<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=us-ascii">
</head>
<body>
Neil,<br>
<br>
My understanding is marking the inode cache reclaimable would make Lustre unusual/unique among Linux file systems. Is that incorrect?<br>
<br>
- Patrick
<hr style="display:inline-block;width:98%" tabindex="-1">
<div id="divRplyFwdMsg" dir="ltr"><font face="Calibri, sans-serif" style="font-size:11pt" color="#000000"><b>From:</b> lustre-discuss <lustre-discuss-bounces@lists.lustre.org> on behalf of NeilBrown <neilb@suse.com><br>
<b>Sent:</b> Monday, April 29, 2019 8:53:43 PM<br>
<b>To:</b> Jacek Tomaka<br>
<b>Cc:</b> lustre-discuss@lists.lustre.org<br>
<b>Subject:</b> Re: [lustre-discuss] Lustre client memory and MemoryAvailable</font>
<div> </div>
</div>
<div class="BodyFragment"><font size="2"><span style="font-size:11pt;">
<div class="PlainText">On Mon, Apr 29 2019, Jacek Tomaka wrote:<br>
<br>
>> so lustre_inode_cache is the real culprit when signal_cache appears to<br>
>> be large.<br>
>> This cache is slaved on the common inode cache, so there should be one<br>
>> entry for each lustre inode that is in memory.<br>
>> These inodes should get pruned when they've been inactive for a while.<br>
><br>
> What triggers the prunning?<br>
><br>
<br>
Memory pressure.<br>
The approx approach is try to free some unused pages and about 1/2000th of<br>
the entries in each slab. Then if that hasn't made enough space<br>
available, try again.<br>
<br>
>>If you look in /proc/sys/fs/inode-nr there should be two numbers:<br>
>> The first is the total number of in-memory inodes for all filesystems.<br>
>> The second is the number of "unused" inodes.<br>
>><br>
>> When you write "3" to drop_caches, the second number should drop down to<br>
>> nearly zero (I get 95 on my desktop, down from 6524).<br>
><br>
> Ok, that is useful to know but echoing 3 to drop_cache or generating memory<br>
> pressure<br>
> clears most of the signal_cache (inode) as well as other lustre objects, so<br>
> this is working fine.<br>
<br>
Oh good, I hadn't remembered clearly what the issue was.<br>
<br>
><br>
> The issue that remains is that they are marked as SUnreclaim vs<br>
> SReclaimable.<br>
<br>
Yes, I think lustre_inode_cache should certainly be flagged as<br>
SLAB_RECLAIM_ACCOUNT.<br>
If the SReclaimable value is too small (and there aren't many<br>
reclaimable pagecache pages), vmscan can decide not to bother. This is<br>
probably a fairly small risk but it is possible that the missing<br>
SLAB_RECLAIM_ACCOUNT flag can result in memory not being reclaimed when<br>
it could be.<br>
<br>
Thanks,<br>
NeilBrown<br>
<br>
<br>
> So i do not think there is a memory leak per se.<br>
><br>
> Regards.<br>
> Jacek Tomaka<br>
><br>
> On Mon, Apr 29, 2019 at 1:39 PM NeilBrown <neilb@suse.com> wrote:<br>
><br>
>><br>
>> Thanks Jacek,<br>
>> so lustre_inode_cache is the real culprit when signal_cache appears to<br>
>> be large.<br>
>> This cache is slaved on the common inode cache, so there should be one<br>
>> entry for each lustre inode that is in memory.<br>
>> These inodes should get pruned when they've been inactive for a while.<br>
>><br>
>> If you look in /proc/sys/fs/inode-nr there should be two numbers:<br>
>> The first is the total number of in-memory inodes for all filesystems.<br>
>> The second is the number of "unused" inodes.<br>
>><br>
>> When you write "3" to drop_caches, the second number should drop down to<br>
>> nearly zero (I get 95 on my desktop, down from 6524).<br>
>><br>
>> When signal_cache stays large even after the drop_caches, it suggest<br>
