearch, Apache Spark, Netflix's Atlas, LinkedIn's Pinot, Metamarkets' Druid, Pilosa, Apache Hive, Apache Tez, Microsoft Visual Studio Team Services and Apache Kylin rely on a specific type of compressed bitmap index called Roaring. We present an optimized software library written in C implementing Roaring bitmaps: CRoaring. It benefits from several algorithms designed for the single-instruction-multiple-data (SIMD) instructions available on commodity processors. In particular, we present vectorized algorithms to compute the intersection, union, difference and symmetric difference between arrays. We benchmark the library against a wide range of competitive alternatives, identifying weaknesses and strengths in our software. Our work is available under a liberal open-source license.

cs.DB↗

arXiv2022

Binary Fuse Filters: Fast and Smaller Than Xor Filters

Bloom and cuckoo filters provide fast approximate set membership while using little memory. Engineers use them to avoid expensive disk and network accesses. The recently introduced xor filters can be faster and smaller than Bloom and cuckoo filters. The xor filters are within 23% of the theoretical lower bound in storage as opposed to 44% for Bloom filters. Inspired by Dietzfelbinger and Walzer, we build probabilistic filters -- called binary fuse filters -- that are within 13% of the storage lower bound -- without sacrificing query speed. As an additional benefit, the construction of the new binary fuse filters can be more than twice as fast as the construction of xor filters. By slightly sacrificing query speed, we further reduce storage to within 8% of the lower bound. We compare the performance against a wide range of competitive alternatives such as Bloom filters, blocked Bloom filters, vector quotient filters, cuckoo filters, and the recent ribbon filters. Our experiments suggest that binary fuse filters are superior to xor filters.

cs.DS↗

arXiv2021

Integer Division by Constants: Optimal Bounds

The integer division of a numerator n by a divisor d gives a quotient q and a remainder r. Optimizing compilers accelerate software by replacing the division of n by d with the division of c * n (or c * n + c) by m for convenient integers c and m chosen so that they approximate the reciprocal: c/m ~= 1/d. Such techniques are especially advantageous when m is chosen to be a power of two and when d is a constant so that c and m can be precomputed. The literature contains many bounds on the distance between c/m and the divisor d. Some of these bounds are optimally tight, while others are not. We present optimally tight bounds for quotient and remainder computations.

cs.DS↗

arXiv2021

Efficient Computation of Positional Population Counts Using SIMD Instructions

In several fields such as statistics, machine learning, and bioinformatics, categorical variables are frequently represented as one-hot encoded vectors. For example, given 8 distinct values, we map each value to a byte where only a single bit has been set. We are motivated to quickly compute statistics over such encodings. Given a stream of k-bit words, we seek to compute k distinct sums corresponding to bit values at indexes 0, 1, 2, ..., k-1. If the k-bit words are one-hot encoded then the sums correspond to a frequency histogram. This multiple-sum problem is a generalization of the population-count problem where we seek the sum of all bit values. Accordingly, we refer to the multiple-sum problem as a positional population-count. Using SIMD (Single Instruction, Multiple Data) instructions from recent Intel processors, we describe algorithms for computing the 16-bit position population count using less than half of a CPU cycle per 16-bit word. Our best approach uses up to 400 times fewer instructions and is up to 50 times faster than baseline code using only regular (non-SIMD) instructions, for sufficiently large inputs.

cs.DS↗

arXiv2021

Decoding billions of integers per second through vectorization

In many important applications -- such as search engines and relational database systems -- data is stored in the form of arrays of integers. Encoding and, most importantly, decoding of these arrays consumes considerable CPU time. Therefore, substantial effort has been made to reduce costs associated with compression and decompression. In particular, researchers have exploited the superscalar nature of modern processors and SIMD instructions. Nevertheless, we introduce a novel vectorized scheme called SIMD-BP128 that improves over previously proposed vectorized approaches. It is nearly twice as fast as the previously fastest schemes on desktop processors (varint-G8IU and PFOR). At the same time, SIMD-BP128 saves up to 2 bits per integer. For even better compression, we propose another new vectorized scheme (SIMD-FastPFOR) that has a compression ratio within 10% of a state-of-the-art scheme (Simple-8b) while being two times faster during decoding.

cs.IR↗

arXiv2020

Hierarchical Bin Buffering: Online Local Moments for Dynamic External Memory Arrays

Local moments are used for local regression, to compute statistical measures such as sums, averages, and standard deviations, and to approximate probability distributions. We consider the case where the data source is a very large I/O array of size n and we want to compute the first N local moments, for some constant N. Without precomputation, this requires O(n) time. We develop a sequence of algorithms of increasing sophistication that use precomputation and additional buffer space to speed up queries. The simpler algorithms partition the I/O array into consecutive ranges called bins, and they are applicable not only to local-moment queries, but also to algebraic queries (MAX, AVERAGE, SUM, etc.). With N buffers of size sqrt{n}, time complexity drops to O(sqrt n). A more sophisticated approach uses hierarchical buffering and has a logarithmic time complexity (O(b log_b n)), when using N hierarchical buffers of size n/b. Using Overlapped Bin Buffering, we show that only a single buffer is needed, as with wavelet-based algorithms, but using much less storage. Applications exist in multidimensional and statistical databases over massive data sets, interactive image processing, and visualization.

