15

Note: before you jump in too fast, yes I read linuxatemyram.com!

I have a server with 64GB RAM.

free -m says that my RAM is full, and it's not because of disk caching:

             total       used       free     shared    buffers     cached
Mem:         64458      64117        340        201         67        331
-/+ buffers/cache:      63719        739
Swap:         1532        383       1149

However, top ordered by memory usage does not add up to 64GB:

KiB Mem:  66005116 total, 65652464 used,   352652 free,    67512 buffers
KiB Swap:  1569780 total,   392656 used,  1177124 free.   337464 cached Mem

  PID USER      PR  NI    VIRT    RES    SHR S  %CPU %MEM     TIME+ COMMAND
 6258 mysql     20   0 38.665g 0.033t   4924 S   1.3 54.3 482:26.21 mysqld
 2293 root      20   0  165896 102116 101964 S   0.0  0.2   0:43.53 systemd-journal
 4909 root      20   0  377548  57840  57548 S   0.0  0.1   0:18.47 rsyslogd
26639 www       20   0  650076  53348  32968 S   0.0  0.1  11:32.27 php-fpm
26640 www       20   0  648344  51912  32984 S   0.0  0.1  11:37.43 php-fpm
26642 www       20   0  648600  51472  32580 S   0.0  0.1  11:37.16 php-fpm
26669 www       20   0  648148  50696  31988 S   0.0  0.1  11:35.24 php-fpm
26643 www       20   0  648452  50616  31628 S   0.0  0.1  11:36.19 php-fpm
26641 www       20   0  648620  50496  31340 S   0.0  0.1  11:36.51 php-fpm
28121 www       20   0  648620  48820  29660 S   0.0  0.1  11:35.75 php-fpm
27231 www       20   0  647508  48804  30760 S   0.0  0.1  11:35.61 php-fpm
28029 www       20   0  648044  48752  30172 S   0.0  0.1  11:37.20 php-fpm
28117 www       20   0  647868  48700  30296 S   0.0  0.1  11:36.45 php-fpm
28122 www       20   0  648340  48568  29676 S   0.0  0.1  11:35.73 php-fpm
 8569 www       20   0  649028  40268  20704 S   0.0  0.1  11:31.50 php-fpm
10126 www       20   0  648432  39420  20700 S   0.0  0.1   9:58.52 php-fpm
22386 www       20   0  647996  39400  20868 S   0.0  0.1  11:25.00 php-fpm
 9643 www       20   0  647976  39220  20704 S   0.0  0.1  11:29.23 php-fpm
23077 www       20   0  647852  39084  20692 S   0.0  0.1  11:11.80 php-fpm
10139 www       20   0  647580  38808  20692 S   0.0  0.1   9:59.94 php-fpm
 6326 www       20   0  647368  38396  20696 S   0.7  0.1   8:32.34 php-fpm
 4727 www       20   0  646128  37304  20692 S   0.0  0.1   8:30.20 php-fpm
 5459 www       20   0  645988  37156  20688 S   0.0  0.1   7:15.13 php-fpm
 2173 www       20   0  645240  36408  20684 S   0.0  0.1   4:39.13 php-fpm
20752 www       20   0  644536  35428  20680 S   0.0  0.1   4:29.78 php-fpm
 5396 www       20   0  644468  35324  20692 S   0.0  0.1   4:14.65 php-fpm
17558 www       20   0  642668  33816  20740 S   0.0  0.1   1:28.34 php-fpm
28133 www       20   0  642780  33636  20704 S   0.0  0.1   0:49.88 php-fpm
10925 www       20   0  479584  29264  11212 S   3.0  0.0   0:00.09 php
26632 root      20   0  552136  26072  19468 S   0.0  0.0   0:37.74 php-fpm
 4946 named     20   0  697996  18748   2104 S   0.0  0.0   3:46.96 named
15609 apache    20   0 2137056   8120   1592 S   0.0  0.0   0:56.18 httpd
 8584 root      20   0  133432   4864   3700 S   0.0  0.0   0:00.08 sshd

MySQL uses 54.3% alone, this is perfectly normal as it has an innodb_buffer_pool_size of 32G. The other processes' memory usage add up to roughly 2.8%, that's a total of 57.1%.