>> that there are lots of lustre inodes that are thought to be still<br>
>> active. I'd have to do a bit of digging to understand what that means,<br>
>> and a lot more to work out why lustre is holding on to inodes longer<br>
>> than you would expect (if that actually is the case).<br>
>><br>
>> If an inode still has cached data pages attached that cannot easily be<br>
>> removed, it will not be purged even if it is unused.<br>
>> So if you see the "unused" number remaining high even after a<br>
>> "drop_caches", that might mean that lustre isn't letting go of cache<br>
>> pages for some reason.<br>
>><br>
>> NeilBrown<br>
>><br>
>><br>
>><br>
>> On Mon, Apr 29 2019, Jacek Tomaka wrote:<br>
>><br>
>> > Wow, Thanks Nathan and NeilBrown.<br>
>> > It is great to learn about slub merging. It is awesome to have a<br>
>> > reproducer.<br>
>> > I am yet to trigger my original problem with slurm_nomerge but<br>
>> > slabinfo tool (in kernel sources) can actually show merged caches:<br>
>> > kernel/3.10.0-693.5.2.el7/tools/slabinfo -a<br>
>> ><br>
>> > :t-0000112 <- sysfs_dir_cache kernfs_node_cache blkdev_integrity<br>
>> > task_delay_info<br>
>> > :t-0000144 <- flow_cache cl_env_kmem<br>
>> > :t-0000160 <- sigqueue lov_object_kmem<br>
>> > :t-0000168 <- lovsub_object_kmem osc_extent_kmem<br>
>> > :t-0000176 <- vvp_object_kmem nfsd4_stateids<br>
>> > :t-0000192 <- ldlm_resources kiocb cred_jar inet_peer_cache key_jar<br>
>> > file_lock_cache kmalloc-192 dmaengine-unmap-16 bio_integrity_payload<br>
>> > :t-0000216 <- vvp_session_kmem vm_area_struct<br>
>> > :t-0000256 <- biovec-16 ip_dst_cache bio-0 ll_file_data kmalloc-256<br>
>> > sgpool-8 filp request_sock_TCP rpc_tasks request_sock_TCPv6<br>
>> > skbuff_head_cache pool_workqueue lov_thread_kmem<br>
>> > :t-0000264 <- osc_lock_kmem numa_policy<br>
>> > :t-0000328 <- osc_session_kmem taskstats<br>
>> > :t-0000576 <- kioctx xfrm_dst_cache vvp_thread_kmem<br>
>> > :t-0001152 <- signal_cache lustre_inode_cache<br>
>> ><br>
>> > It is not on a machine that had the problem i described before but the<br>
>> > kernel version is the same so I am assuming the cache merges are the<br>
>> same.<br>
>> ><br>
>> > Looks like signal_cache points to lustre_inode_cache.<br>
>> > Regards.<br>
>> > Jacek Tomaka<br>
>> ><br>
>> ><br>
>> > On Thu, Apr 25, 2019 at 7:42 AM NeilBrown <neilb@suse.com> wrote:<br>
>> ><br>
>> >><br>
>> >> Hi,<br>
>> >> you seem to be able to reproduce this fairly easily.<br>
>> >> If so, could you please boot with the "slub_nomerge" kernel parameter<br>
>> >> and then reproduce the (apparent) memory leak.<br>
>> >> I'm hoping that this will show some other slab that is actually using<br>
>> >> the memory - a slab with very similar object-size to signal_cache that<br>
>> >> is, by default, being merged with signal_cache.<br>
>> >><br>
>> >> Thanks,<br>
>> >> NeilBrown<br>
>> >><br>
>> >><br>
>> >> On Wed, Apr 24 2019, Nathan Dauchy - NOAA Affiliate wrote:<br>
>> >><br>
>> >> > On Mon, Apr 15, 2019 at 9:18 PM Jacek Tomaka <jacekt@dug.com> wrote:<br>
>> >> ><br>
>> >> >><br>
>> >> >> >signal_cache should have one entry for each process (or<br>
>> thread-group).<br>
>> >> >><br>
>> >> >> That is what i thought as well, looking at the kernel source,<br>
>> >> allocations<br>
>> >> >> from<br>
>> >> >> signal_cache happen only during fork.<br>
>> >> >><br>
>> >> >><br>
>> >> > I was recently chasing an issue with clients suffering from low memory<br>
>> >> and<br>
>> >> > saw that "signal_cache" was a major player. But the workload on those<br>
