FLASH: Randomized Algorithms Accelerated over CPU-GPU for Ultra-High Dimensional Similarity Search. Wang, Y., Shrivastava, A., & Ryu, J. arXiv:1709.01190 [cs], September, 2017. arXiv: 1709.01190![FLASH: Randomized Algorithms Accelerated over CPU-GPU for Ultra-High Dimensional Similarity Search [link]](https://bibbase.org/img/filetypes/link.svg "link")
Paper abstract bibtex We present FLASH (\\textbackslashbf F\ast \\textbackslashbf L\SH \\textbackslashbf A\lgorithm for \\textbackslashbf S\imilarity search accelerated with \\textbackslashbf H\PC (High-Performance Computing)), a similarity search system for ultra-high dimensional datasets on a single machine, which does not require similarity computation. Our system is an auspicious illustration of the power of randomized algorithms carefully tailored for high-performance computing platforms. We leverage LSH style randomized indexing procedure and combine it with several principled techniques, such as reservoir sampling, recent advances in one-pass minwise hashing, and count based estimations. The combination, while retaining sound theoretical guarantees, reduces the computational as well as parallelization overhead of our proposal. We provide CPU and hybrid CPU-GPU implementations of FLASH for replicability of our results https://github.com/RUSH-LAB/Flash. We evaluate FLASH on several real high dimensional datasets coming from different domains including text, malicious URL, click-through prediction, social networks, etc. Our experiments shed new light on the difficulties associated with datasets having several millions of dimensions. Current state-of-the-art implementations either fail on the presented scale or are orders of magnitude slower than our system. FLASH is capable of computing an approximate k-NN graph, from scratch, over full webspam dataset (1.3 billion nonzeros) in less than 10 seconds. Computing full k-NN graph in less than 10 seconds on webspam dataset, using brute-force (\$n{\textasciicircum}2D\$), will require at least 20 TFLOPS. We hope that FLASH gets adopted in practice.
@article{wang_flash:_2017-1,
title = {{FLASH}: {Randomized} {Algorithms} {Accelerated} over {CPU}-{GPU} for {Ultra}-{High} {Dimensional} {Similarity} {Search}},
shorttitle = {{FLASH}},
url = {http://arxiv.org/abs/1709.01190},
abstract = {We present FLASH (\{{\textbackslash}bf F\}ast \{{\textbackslash}bf L\}SH \{{\textbackslash}bf A\}lgorithm for \{{\textbackslash}bf S\}imilarity search accelerated with \{{\textbackslash}bf H\}PC (High-Performance Computing)), a similarity search system for ultra-high dimensional datasets on a single machine, which does not require similarity computation. Our system is an auspicious illustration of the power of randomized algorithms carefully tailored for high-performance computing platforms. We leverage LSH style randomized indexing procedure and combine it with several principled techniques, such as reservoir sampling, recent advances in one-pass minwise hashing, and count based estimations. The combination, while retaining sound theoretical guarantees, reduces the computational as well as parallelization overhead of our proposal. We provide CPU and hybrid CPU-GPU implementations of FLASH for replicability of our results https://github.com/RUSH-LAB/Flash. We evaluate FLASH on several real high dimensional datasets coming from different domains including text, malicious URL, click-through prediction, social networks, etc. Our experiments shed new light on the difficulties associated with datasets having several millions of dimensions. Current state-of-the-art implementations either fail on the presented scale or are orders of magnitude slower than our system. FLASH is capable of computing an approximate k-NN graph, from scratch, over full webspam dataset (1.3 billion nonzeros) in less than 10 seconds. Computing full k-NN graph in less than 10 seconds on webspam dataset, using brute-force (\$n{\textasciicircum}2D\$), will require at least 20 TFLOPS. We hope that FLASH gets adopted in practice.},
urldate = {2018-02-08TZ},
journal = {arXiv:1709.01190 [cs]},
author = {Wang, Yiqiu and Shrivastava, Anshumali and Ryu, Junghee},
month = sep,
year = {2017},
note = {arXiv: 1709.01190},
keywords = {⛔ No DOI found}
}
Downloads: 0
