Fix bad latency reporting for rated IO jobs
authorJens Axboe <jaxboe@fusionio.com>
Tue, 30 Aug 2011 21:34:14 +0000 (15:34 -0600)
committerJens Axboe <jaxboe@fusionio.com>
Tue, 30 Aug 2011 21:34:14 +0000 (15:34 -0600)
If we are going to sleep to satisfy a given rate limit, then
ensure that we flush pending IO first. Otherwise they end up
including the sleep time in the calculated completion latencies.

Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
io_u.c

diff --git a/io_u.c b/io_u.c
index 6a53bda..7709df5 100644 (file)
--- a/io_u.c
+++ b/io_u.c
@@ -486,6 +486,16 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
        } else
                usec = td->rate_pending_usleep[ddir];
 
+       /*
+        * We are going to sleep, ensure that we flush anything pending as
+        * not to skew our latency numbers
+        */
+       if (td->cur_depth) {
+               int fio_unused ret;
+
+               ret = io_u_queued_complete(td, td->cur_depth, NULL);
+       }
+
        fio_gettime(&t, NULL);
        usec_sleep(td, usec);
        usec = utime_since_now(&t);