cs.DS↗

arXiv2020

SIMD Compression and the Intersection of Sorted Integers

Sorted lists of integers are commonly used in inverted indexes and database systems. They are often compressed in memory. We can use the SIMD instructions available in common processors to boost the speed of integer compression schemes. Our S4-BP128-D4 scheme uses as little as 0.7 CPU cycles per decoded integer while still providing state-of-the-art compression. However, if the subsequent processing of the integers is slow, the effort spent on optimizing decoding speed can be wasted. To show that it does not have to be so, we (1) vectorize and optimize the intersection of posting lists; (2) introduce the SIMD Galloping algorithm. We exploit the fact that one SIMD instruction can compare 4 pairs of integers at once. We experiment with two TREC text collections, GOV2 and ClueWeb09 (Category B), using logs from the TREC million-query track. We show that using only the SIMD instructions ubiquitous in all modern CPUs, our techniques for conjunctive queries can double the speed of a state-of-the-art approach.

cs.IR↗

arXiv2020

Xor Filters: Faster and Smaller Than Bloom and Cuckoo Filters

The Bloom filter provides fast approximate set membership while using little memory. Engineers often use these filters to avoid slow operations such as disk or network accesses. As an alternative, a cuckoo filter may need less space than a Bloom filter and it is faster. Chazelle et al. proposed a generalization of the Bloom filter called the Bloomier filter. Dietzfelbinger and Pagh described a variation on the Bloomier filter that can be used effectively for approximate membership queries. It has never been tested empirically, to our knowledge. We review an efficient implementation of their approach, which we call the xor filter. We find that xor filters can be faster than Bloom and cuckoo filters while using less memory. We further show that a more compact version of xor filters (xor+) can use even less space than highly compact alternatives (e.g., Golomb-compressed sequences) while providing speeds competitive with Bloom filters.

cs.DS↗

arXiv2019

Faster Remainder by Direct Computation: Applications to Compilers and Software Libraries

On common processors, integer multiplication is many times faster than integer division. Dividing a numerator n by a divisor d is mathematically equivalent to multiplication by the inverse of the divisor (n / d = n x 1/d). If the divisor is known in advance---or if repeated integer divisions will be performed with the same divisor---it can be beneficial to substitute a less costly multiplication for an expensive division. Currently, the remainder of the division by a constant is computed from the quotient by a multiplication and a subtraction. But if just the remainder is desired and the quotient is unneeded, this may be suboptimal. We present a generally applicable algorithm to compute the remainder more directly. Specifically, we use the fractional portion of the product of the numerator and the inverse of the divisor. On this basis, we also present a new, simpler divisibility algorithm to detect nonzero remainders. We also derive new tight bounds on the precision required when representing the inverse of the divisor. Furthermore, we present simple C implementations that beat the optimized code produced by state-of-art C compilers on recent x64 processors (e.g., Intel Skylake and AMD Ryzen), sometimes by more than 25%. On all tested platforms including 64-bit ARM and POWER8, our divisibility-test functions are faster than state-of-the-art Granlund-Montgomery divisibility-test functions, sometimes by more than 50%.

cs.MS↗

arXiv2019

Base64 encoding and decoding at almost the speed of a memory copy

Many common document formats on the Internet are text-only such as email (MIME) and the Web (HTML, JavaScript, JSON and XML). To include images or executable code in these documents, we first encode them as text using base64. Standard base64 encoding uses 64~ASCII characters: both lower and upper case Latin letters, digits and two other symbols. We show how we can encode and decode base64 data at nearly the speed of a memory copy (memcpy) on recent Intel processors, as long as the data does not fit in the first-level (L1) cache. We use the SIMD (Single Instruction Multiple Data) instruction set AVX-512 available on commodity processors. Our implementation generates several times fewer instructions than previous SIMD-accelerated base64 codecs. It is also more versatile, as it can be adapted---even at runtime---to any base64 variant by only changing constants.

cs.DC↗

arXiv2018

Fast Random Integer Generation in an Interval

In simulations, probabilistic algorithms and statistical tests, we often generate random integers in an interval (e.g., [0,s)). For example, random integers in an interval are essential to the Fisher-Yates random shuffle. Consequently, popular languages like Java, Python, C++, Swift and Go include ranged random integer generation functions as part of their runtime libraries. Pseudo-random values are usually generated in words of a fixed number of bits (e.g., 32 bits, 64 bits) using algorithms such as a linear congruential generator. We need functions to convert such random words to random integers in an interval ([0,s)) without introducing statistical biases. The standard functions in programming languages such as Java involve integer divisions. Unfortunately, division instructions are relatively expensive. We review an unbiased function to generate ranged integers from a source of random words that avoids integer divisions with high probability. To establish the practical usefulness of the approach, we show that this algorithm can multiply the speed of unbiased random shuffling on x64 processors. Our proposed approach has been adopted by the Go language for its implementation of the shuffle function.

