Abstract
We introduce manylogs, a simple and novel concept of logging that deploys many scattered logs on disk such that small random writes can be appended into any log near the current disk head position (e.g., the location of last large I/O). The benefit is two-fold: the small writes attain fast durability while the large I/Os still sustain large bandwidth. Manylogs also inspire a new principle: decoupling of durability and location constraints. For example, we can put journal blocks in any scattered log as they only need durability but not location constraints (i.e., eventually journal blocks will be freed). We can also allow applications to specify which files that require only durability but not location constraints (e.g., application commit-log files). We show the power of manylogs for Conventional-Magnetic Recording file system (MLFS) and block-level (MLB) layers and also Shingled-Magnetic Recording (MLSMR) layer. With microand macro-benchmarks, compared to Linux ext3, MLFS provides up to 15x (3.7x on average) bandwidth improvement and up to 22x (5.7x) faster sync latency. With real-world traces, MLB and MLSMR deliver 1.9x and 1.3x I/O latency speed-up on average respectively.