자료유형  

 학위논문

Control Number  

0017162967

International Standard Book Number  

9798384344568

Dewey Decimal Classification Number  

001

Main Entry-Personal Name  

Raghavan, Deepti.

Publication, Distribution, etc. (Imprint  

[S.l.] : Stanford University., 2024

Publication, Distribution, etc. (Imprint  

Ann Arbor : ProQuest Dissertations & Theses, 2024

Physical Description  

137 p.

General Note  

Source: Dissertations Abstracts International, Volume: 86-04, Section: B.

General Note  

Advisor: Levis, Philip;Zaharia, Matei.

Dissertation Note  

Thesis (Ph.D.)--Stanford University, 2024.

Summary, Etc.  

요약Software serialization is critical for many datacenter applications, but serialization is costly in today's datacenters. Datacenter networks have become at least 20x faster in the last 15 years, while CPU performance has stagnated over the same time period. Serialization is fundamentally about data movement: serialization libraries coalesce and flatten in-memory data structures into a single transmittable buffer. Unfortunately, the memory copies and allocations required for data movement cause application performance to fall behind modern network capabilities.This dissertation describes Cornflakes, an efficient serialization and networking stack for datacenter applications, based on three key ideas. First, Cornflakes partially offloads serialization to commodity hardware by using NIC scatter-gather. Cornflakes uses NIC scatter-gather to perform serialization's memory copies, but keeps the rest of serialization in software. Second, because zero-copy is sometimes slower or impossible, Cornflakes is hybrid and transparently chooses between hardware and software copies. Zero-copy requires memory coordination, which adds bookkeeping overhead that can sometimes outweigh the benefits of removing copies. Third, the serialization and networking APIs in Cornflakes are tightly co-designed. This minimizes the overhead of translating between data representations in the stack.Cornflakes shows that it is possible to accelerate datacenter serialization compared to existing software approaches, without new hardware. Cornflakes supports a UDP networking interface on top of Mellanox and Intel NICs and partially supports a TCP networking interface as well. In a custom key-value store application serving a Twitter-like cache trace, Cornflakes achieves 15.4% higher throughput than prior general-purpose serialization libraries. When integrated into Redis, Cornflakes achieves 8.8% higher throughput compared to Redis serialization, on the same trace.

Subject Added Entry-Topical Term  

Software.

Subject Added Entry-Topical Term  

Computer science.

Added Entry-Corporate Name  

Stanford University.

Host Item Entry  

Dissertations Abstracts International. 86-04B.

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:655428