Bug #37673
Updated by Kefu Chai over 5 years ago
* 6 osd cluster
* separated cluster network on eth0, and public network on eth1.
* a rados client accessing from public network
# inject 2 seconds delay using tc to eth0. like @tc qdisc add dev eth0 root netem delay 2000ms@
# put a 128 MB file to into 3 replica pool
slow requests are observed. quite a few of them had over 20 seconds' delay between "initiated" and "queued_for_pg"
<pre>
{
"time": "2018-12-17 05:36:11.977528",
"event": "initiated"
},
{
"time": "2018-12-17 05:36:47.991316",
"event": "queued_for_pg"
},
</pre>
found on luminous. need to reproduce this on master.