>> >> > clients was not doing a lot of forking. (and I don't *think*<br>
>> threading<br>
>> >> > either) Rather it was a LOT of metadata read operations.<br>
>> >> ><br>
>> >> > You can see the symptoms by a simple "du" on a Lustre file system:<br>
>> >> ><br>
>> >> > # grep signal_cache /proc/slabinfo<br>
>> >> > signal_cache 967 1092 1152 28 8 : tunables 0 0<br>
>> >> 0<br>
>> >> > : slabdata 39 39 0<br>
>> >> ><br>
>> >> > # du -s /mnt/lfs1/projects/foo<br>
>> >> > 339744908 /mnt/lfs1/projects/foo<br>
>> >> ><br>
>> >> > # grep signal_cache /proc/slabinfo<br>
>> >> > signal_cache 164724 164724 1152 28 8 : tunables 0 0<br>
>> >> 0<br>
>> >> > : slabdata 5883 5883 0<br>
>> >> ><br>
>> >> > # slabtop -s c -o | head -n 20<br>
>> >> > Active / Total Objects (% used) : 3660791 / 3662863 (99.9%)<br>
>> >> > Active / Total Slabs (% used) : 93019 / 93019 (100.0%)<br>
>> >> > Active / Total Caches (% used) : 72 / 107 (67.3%)<br>
>> >> > Active / Total Size (% used) : 836474.91K / 837502.16K (99.9%)<br>
>> >> > Minimum / Average / Maximum Object : 0.01K / 0.23K / 12.75K<br>
>> >> ><br>
>> >> > OBJS ACTIVE USE OBJ SIZE SLABS OBJ/SLAB CACHE SIZE NAME<br>
>> >> ><br>
>> >> > 164724 164724 100% 1.12K 5883 28 188256K signal_cache<br>
>> >> ><br>
>> >> > 331712 331712 100% 0.50K 10366 32 165856K ldlm_locks<br>
>> >> ><br>
>> >> > 656896 656896 100% 0.12K 20528 32 82112K kmalloc-128<br>
>> >> ><br>
>> >> > 340200 339971 99% 0.19K 8100 42 64800K kmalloc-192<br>
>> >> ><br>
>> >> > 162838 162838 100% 0.30K 6263 26 50104K osc_object_kmem<br>
>> >> ><br>
>> >> > 744192 744192 100% 0.06K 11628 64 46512K kmalloc-64<br>
>> >> ><br>
>> >> > 205128 205128 100% 0.19K 4884 42 39072K dentry<br>
>> >> ><br>
>> >> > 4268 4256 99% 8.00K 1067 4 34144K kmalloc-8192<br>
>> >> ><br>
>> >> > 162978 162978 100% 0.17K 3543 46 28344K vvp_object_kmem<br>
>> >> ><br>
>> >> > 162792 162792 100% 0.16K 6783 24 27132K<br>
>> >> kvm_mmu_page_header<br>
>> >> ><br>
>> >> > 162825 162825 100% 0.16K 6513 25 26052K sigqueue<br>
>> >> ><br>
>> >> > 16368 16368 100% 1.02K 528 31 16896K nfs_inode_cache<br>
>> >> ><br>
>> >> > 20385 20385 100% 0.58K 755 27 12080K inode_cache<br>
>> >> ><br>
>> >> ><br>
>> >> > Repeat that for more (and bigger) directories and slab cache added up<br>
>> to<br>
>> >> > more than half the memory on this 24GB node.<br>
>> >> ><br>
>> >> > This is with CentOS-7.6 and lustre-2.10.5_ddn6.<br>
>> >> ><br>
>> >> > I worked around the problem by tackling the "ldlm_locks" memory usage<br>
>> >> with:<br>
>> >> > # lctl set_param ldlm.namespaces.lfs*.lru_max_age=10000<br>
>> >> ><br>
>> >> > ...but I did not find a way to reduce the "signal_cache".<br>
>> >> ><br>
>> >> > Regards,<br>
>> >> > Nathan<br>
>> >><br>
>> ><br>
>> ><br>
>> > --<br>
>> > *Jacek Tomaka*<br>
>> > Geophysical Software Developer<br>
>> ><br>
>> ><br>
>> ><br>
>> ><br>
>> ><br>
>> ><br>
>> > *DownUnder GeoSolutions*<br>
>> > 76 Kings Park Road<br>
>> > West Perth 6005 WA, Australia<br>
>> > *tel *+61 8 9287 4143 <+61%208%209287%204143><br>
>> > jacekt@dug.com<br>
>> > *www.dug.com <<a href="http://www.dug.com">http://www.dug.com</a>>*<br>
>><br>
><br>
><br>
> -- <br>
> *Jacek Tomaka*<br>
> Geophysical Software Developer<br>
><br>
><br>
><br>
><br>
><br>
><br>
> *DownUnder GeoSolutions*<br>
> 76 Kings Park Road<br>
> West Perth 6005 WA, Australia<br>
> *tel *+61 8 9287 4143 <+61%208%209287%204143><br>
> jacekt@dug.com<br>
> *www.dug.com <<a href="http://www.dug.com">http://www.dug.com</a>>*<br>
</div>
</span></font></div>
</body>
</html>