{"_id":"LADbNGtjD8ZZo3vXE","bibbaseid":"wang-shrivastava-ryu-flashrandomizedalgorithmsacceleratedovercpugpuforultrahighdimensionalsimilaritysearch-2017","authorIDs":[],"author_short":["Wang, Y.","Shrivastava, A.","Ryu, J."],"bibdata":{"bibtype":"article","type":"article","title":"FLASH: Randomized Algorithms Accelerated over CPU-GPU for Ultra-High Dimensional Similarity Search","shorttitle":"FLASH","url":"http://arxiv.org/abs/1709.01190","abstract":"We present FLASH (\\\\textbackslashbf F\\ast \\\\textbackslashbf L\\SH \\\\textbackslashbf A\\lgorithm for \\\\textbackslashbf S\\imilarity search accelerated with \\\\textbackslashbf H\\PC (High-Performance Computing)), a similarity search system for ultra-high dimensional datasets on a single machine, which does not require similarity computation. Our system is an auspicious illustration of the power of randomized algorithms carefully tailored for high-performance computing platforms. We leverage LSH style randomized indexing procedure and combine it with several principled techniques, such as reservoir sampling, recent advances in one-pass minwise hashing, and count based estimations. The combination, while retaining sound theoretical guarantees, reduces the computational as well as parallelization overhead of our proposal. We provide CPU and hybrid CPU-GPU implementations of FLASH for replicability of our results https://github.com/RUSH-LAB/Flash. We evaluate FLASH on several real high dimensional datasets coming from different domains including text, malicious URL, click-through prediction, social networks, etc. Our experiments shed new light on the difficulties associated with datasets having several millions of dimensions. Current state-of-the-art implementations either fail on the presented scale or are orders of magnitude slower than our system. FLASH is capable of computing an approximate k-NN graph, from scratch, over full webspam dataset (1.3 billion nonzeros) in less than 10 seconds. Computing full k-NN graph in less than 10 seconds on webspam dataset, using brute-force (\\$n{\\textasciicircum}2D\\$), will require at least 20 TFLOPS. We hope that FLASH gets adopted in practice.","urldate":"2018-02-08TZ","journal":"arXiv:1709.01190 [cs]","author":[{"propositions":[],"lastnames":["Wang"],"firstnames":["Yiqiu"],"suffixes":[]},{"propositions":[],"lastnames":["Shrivastava"],"firstnames":["Anshumali"],"suffixes":[]},{"propositions":[],"lastnames":["Ryu"],"firstnames":["Junghee"],"suffixes":[]}],"month":"September","year":"2017","note":"arXiv: 1709.01190","keywords":"⛔ No DOI found","bibtex":"@article{wang_flash:_2017-1,\n\ttitle = {{FLASH}: {Randomized} {Algorithms} {Accelerated} over {CPU}-{GPU} for {Ultra}-{High} {Dimensional} {Similarity} {Search}},\n\tshorttitle = {{FLASH}},\n\turl = {http://arxiv.org/abs/1709.01190},\n\tabstract = {We present FLASH (\\{{\\textbackslash}bf F\\}ast \\{{\\textbackslash}bf L\\}SH \\{{\\textbackslash}bf A\\}lgorithm for \\{{\\textbackslash}bf S\\}imilarity search accelerated with \\{{\\textbackslash}bf H\\}PC (High-Performance Computing)), a similarity search system for ultra-high dimensional datasets on a single machine, which does not require similarity computation. Our system is an auspicious illustration of the power of randomized algorithms carefully tailored for high-performance computing platforms. We leverage LSH style randomized indexing procedure and combine it with several principled techniques, such as reservoir sampling, recent advances in one-pass minwise hashing, and count based estimations. The combination, while retaining sound theoretical guarantees, reduces the computational as well as parallelization overhead of our proposal. We provide CPU and hybrid CPU-GPU implementations of FLASH for replicability of our results https://github.com/RUSH-LAB/Flash. We evaluate FLASH on several real high dimensional datasets coming from different domains including text, malicious URL, click-through prediction, social networks, etc. Our experiments shed new light on the difficulties associated with datasets having several millions of dimensions. Current state-of-the-art implementations either fail on the presented scale or are orders of magnitude slower than our system. FLASH is capable of computing an approximate k-NN graph, from scratch, over full webspam dataset (1.3 billion nonzeros) in less than 10 seconds. Computing full k-NN graph in less than 10 seconds on webspam dataset, using brute-force (\\$n{\\textasciicircum}2D\\$), will require at least 20 TFLOPS. We hope that FLASH gets adopted in practice.},\n\turldate = {2018-02-08TZ},\n\tjournal = {arXiv:1709.01190 [cs]},\n\tauthor = {Wang, Yiqiu and Shrivastava, Anshumali and Ryu, Junghee},\n\tmonth = sep,\n\tyear = {2017},\n\tnote = {arXiv: 1709.01190},\n\tkeywords = {⛔ No DOI found}\n}\n\n","author_short":["Wang, Y.","Shrivastava, A.","Ryu, J."],"key":"wang_flash:_2017-1","id":"wang_flash:_2017-1","bibbaseid":"wang-shrivastava-ryu-flashrandomizedalgorithmsacceleratedovercpugpuforultrahighdimensionalsimilaritysearch-2017","role":"author","urls":{"Paper":"http://arxiv.org/abs/1709.01190"},"keyword":["⛔ No DOI found"],"downloads":0},"bibtype":"article","biburl":"https://bibbase.org/zotero/ky25","creationDate":"2019-05-11T17:47:04.275Z","downloads":0,"keywords":["⛔ no doi found"],"search_terms":["flash","randomized","algorithms","accelerated","over","cpu","gpu","ultra","high","dimensional","similarity","search","wang","shrivastava","ryu"],"title":"FLASH: Randomized Algorithms Accelerated over CPU-GPU for Ultra-High Dimensional Similarity Search","year":2017,"dataSources":["XxiQtwZYfozhQmvGR"]}