cs.DS↗

arXiv2018

Xorshift1024*, Xorshift1024+, Xorshift128+ and Xoroshiro128+ Fail Statistical Tests for Linearity

L'Ecuyer & Simard's Big Crush statistical test suite has revealed statistical flaws in many popular random number generators including Marsaglia's Xorshift generators. Vigna recently proposed some 64-bit variations on the Xorshift scheme that are further scrambled (i.e., Xorshift1024*, Xorshift1024+, Xorshift128+, Xoroshiro128+). Unlike their unscrambled counterparts, they pass Big Crush when interleaving blocks of 32 bits for each 64-bit word (most significant, least significant, most significant, least significant, etc.). We report that these scrambled generators systematically fail Big Crush---specifically the linear-complexity and matrix-rank tests that detect linearity---when taking the 32 lowest-order bits in reverse order from each 64-bit word.

cs.DS↗

arXiv2018

Strongly universal string hashing is fast

We present fast strongly universal string hashing families: they can process data at a rate of 0.2 CPU cycle per byte. Maybe surprisingly, we find that these families---though they require a large buffer of random numbers---are often faster than popular hash functions with weaker theoretical guarantees. Moreover, conventional wisdom is that hash functions with fewer multiplications are faster. Yet we find that they may fail to be faster due to operation pipelining. We present experimental results on several processors including low-powered processors. Our tests include hash functions designed for processors with the Carry-Less Multiplication (CLMUL) instruction set. We also prove, using accessible proofs, the strong universality of our families.

cs.DB↗

arXiv2018

Faster Population Counts Using AVX2 Instructions

Counting the number of ones in a binary stream is a common operation in database, information-retrieval, cryptographic and machine-learning applications. Most processors have dedicated instructions to count the number of ones in a word (e.g., popcnt on x64 processors). Maybe surprisingly, we show that a vectorized approach using SIMD instructions can be twice as fast as using the dedicated instructions on recent Intel processors. The benefits can be even greater for applications such as similarity measures (e.g., the Jaccard index) that require additional Boolean operations. Our approach has been adopted by LLVM: it is used by its popular C compiler (clang).

cs.DS↗

arXiv2018

Consistently faster and smaller compressed bitmaps with Roaring

Compressed bitmap indexes are used in databases and search engines. Many bitmap compression techniques have been proposed, almost all relying primarily on run-length encoding (RLE). However, on unsorted data, we can get superior performance with a hybrid compression technique that uses both uncompressed bitmaps and packed arrays inside a two-level tree. An instance of this technique, Roaring, has recently been proposed. Due to its good performance, it has been adopted by several production platforms (e.g., Apache Lucene, Apache Spark, Apache Kylin and Druid). Yet there are cases where run-length encoded bitmaps are smaller than the original Roaring bitmaps---typically when the data is sorted so that the bitmaps contain long compressible runs. To better handle these cases, we build a new Roaring hybrid that combines uncompressed bitmaps, packed arrays and RLE compressed segments. The result is a new Roaring format that compresses better. Overall, our new implementation of Roaring can be several times faster (up to two orders of magnitude) than the implementations of traditional RLE-based alternatives (WAH, Concise, EWAH) while compressing better. We review the design choices and optimizations that make these good results possible.

cs.DB↗

arXiv2018

Apache Calcite: A Foundational Framework for Optimized Query Processing Over Heterogeneous Data Sources

Apache Calcite is a foundational software framework that provides query processing, optimization, and query language support to many popular open-source data processing systems such as Apache Hive, Apache Storm, Apache Flink, Druid, and MapD. Calcite's architecture consists of a modular and extensible query optimizer with hundreds of built-in optimization rules, a query processor capable of processing a variety of query languages, an adapter architecture designed for extensibility, and support for heterogeneous data models and stores (relational, semi-structured, streaming, and geospatial). This flexible, embeddable, and extensible architecture is what makes Calcite an attractive choice for adoption in big-data frameworks. It is an active project that continues to introduce support for the new types of data sources, query languages, and approaches to query processing and optimization.

cs.DB↗

arXiv2017

Stream VByte: Faster Byte-Oriented Integer Compression

Arrays of integers are often compressed in search engines. Though there are many ways to compress integers, we are interested in the popular byte-oriented integer compression techniques (e.g., VByte or Google's Varint-GB). They are appealing due to their simplicity and engineering convenience. Amazon's varint-G8IU is one of the fastest byte-oriented compression technique published so far. It makes judicious use of the powerful single-instruction-multiple-data (SIMD) instructions available in commodity processors. To surpass varint-G8IU, we present Stream VByte, a novel byte-oriented compression technique that separates the control stream from the encoded data. Like varint-G8IU, Stream VByte is well suited for SIMD instructions. We show that Stream VByte decoding can be up to twice as fast as varint-G8IU decoding over real data sets. In this sense, Stream VByte establishes new speed records for byte-oriented integer compression, at times exceeding the speed of the memcpy function. On a 3.4GHz Haswell processor, it decodes more than 4 billion differentially-coded integers per second from RAM to L1 cache.

cs.IR↗