東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁

切換: 標籤 | MARC模式 | ISBD

FindBook

Google Book

Amazon

博客來

Data-Driven Statistical Sharding for Industry-Scale Neural Recommendation.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Data-Driven Statistical Sharding for Industry-Scale Neural Recommendation./
作者:	Sethi, Geet.
面頁冊數:	1 online resource (111 pages)
附註:	Source: Dissertations Abstracts International, Volume: 84-12, Section: A.
Contained By:	Dissertations Abstracts International84-12A.
標題:	Internships. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30462697click for full text (PQDT)
ISBN:	9798379652869

Data-Driven Statistical Sharding for Industry-Scale Neural Recommendation.
Sethi, Geet.

Data-Driven Statistical Sharding for Industry-Scale Neural Recommendation. - 1 online resource (111 pages)

Source: Dissertations Abstracts International, Volume: 84-12, Section: A.

Thesis (Ph.D.)--Stanford University, 2023.

Includes bibliographical references

Deep learning based recommendation models (DLRMs) form the backbone of many internet-scale services such as web search, social media, and video streaming. Primarily composed of massive embedding tables, potentially terabytes in size, these models require immense system resources to train and the solving of the sharding problem. The sharding problem is the task of partitioning and placing the embedding table parameters throughout the target system memory topology such that training throughput is maximized.This dissertation: (1) Characterizes and derives statistics from DLRM training data which can be used to accurately and granularly predict the memory demands of individual embedding table rows; (2) Presents RecShard, a mixed-integer linear program based approach which uses these statistics to solve the sharding problem for capacity constrained single-node systems, where parameters must be placed across high-performance GPU HBM and much slower CPU DRAM; reducing accesses to the latter by orders of magnitude; and (3) Presents FlexShard, a precise row-level sharding algorithm which focuses on sharding emerging sequence-based DLRMs across multi-node GPU training clusters; leveraging these statistics to significantly reduce inter-node communication demand, the bottleneck of scale-out DLRM training.The size of industry-scale DLRMs requires sharding to be performed; however the skewed power-law nature of DLRM training data causes imprecise partitioning and placement decisions to result in imbalanced load across the system memory topology. The contributions of this dissertation provide a foundation upon which one can reason about the access patterns to fine-grained regions of DLRM memory; as well as two novel sharding techniques built upon this foundation. These techniques demonstrate significant improvements over the prior state-of-the-art on real-world production data and system deployments.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2023

Mode of access: World Wide Web

ISBN: 9798379652869Subjects--Topical Terms:

3560137
Internships.
Index Terms--Genre/Form:

542853
Electronic books.

Data-Driven Statistical Sharding for Industry-Scale Neural Recommendation.
LDR:03163nmm a2200325K 4500 001 2363202
005 20231116093823.5
006 m o d
007 cr mn ---uuuuu
008 241011s2023 xx obm 000 0 eng d
020 $a 9798379652869
035 $a (MiAaPQ)AAI30462697
035 $a (MiAaPQ)STANFORDzs617qp8476
035 $a AAI30462697
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Sethi, Geet. $3 3703952
245 1 0 $a Data-Driven Statistical Sharding for Industry-Scale Neural Recommendation.
264 0 $c 2023
300 $a 1 online resource (111 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 84-12, Section: A.
500 $a Advisor: Trippel, Caroline;Wu, Carole-Jean;Kozyrakis, Christos.
502 $a Thesis (Ph.D.)--Stanford University, 2023.
504 $a Includes bibliographical references
520 $a Deep learning based recommendation models (DLRMs) form the backbone of many internet-scale services such as web search, social media, and video streaming. Primarily composed of massive embedding tables, potentially terabytes in size, these models require immense system resources to train and the solving of the sharding problem. The sharding problem is the task of partitioning and placing the embedding table parameters throughout the target system memory topology such that training throughput is maximized.This dissertation: (1) Characterizes and derives statistics from DLRM training data which can be used to accurately and granularly predict the memory demands of individual embedding table rows; (2) Presents RecShard, a mixed-integer linear program based approach which uses these statistics to solve the sharding problem for capacity constrained single-node systems, where parameters must be placed across high-performance GPU HBM and much slower CPU DRAM; reducing accesses to the latter by orders of magnitude; and (3) Presents FlexShard, a precise row-level sharding algorithm which focuses on sharding emerging sequence-based DLRMs across multi-node GPU training clusters; leveraging these statistics to significantly reduce inter-node communication demand, the bottleneck of scale-out DLRM training.The size of industry-scale DLRMs requires sharding to be performed; however the skewed power-law nature of DLRM training data causes imprecise partitioning and placement decisions to result in imbalanced load across the system memory topology. The contributions of this dissertation provide a foundation upon which one can reason about the access patterns to fine-grained regions of DLRM memory; as well as two novel sharding techniques built upon this foundation. These techniques demonstrate significant improvements over the prior state-of-the-art on real-world production data and system deployments.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2023
538 $a Mode of access: World Wide Web
650 4 $a Internships. $3 3560137
650 4 $a Ablation. $3 3562462
650 4 $a Verbal communication. $3 3560678
650 4 $a Communication. $3 524709
655 7 $a Electronic books. $2 lcsh $3 542853
690 $a 0459
710 2 $a ProQuest Information and Learning Co. $3 783688
710 2 $a Stanford University. $3 754827
773 0 $t Dissertations Abstracts International $g 84-12A.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30462697 $z click for full text (PQDT)