Crash consistency for offsite metadata during DRAM harvesting

Establish a mechanism to guarantee crash consistency of a borrower's flash translation layer (FTL) mapping table when the mapping entries are cached in a lender SSD's DRAM as part of DRAM harvesting over Compute eXpress Link (CXL), ensuring correctness of the offsite metadata despite failures or interruptions.

Background

XBOF enables inter-SSD DRAM harvesting over CXL, allowing a borrower SSD to cache portions of its FTL mapping table in a lender SSD’s DRAM to reduce miss penalties and improve performance. This places critical metadata offsite, outside the borrower’s local power-loss protection domain.

Enterprise SSDs typically provide power loss protection (PLP) by flushing dirty data and metadata from caches and DRAM to flash upon power failure. When the borrower’s mapping table entries reside only in a lender’s DRAM, these entries may not be flushable by the borrower on power loss or device removal, raising crash consistency and data recovery challenges.

References

While this idea can be facilitated by the cache-coherent capability of CXL, there still remains an open question in practice, that is, how to guarantee crash consistency of the offsite metadata (i.e., the mapping table stored in the lender's DRAM).

— XBOF: A Cost-Efficient CXL JBOF with Inter-SSD Compute Resource Sharing  (2509.10251 - Yi et al., 12 Sep 2025) in Section 4.5 Persistent DRAM Harvesting, Crash consistency