On Thu, Mar 29, 2018 at 11:55 PM, 陶冬冬 <tdd21151186@xxxxxxxxx> wrote: > > I have verified that by adding some debug log, and found it’s truly caused by long wait in mqueue for rejoin. > but, it seems this kind laggy does no harm ? > It's harmful. I think we should move long work into worker thread. So it does not block message dispatch >> 在 2018年3月29日，上午10:08，Yan, Zheng <ukernel@xxxxxxxxx> 写道： >> >> #2 0x00005555557134eb in MDCache::process_imported_caps >> (this=this@entry=0x5555566a9000) at >> /home/zhyan/Ceph/ceph-2/src/mds/MDCache.cc:5397 >> >> #3 0x00005555557149a3 in MDCache::rejoin_start >> (this=this@entry=0x5555566a9000, rejoin_done_=0x5555566b5c80) at >> /home/zhyan/Ceph/ceph-2/src/mds/MDCache.cc:3950 >> >> #4 0x0000555555663f93 in MDSRank::rejoin_start >> (this=this@entry=0x5555566a8000) at >> /home/zhyan/Ceph/ceph-2/src/mds/MDSRank.cc:1425 >> >> #5 0x00005555556747e3 in MDSRankDispatcher::handle_mds_map >> (this=0x5555566a8000, m=m@entry=0x555556527400, >> oldmap=oldmap@entry=0x5555564b2dc0) at >> /home/zhyan/Ceph/ceph-2/src/mds/MDSRank.cc:1683 >> >> #6 0x0000555555655a1e in MDSDaemon::handle_mds_map >> (this=this@entry=0x55555651e000, m=m@entry=0x555556527400) at >> /home/zhyan/Ceph/ceph-2/src/mds/MDSDaemon.cc:1000 >> >> #7 0x00005555556582fc in MDSDaemon::handle_core_message >> (this=this@entry=0x55555651e000, m=m@entry=0x555556527400) at >> /home/zhyan/Ceph/ceph-2/src/mds/MDSDaemon.cc:1204 >> >> #8 0x0000555555658aa3 in MDSDaemon::ms_dispatch (this=0x55555651e000, >> m=0x555556527400) at /home/zhyan/Ceph/ceph-2/src/mds/MDSDaemon.cc:1158 >> >> #9 0x00007fffef283eda in Messenger::ms_deliver_dispatch >> (m=0x555556527400, this=0x55555651a000) at >> /home/zhyan/Ceph/ceph-2/src/msg/Messenger.h:667 >> >> #10 DispatchQueue::entry (this=0x55555651a1a8) at >> /home/zhyan/Ceph/ceph-2/src/msg/DispatchQueue.cc:201 >> >> #11 0x00007fffef31dcfd in DispatchQueue::DispatchThread::entry >> (this=<optimized out>) at >> /home/zhyan/Ceph/ceph-2/src/msg/DispatchQueue.h:101 >> >> #12 0x00007fffed4f336d in start_thread () from /lib64/libpthread.so.0 >> >> #13 0x00007fffec547b4f in clone () from /lib64/libc.so.6 >> >> >> I think the real cause is that MDCache::process_imported_caps() takes >> too long. prevent messenger from dispatch other message. >> >> On Thu, Mar 29, 2018 at 9:37 AM, Yan, Zheng <ukernel@xxxxxxxxx> wrote: >>> On Thu, Mar 29, 2018 at 9:09 AM, 陶冬冬 <tdd21151186@xxxxxxxxx> wrote: >>>> Patrick, we are using 32G memory for the mds. >>>> Zheng, calling mds->heartbeat_reset() could make the healthy check pass so >>>> that monitor won’t kick it out. >>>> more frustrating me is about the laggy issue, from the monitor log, i can >>>> actually see the MDSBeacon are sent without delay. >>>> but about 50 seconds later, mds start handling that MDSBeacon message. >>>> so i’m wondering would that possible the message stayed in the mqueue for >>>> that long time (if the previous message is MDSMap with rejoin state, ant >>>> that rejoin take long time) >>>> (meantime, i will do some more investigating about this issue) >>>> >>> >>> If it's really caused by long wait in mqueue. we should limit >>> concurrent open_ino >>> started by MDCache::process_imported_caps() >>> >>> >>>> 在 2018年3月29日，上午8:10，Yan, Zheng <ukernel@xxxxxxxxx> 写道： >>>> >>>> On Wed, Mar 28, 2018 at 11:14 PM, 陶冬冬 <tdd21151186@xxxxxxxxx> wrote: >>>> >>>> Hi Zheng & Patrick, >>>> >>>> we are using v12.2.2. >>>> Recently we’ve met an mds laggy issue (significantly, about 50 seconds) >>>> i’ve traced the monitor and mds log and found that the MMDSBeacon message >>>> was actually sent to mds 50 seconds ago. >>>> so, looks like monitor isn’t laggy , and more worse is that i also found >>>> that the mds’s health check is failed and eventually monitor >>>> just kicked out this mds and make it respawn. >>>> by the way, this happened at rejoin phase. >>>> >>>> Following is my analysis : >>>> The mds health check failure is because the mds tick thread could not get >>>> the mds_lock due to rejoin. (i found rejoin has many missing ino needed to >>>> fetch) >>>> and this leads the mqueue of the DispatchQueue consumed by Dispatcher got >>>> very slow, eventually make MMDSBeacon in mqueue got dispatched after a big >>>> delay. >>>> >>>> >>>> how about calling mds->heartbeat_reset() in the loop that fetch inodes >>>> >>>> >>>> >>>> i want to know if my analysis make sense to you ? if so, i’m wondering can >>>> we make MMDSBeacon fast dispatch. >>>> >>>> Regards, >>>> Dongdong >>>> >>>> > -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
