Impact of Read Operations for Transparent DAX Mappings
- Typ:Bachelorarbeit
- Datum:06.08.2026
- Betreuung:
Prof. Dr. Frank Bellosa
Yussuf Khalil
- Bearbeitung:Claas Bastian Wittig
- Links:PDF
-
Abstract
Compute Express Link (CXL) is a recent interconnect standard providing new possibilities for system expansion such as extending system memory without occupying additional DIMM slot capacity. From these possibilities, hybrid SSDs emerged, which provide persistent memory semantics in addition to block-based interfaces. Hybrid SSDs implement this using a small on-device memory cache that they fill from persistent backing storage if the requested data is not in cache, introducing CPU stalls on cache misses. Previous research showed that the OS needs to actively manage the on-device cache when it is used with DAX, because DAX was not designed with the CPU stalls in mind that hybrid SSDs exhibit. Transparent DAX mappings (TDMs) address this issue while also providing unoptimized applications with the improvements that hybrid SSDs persistent memory semantics offer by transparently handling read and write requests through DAX mappings. Evaluation of TDMs focuses on handling write operations, which is why this thesis investigates what impact handling read operations can have.
To answer this question, we design tooling for observing the read and write system call usage and incurred read amplification. During analysis, we focus on detecting scenarios with read amplification, high read bandwidth, and recurring access patterns. Based on these findings, we sketch four policy ideas, which select what files or file regions TDMs handle. We evaluate the two policies that focus on alleviating system call overhead with MariaDB as the application driven by the Yahoo! Cloud Serving Benchmark, using our TDM prototype, which emulates hybrid SSD access latencies with Intel Optane DC Persistent Memory.
We observe during analysis that none of the three investigated applications induce substantial read amplification the operating system can detect. This suggests TDMs are not well suited to alleviate read amplification with these applications. We attribute the evaluation results to alleviated system call overhead. When handling read and write operations, TDMs cause up to 10 times higher benchmark throughput with read-and-write workloads and 6 times higher throughput with read-only workloads. Even though hybrid SSD emulation exhibits different characteristics than a real device, the performance improvement indicates that read operations can affect the application’s performance when TDMs handle them.BibTex:
@bachelorthesis{wittig26TransparentDAXMappings,
author = {Claas Bastian Wittig},
title = {Impact of Read Operations for Transparent DAX Mappings},
type = {Bachelor Thesis},
year = 2026,
month = aug# "06",
school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
}