Okay, that sounds like something else is happening then. You might try turning up the debug to 20 and seeing what is happening during those 4 seconds, as that's definitely more time than it should take. -Greg On Mon, Nov 26, 2018 at 10:37 AM Marvin Zhang <fanzier@xxxxxxxxx> wrote: > > I tested it in 12.2.8, and the debug_mds log level is 10/10 > I don't config the mds_cache_memory_limit, I think it will be 1GB. > Here is the log(I set MDS log as 10) > .... > 2018-11-23 02:01:02.995067 7f40535fe700 7 src/mds/MDCache.cc:6517 > mds.0.cache trim bytes_used=1GB limit=1GB reservation=0.05% count=0 > 2018-11-23 02:01:02.995084 7f40535fe700 7 src/mds/MDCache.cc:6456 > mds.0.cache trim_lru trimming 0 items from LRU size=694884 mid=472796 > pintail=4942 pinned=19460 > 2018-11-23 02:01:02.995409 7f404e7fb700 10 src/mds/MDSContext.cc:51 > MDSIOContextBase::complete: 21C_IO_Dir_OMAP_Fetched > 2018-11-23 02:01:04.207192 7f40535fe700 7 src/mds/MDCache.cc:6502 > mds.0.cache trim_lru trimmed 262781 items > 2018-11-23 02:01:04.207211 7f40535fe700 10 src/mds/MDCache.cc:7426 > mds.0.cache trim_client_leases > .... > > On Mon, Nov 26, 2018 at 10:43 PM Gregory Farnum <gfarnum@xxxxxxxxxx> wrote: > > > > On Mon, Nov 26, 2018 at 7:00 AM Marvin Zhang <fanzier@xxxxxxxxx> wrote: > > > > > > Hi CephFs Experts, > > > I found that MDCache::trim_lru() will take more than 2s sometime, is > > > it as expected? And it seems like to hold mds_lock at that time, which > > > cause other client request can't be processed. > > > Thanks, > > > Marvin > > > > It's definitely not supposed to, but we've started noticing some > > performance issues with extremely large MDS cache sizes. How much > > memory did you give the cache, and what is your system configuration?
