Design and Implementation of a Scalable Naming Service in Shared Memory

Offer Description

>>> Complete PhD offer in PDF <<<

Context
The CXL [1] standard will profoundly impact resource management in data centers. CXL defines a cache coherency domain that not only includes system memory and CPUs, but also PCIe devices. It opens the way to fully disaggregated data centers, as the PCIe buses of a cluster of machines can be connected through a CXL fabric [2, 3] which allows the loads and stores emitted by a processor to be transparently routed to the memory of the receiver through a cluster-scale cache-coherency protocol. At the software level, far memory located in another machine can be accessed as transparently as local memory: a simple statement such as a = 42 can be routed seamlessly to any memory of any machine connected to the CXL fabric.

In this context, the traditional design in which independent storage nodes are accessed by compute nodes over a network becomes inadequate. This approach has been dominant in the past due to its ability to independently scale computation and storage. However, it incurs significant costs from data exchange and transformation which can be avoided by taking advantage of the efficient cluster-scale cache-coherency protocol provided by CXL.

PhD Topic
In this project, we aim to reimagine the architecture of cloud applications in the CXL era. Our approach decouples memory from processes, enabling global memory sharing across processes, similar to how threads share memory in multi-threaded applications. However, unlike the multi-threaded model, memory objects in this design can persist beyond the lifespan of individual processes, acting as long-term storage for ephemeral processes that are launched on demand to serve clients or handle large-scale data analytics. Since any process can directly access the global memory, the architecture avoids the high cost of transforming data when it is exchanged between the processes.

Central to this architecture is a naming service, which, at a high level, is reminiscent of a classical file system. This naming service must make it possible for each process to retrieve objects produced by other processes. It must reside in shared memory and scale to thousands of processes. As a PhD student, you will study how such a naming service can be designed, leveraging low-level hardware features such as virtualization extensions to alleviate bottlenecks.

References
[1] Debendra Das Sharma. Compute Express Link (CXL): enabling heterogeneous data-centric computing with heterogeneous memory hierarchy. IEEE Micro, 2022.

[2] Donghyun Gouk, Sangwon Lee, Miryeong Kwon, and Myoungsoo Jung. Direct access, high-performance memory disaggregation with DirectCXL. In Proceedings of the USENIX Annual Technical Conference, USENIX ATC’22, 2022.

[3] Huaicheng Li, Daniel S. Berger, Lisa Hsu, Daniel Ernst, Pantea Zardoshti, Stanko Novakovic, Monish Shah, Samir Rajadnya, Scott Lee, Ishwar Agarwal, Mark D. Hill, Marcus Fontoura, and Ricardo Bianchini. Pond: CXL-based memory pooling systems for cloud platforms. In Proceedings of the conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS’23, 2023.

Where to apply

Website https://jobrxiv.org/job/inria-27778-phd-position-design-and-implementation-of-a…

Requirements

Additional Information

Work Location(s)

Number of offers available 1

Company/Institute Inria

Country France

Contact

Website https://inria.fr/