Beyond PMEM, Libfabric & x86_64

While DAOS was originally developed to store all metadata in Optane Persistent Memory on modern Intel CPUs with client/server communications over libfabric, the architecture and support matrix have evolved over time to support a broader set of configurations as the open-source community expands.

The metadata-on-SSD effort was initiated in early 2022 to allow the DAOS engine to store metadata on SSD. This capability was landed as a feature preview in DAOS v2.4.0 and will be ready for prime time in the upcoming 2.6 DAOS version. Michael Hennecke, from Intel, shared very encouraging results with the metadata-on-SSD feature at the IXPUG conference. It is worth noting that the DAOS engine will still keep the ability to store metadata in persistent memory for both existing customers and for future CXL2.0 SSDs supporting persistent memory once generally available.

With the 2.4.0 version, the UCX communication middleware is now officially supported by DAOS in addition to libfabric. An extra package is now shipped with the DAOS releases providing the UCX Mercury plugin. The support of UCX allows DAOS to use some features on Mellanox hardware that were not accessible so far like the DC_X and UD_X transports. This should also allow DAOS to eventually leverage multi-rails and GPGPU optimizations available in UCX.

Finally, the DAOS community has also been collaborating with Linaro to port and test DAOS on the ARM64 architecture. Some ARM64 VMs have been integrated into the CI DAOS pipeline to test both the DAOS client library and engine on ARM64. Riken and Intel presented an evaluation of DAOS on ARM64 at the ARM HPC User Group. An IO500 submission with DAOS running on ARM64 was also done by Riken last November.

It is a very exciting time for the DAOS community!