AiSIO Technical Report#
Note
For more recent information while this report is being updated, see the presentations of AiSIO, “AiSIO: Orchestrating Storage I/O Across CPUs and Accelerators”:
Warning
This report is outdated, about nine months behind, and is being reworked. Since it was written, significant changes have made it into the software ecosystem AiSIO builds on, and its experiments and integrations have evolved with them. Read it with that in mind:
The experiments do not exercise HOMI. Their results describe the I/O paths without the control plane the architecture is built around.
The architecture diagrams and some names predate the current components.
The experiment that exercises the whole stack through HOMI today is
tasks/bench_io_modes.yaml, with xnvmeperf as the client, in four modes
taken in order: CPU-initiated I/O, CPU-initiated P2P I/O, accelerator-assisted
P2P I/O and accelerator-initiated P2P I/O, as Terminology defines
them. The README walks through running it.
Where the report and the repository disagree, the repository is right: the
task files in tasks/ and the configurations in configs/ are what runs.
Section This Report lists the known shortcomings and what is
being done about them.
- AiSIO Technical Report
- Abstract
- Introduction
- Background
- Architecture
- Related Work
- Implementation
- Environments
- Experimental Framework
- Experiments
- Conclusion
- Future Work
- References