Meta’s CRAM keeps compressed Linux pages in RAM and reports big speed gains

The approach is still research, but it could change how Linux systems handle compressed memory under load.

Meta researcher Gregory Price and his team have presented CRAM, a Linux memory-compression approach that keeps compressed pages in RAM instead of using swap, and the idea could significantly change how compressed memory is implemented on servers and other Linux machines. The project claims very large read-side gains, with the presentation citing up to 452x better performance than ZRAM and a worst-case rate of 489 million operations per second versus 1.1 million for ZRAM. When writes are included, the advantage shrinks but still remains substantial, with the article describing a 5.4x speedup in a test where 20% of operations were writes. The design uses a private NUMA node and Linux’s existing memory-management behavior, while a “Chicken Bit” is meant to prevent cascading allocation failures while CRAM is under pressure. Some core questions are still unresolved, especially how to account for logical capacity when physical data is compressed, so the work is still presented as ongoing research rather than a finished kernel feature.