Title:
Efficient programming of massive-memory machines

Thumbnail Image
Author(s)
Merritt, Alexander Marshall
Authors
Advisor(s)
Gavrilovska, Ada
Advisor(s)
Editor(s)
Associated Organization(s)
Organizational Unit
Organizational Unit
Series
Supplementary to
Abstract
New and emerging memory technologies combined with enormous growths in data collection and mining within industry are giving rise to servers with massive pools of main memory — terabytes of memory, disaggregated bandwidth across tens of sockets, and hundreds of cores. But, these systems are proving difficult to program efficiently, posing scalability challenges for all layers in the software stack, specifically in managing in-memory data sets. Larger and longer-lived data sets managed by key-value stores require minimizing over- commitments of memory, but current designs trade off performance scalability and memory bloat. Furthermore, opaque operating system abstractions like virtual memory and ill-matched, non-portable interfaces used to manipulate them make the expression of semantic relationships between applications and their data difficult: sharing in-memory data sets requires careful control over internal address mappings, but mmap, ASLR, and friends remove this control. To explore and address these challenges, this dissertation is composed of two pieces: (1) We introduce and compare a new design for key-value stores, a multi-head log-structured allocator whose design makes explicit use of a machine’s configuration to support linear scalability of common read- and write-heavy access patterns. Our implementation of this design, called Nibble, is written in 4k lines of Rust. (2) Going beyond key-value stores, the second part of this dissertation introduces new general support within the operating system enabling applications to more explicitly manage and share pointer-based in-memory data: we introduce explicit control over address space allocation and layout by promoting address spaces as an explicit abstraction. Processes may associate with multiple address spaces, and threads may arbitrarily switch between them to access infinite data set sizes without encountering typical bottlenecks from legacy mmap interfaces. Our implementation of this design is in DragonFly BSD.
Sponsor
Date Issued
2017-08-24
Extent
Resource Type
Text
Resource Subtype
Dissertation
Rights Statement
Rights URI