3 ms·
On at least one drive I saw, the flush command was instead interpreted as a barrier to commands being committed to the log in controller DRAM, which could cut i
by monocasa 5y ago
On at least one drive I saw, the flush command was instead interpreted as a barrier to commands being committed to the log in controller DRAM, which could cut into parallelization, and therefore throughput, looking like a latency spike but not a flush out of the cache.
- marcan_42 5y agoMy test is single threaded, and thus has no parallelism to begin with.
- monocasa 5y agoThe drive controller is internally parallel. The write is just a job queue submission, so the next write hits while it's still processing previous requests.
- marcan_42 5y agoPeople have tested this stuff on storage devices with torture tests. Can you point at an example of a modern (directly attached) NVMe drive from a reputable vendor that cheats at this? FWIW, macOS also has F_BARRIERFSYNC, which is still much slower than full syncs on the competition.