Jens, here are a few latency cleanup patches. The main improvement is to avoid recording issue_time twice for the ioengines that record issue_time themselves when their commit hook is called. Recording issue_time twice underestimates clat since the first issue_time value is used for slat and then the later second issue_time is used when calculating clat. I'm not able to test the rdma and librpma ioengines but the fix for libaio and io_uring appears to work properly. Vincent Fu (5): ioengines: add helper for trims with async ioengines ioengines: don't record issue_time if ioengines already do it HOWTO: improve description of latency measures ioengines: update last_issue if we set issue_time ioengines: clean up latency accounting for 3 ioengines HOWTO.rst | 29 ++++++++++++++++----------- engines/io_uring.c | 13 ++++++++++-- engines/libaio.c | 9 ++++++++- engines/librpma_apm.c | 2 +- engines/librpma_fio.c | 9 ++++++++- engines/librpma_gpspm.c | 2 +- engines/rdma.c | 9 ++++++++- ioengines.c | 44 ++++++++++++++++++++++++----------------- ioengines.h | 2 ++ 9 files changed, 83 insertions(+), 36 deletions(-) -- 2.25.1
