b5c99cc908
Summary: FILE_FLAG_WRITE_THROUGH is for disabling device on-board cache in windows API, which should be disabled if user doesn't need system cache. There was a perf issue related with this, we found during memtable flush, the high percentile latency jumps significantly. During profiling, we found those high latency (P99.9) read requests got queue-jumped by write requests from memtable flush and takes 80ms or even more time to wait, even when SSD overall IO throughput is relatively low. After enabling FILE_FLAG_WRITE_THROUGH, we rerun the test found high percentile latency drops a lot without observable impact on writes. Scenario 1: 40MB/s + 40MB/s R/W compaction throughput Original | FILE_FLAG_WRITE_THROUGH | Percentage reduction --------------------------------------------------------------- P99.9 | 56.897 ms | 35.593 ms | -37.4% P99 | 3.905 ms | 3.896 ms | -2.8% Scenario 2: 14MB/s + 14MB/s R/W compaction throughput, cohosted with 100+ other rocksdb instances have manually triggered memtable flush operations (memtable is tiny), creating a lot of randomized the small file writes operations during test. Original | FILE_FLAG_WRITE_THROUGH | Percentage reduction --------------------------------------------------------------- P99.9 | 86.227 ms | 50.436 ms | -41.5% P99 | 8.415 ms | 3.356 ms | -60.1% Closes https://github.com/facebook/rocksdb/pull/3225 Differential Revision: D6624174 Pulled By: miasantreble fbshipit-source-id: 321b86aee9d74470840c70e5d0d4fa9880660a91 |
||
|---|---|---|
| .. | ||
| env_default.cc | ||
| env_win.cc | ||
| env_win.h | ||
| io_win.cc | ||
| io_win.h | ||
| port_win.cc | ||
| port_win.h | ||
| win_jemalloc.cc | ||
| win_logger.cc | ||
| win_logger.h | ||
| win_thread.cc | ||
| win_thread.h | ||
| xpress_win.cc | ||
| xpress_win.h | ||