Meta engineer's CRAM project speeds up Linux memory compression up to 452x
Gregory Price and a team at Meta have built CRAM, a new Linux memory-compression technique that keeps compressed data directly in RAM instead of routing it through the swap layer like existing tools zswap and ZRAM. CRAM uses a private NUMA node to let Linux manage compressed memory with its normal memory semantics, including migration and ballooning, and includes a mechanism called the 'Chicken Bit' to pause CRAM during allocation management. Early benchmarks cited show read performance gains of up to 452 times over ZRAM.