The word “Fabrics” is perhaps the most used or abused word in Computing today. A Fabric is, as its name implies, a web of multiple devices in a cluster using some physical level connectivity with transport and protocol layers built on top of it. Computing, in general, and High Performance Computing in particular have moved over the years from proprietary cluster interconnects to more popularly, InfiniBand and Ethernet.

However, the communication fabric closer to the CPUs or Accelerators (including GPUs) is where the focus has been of late with the AI era. Compute Express Link (CXL) is a memory semantic fabric that has been developed to solve multiple challenges in computing related to capacity and bandwidth of memory, as well as the need to disaggregate memory. When it comes to connecting GPUs with each other, NVLink has become the most popular communication fabric. However, there have been recent advancements in this space with the formation of the Ultra Accelerator Link, or UALink, consortium to connect accelerators and allow sharing of memory between them and CPUs. Ethernet is further being enhanced by the Ultra Ethernet Consortium (UEC) to build a low-latency, lossless variant of Ethernet.

The Fabrics for HPC/AI workshop will be a half-day workshop with the aim of exploring the novel research ideas around this very rich space of fabrics and the research happening in operating systems, virtualization and manageability in related areas.

Papers are solicited from the areas, including, but not limited to:

Hardware Architectures

• Memory Expansion and Pooling (via technologies like CXL)
• Ethernet challenges as a High Performance Fabric
• Fabrics for next gen AI/ML workloads

Software Architetcures

• CXL emulation
• Operating system support for Tiered Memory
• Virtualization for Fabrics

Applications and Use-Cases

• Benchmarking Applications
• Workload characterization
• Analysis and Profiling of Fabric Performance

Control Plane Software for Management

• Fabric Management e.g.  RedFish (DMTF)/OFA (Open Fabrics)
• In-band and OOB Management of Fabric devices (CXL/Ethernet/Accelerators)

Submission Guidelines

General guidelines: Manuscripts submitted to FABRICS 2024 should not have been previously published or be under review for a different workshop, conference or journal. Abstracts should contain no more than 300 words and must be submitted along with the paper by the paper submission deadline. The title and abstract submitted by this deadline should have sufficient detail and not just be a placeholder. Submissions should have the final list of authors, as changes may not be feasible later. Submitted papers must represent original unpublished research that is not currently under review for any other conference or journal. Papers not following these guidelines will be rejected without review and further action may be taken, including (but not limited to) notifications sent to the heads of the institutions of the authors. 

Length of submission: Submitted manuscripts can be full papers or short papers.

• Full papers may not exceed six (6) single-spaced double-column pages using 10-point size font on 8.5×11 inch pages (IEEE conference style) plus one extra page for references. 
• Short papers may not exceed four (4) plus one extra page for references in the same format. 

Single blind policy: All submitted manuscripts will be reviewed by the Program Committee under a single-blind review process, so the submitted paper should NOT list any authors or their affiliations.

Committees

Organizing committee

• Mohan Parthasarathy (HPE)
• Sunita Jain (AMD)
• Ajay Joshi (Micron)
• Dr Badrinath Ramamurthy (IIITB)

Contact

All questions about submissions should be emailed to Parthasarathy, Mohan (HPE) <pmonty@gmail.com>