Meta团队开发的新型CRAM内存压缩方案在Linux Plumbers大会上公布，在纯读场景下相比现有ZRAM解决方案实现了显著的性能提升1。该技术由Gregory Price及其团队研发，通过创新的架构设计大幅提升了压缩内存的读取效率1。
CRAM采用私有NUMA节点架构，将压缩数据直接保留在内存中而非交换层，避免了传统交换机制带来的性能开销1。在纯读模式下，CRAM达到每秒489百万操作，相比ZRAM的110万操作每秒提升452倍1。即使在启用写入的混合负载场景下（含20%写入操作），CRAM仍保持5.4倍的性能优势1。该方案还引入了Chicken Bit机制用于防止级联故障1。
Meta's engineering team has unveiled a novel memory compression technique called CRAM that substantially outperforms existing solutions for reading compressed data from memory.1 Presented at the Linux Plumbers conference in Prague, the new method, developed by Gregory Price and his team at Meta, achieves a 452-fold performance improvement over the current ZRAM approach in read-only scenarios.1
CRAM operates by maintaining compressed data directly in memory through a private NUMA node architecture, circumventing the performance penalties associated with traditional swap operations.1 In pure read mode, CRAM delivers 489 million operations per second, compared to ZRAM's 1.1 million operations per second.1 Even when write operations are enabled—simulating a 20 percent write workload—CRAM maintains a 5.4-fold performance advantage over ZRAM.1 The system incorporates a Chicken Bit mechanism designed to prevent cascading failures.1
评论
还没有评论，欢迎留下第一条。