Cephalocon 2024

A cyclic redundancy check (CRC) is an error-detecting code widely used in digital networks and storage devices to identify accidental changes in digital data. Ceph employs CRC checksums to ensure data correctness during network transmission and to maintain data integrity in Bluestore. For RBD 4K random write and random read workloads, we observed that the CRC32C operation size primarily ranges from 4 to 292 bytes. Specifically, for 4K random writes, 80% of the operations are smaller than 292 bytes, while for 4K random reads, 83% are smaller than 254 bytes. The most common sizes are 28 bytes and 41 bytes, accounting for nearly 40%; these correspond to 'preamble_block_t' and 'ceph_msg_header2', respectively. Given the characteristics of the CRC32C operation, we optimized it for smaller sizes by using inline assembly code to unfold the CRC32C function 'ceph_crc32c_aarch64'. Additionally, we optimized the encoding procedure to minimize the occurrence of small-sized data segments. We tested these optimizations on the RBD 4K random write and random read workloads, resulting in a 2% improvement in throughput.