Moses: Heap Partitioning for Semantic Data Tiering
Open Access
- 3 November 2024
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 25-32
Abstract
The adoption of emerging memory technologies has long been impeded by the lack of interfaces communicating memory placement semantics not only between applications, system software, and diagnostics tools but also between multiple tenants in a system. With the support of NUMA and memory tiering in Linux, applications can explicitly place pages based on domain knowledge. Yet, the internal reasoning behind placing a specific allocation into a specific memory resource remain largely hidden from the outside world. At the application level, different allocations exhibit different longevity, which is very prominent in the domain of In-Memory Database Systems (IMDBS), where short-lived temporary allocations required by database operators are distinct from long-lived table data. Consequently, access patterns or non-functional requirements on memory resources like bandwidth and latency vary considerably. To allow for more fine-grained control over the allocation behavior of an application, we propose to partition the heap into memory regions with associated semantics, that are exposed in a hierarchical file system. Our prototype Moses enables domain experts (i.e., programmers) to express and refine memory placement requirements. We demonstrate how this can be applied to the in-memory database system Hyrise. Using Moses, we can achieve semantic data tiering, transparent to extrinsic operating system tools. We subsequently reduce the resident set size of the table data of the TPC-H benchmark by nearly 65 % at a moderate execution time overhead of less than 5 %.Keywords
Funding Information
- Bundesministerium für Forschung, Technologie und Raumfahrt (01IS22092)
- Deutsche Forschungsgemeinschaft (502228341)
This publication has 11 references indexed in Scilit:
- Persistent Memory Research in the Post-Optane Era2023
- MEMTIS: Efficient Memory Tiering with Dynamic Page Classification and Page Size Determination2023
- Carbon-Aware Memory Placement2023
- Pond: CXL-Based Memory Pooling Systems for Cloud Platforms2023
- Improving the accessibility of NUMA‐aware C++ application development based on the PGASUS frameworkConcurrency and Computation: Practice and Experience, 2022
- Nimble Page Management for Tiered Memory Systems2019
- MemBrain: Automated Application Guidance for Hybrid Memory Systems2018
- Adaptive NUMA-aware data placement and task scheduling for analytical workloads in main-memory column-storesProceedings of the VLDB Endowment, 2016
- The flame graphCommunications of the ACM, 2016
- Traffic managementACM SIGPLAN Notices, 2013