Transactional Updates for Virtual-Memory Objects on (Block|Byte)-Addressable Persistent Storage
- Typ der Arbeit: Masterarbeit
- Status der Arbeit: laufend
- Projekte: ParPerOS
- Betreuer: Alexander Halbuer, Daniel Lohmann
- Bearbeiter: Oscar Tylkowski

Exploring the cost of persistent snapshots on storage devices with different guarantees. [generated with AI].
Context
Intel Optane promised to make all system state trivially durable by making volatile main memory persistent. Ultimately, it failed to achieve a breakthrough because the required hardware was rare, its strong persistence guarantees (eADR) were difficult to provide, crash-consistent algorithms were complex, and production ended early. There may have been other contributing factors. Nevertheless, although no killer application emerged to sustain the Optane business during its short period of availability, the question of how to bridge the gap between fast, volatile working memory and block-addressable persistent storage remains.
More recent, the Compute Express Link (CXL) interconnect has emerged, allowing device memory to be mapped into the host's physical address space and accessed through CPU loads and stores via the CXL.mem protocol. This capability gives rise to a new class of SSDs: memory-semantic CXL-SSDs, which expose their entire flash capacity through CXL.mem by mapping it into the host's physical address space. To serve byte-granular loads and stores, these devices place an on-device DRAM cache in front of the flash. Via global persistent flush (GPF) these devices also provide persistency of data as soon as the store hits the host's CPU cache.
However, this hardware is still difficult to obtain. This raises the question: what are the costs of implementing semantically persistent memory objects using conventional, high-IOPS NVMe SSDs?
Problem
The Morsel abstraction defines virtual-memory objects as subtrees of the page-table tree. The page tables included in an object are part of that object and fully describe its layout without external metadata. This approach makes it straightforward to persist such objects on Intel Optane with strong guarantees (eADR, where CPU caches are part of the persistence domain). Changes to the page tables can simply be made atomic, since visibility is equivalent to persistence in this model. However, this makes the layout of Morsel objects persistent; it does not automatically persist all user data. Applications would still need to use appropriate crash-consistent algorithms.
Without strong eADR guarantees, explicit cache flushes are also required to make data truly persistent. This adds steps to the workflow, degrades performance, and breaks the visibility equals persistence paradigm. Although this may be feasible for the operating system that implements the memory abstraction, it makes adopting persistence features in applications more complicated. Therefore, the operating system should provide a suitable abstraction for data transactions.
Morsels already offer an O(1) snapshot mechanism based on copy-on-write. Could this mechanism be used to track changes between two synchronization points and atomically propagate those changes to the underlying persistent medium?
Goal
The goal of this thesis is to explore the costs of object persistence under different guarantees:
- Intel Optane (with or without eADR): all main memory is persistent, caches might be persistent.
- Memory-semantic SSD: all device memory is semantically persistent, but the device internally uses a tiered memory hierarchy.
- Conventional NVMe SSD: high IOPS, but block-level access only.
More specific, this thesis aims to answer the following questions:
-
How to efficiently propagate changed data to an underlying block-storage device? And how to deal with the persistence guarantees that an SSD provides to be resistant to inconsistent intermediate states that can be caused by unexpected power losses.
-
What overhead do we incur by ensuring atomic updates? Providing atomic updates certainly leads to management overhead. How costly is that guarantee?
-
Do CXL-SSDs provide a lower overhead way to persist morsels? CXL-SSDs provide byte-granular access to persistent storage while still providing atomics. Does this capability allow for a lower overhead implementation of atomic morsel updates?
-
How do CXL-SSDs compare to Optane NVDIMMs in providing morsel persistence? Intel Optane (now discontinued) chips provide persistent RAM (NVRAM). Does the morsels implementation on Intel Optane differ to the implementation on CXL-SSDs? If yes, how do they differ and why?
Topics: virtual memory, MMU, NVMe SSD, CXL
