On 3/6/19 2:41 AM, 韦皓诚 wrote:
Hi~
We use gdbpmp to profile wall time in osd and find that
pthread_cond_timedwait cost most of the real time(65.8%) in tp_osd_tp.
This means that the queue is empty in most of the time. But when we
use perfcount to analysis data reading, we find that the latency
between enqueue and dequeue is not negligible( about 40% in
op_r_latency). That may mean the queue is busy. So what cause the
contradiction between them?
Thank you~
Hi,
Can you post the gdbpmp trace and perfcount data? (feel free to send to
me if you don't want to post publicly). I'm not sure I have an answer
for you, but I'd start out by looking at what's happening the other 36%
of the time in tp_osd_tp.
Mark