Where are the 32% remaining?


Edit: contents of /proc/meminfo:

MemTotal:       66005116 kB
MemFree:          353272 kB
Buffers:           66328 kB
Cached:           736620 kB
SwapCached:        11348 kB
Active:         34396680 kB
Inactive:        2651132 kB
Active(anon):   34223240 kB
Inactive(anon):  2228020 kB
Active(file):     173440 kB
Inactive(file):   423112 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:       1569780 kB
SwapFree:        1177448 kB
Dirty:               328 kB
Writeback:             0 kB
AnonPages:      36234364 kB
Mapped:           125208 kB
Shmem:            206396 kB
Slab:           28058904 kB
SReclaimable:   28010224 kB
SUnreclaim:        48680 kB
KernelStack:        2760 kB
PageTables:        94780 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:    34572336 kB
Committed_AS:   38572348 kB
VmallocTotal:   34359738367 kB
VmallocUsed:      382304 kB
VmallocChunk:   34359353572 kB
HardwareCorrupted:     0 kB
DirectMap4k:        9000 kB
DirectMap2M:     2054144 kB
DirectMap1G:    67108864 kB

Output of slabtop:

 Active / Total Objects (% used)    : 147380425 / 147413026 (100.0%)
 Active / Total Slabs (% used)      : 7005839 / 7005839 (100.0%)
 Active / Total Caches (% used)     : 71 / 144 (49.3%)
 Active / Total Size (% used)       : 27615020.12K / 27627490.91K (100.0%)
 Minimum / Average / Maximum Object : 0.01K / 0.19K / 16.12K

  OBJS ACTIVE  USE OBJ SIZE  SLABS OBJ/SLAB CACHE SIZE NAME
146851887 146851887  12%    0.19K 6992947       21  27971788K dentry
124936 124936 100%    0.07K   2231       56      8924K Acpi-ParseExt
105144 105144 100%    0.10K   2696       39     10784K buffer_head
 49920  49172  98%    0.06K    780       64      3120K kmalloc-64
 29916  29916 100%    0.11K    831       36      3324K sysfs_dir_cache
 29856  29661  99%    0.12K    933       32      3732K kmalloc-128
 21450  21128  98%    0.18K    975       22      3900K vm_area_struct
 19328  19328 100%    0.03K    151      128       604K kmalloc-32
 18258  13383  73%    0.93K    537       34     17184K ext4_inode_cache
 17952  11651  64%    0.04K    176      102       704K ext4_extent_status
 16828   6513  38%    0.55K    601       28      9616K radix_tree_node
 14400  13996  97%    0.06K    225       64       900K anon_vma
 11645   7903  67%    0.05K    137       85       548K shared_policy_node
 10710   7006  65%    0.19K    510       21      2040K kmalloc-192
 10608  10608 100%    0.04K    104      102       416K Acpi-Namespace
  9728   9728 100%    0.01K     19      512        76K kmalloc-8
...
  • 3
    read linuxatemyram.com again. Then read kernel.org/doc/gorman/html/understand – symcbean Jan 5 '15 at 10:57
  • I read it again, and it ends up with suggesting that free -m is the way to go to really get the memory usage. Also, the fact that the swap usage is not 0 made me a bit dubious. I've checked your link, but it's so much data to read that I just cannot handle all of this at the moment I'm afraid. – Benjamin Jan 5 '15 at 11:03
  • Is this a physical server? – Optichip Jan 5 '15 at 11:07
  • @Optichip It is a physical server. – Benjamin Jan 5 '15 at 11:49
  • 5
    @AndrewSchulman I don't think this is a question is a duplicate of that particular question. This question is about memory used for slabs. The other question doesn't mention slabs at all. – kasperd Jan 5 '15 at 15:48
23
  OBJS ACTIVE  USE OBJ SIZE  SLABS OBJ/SLAB CACHE SIZE NAME
146851887 146851887  12%    0.19K 6992947       21  27971788K dentry

You say it's not because of disk caching, but clearly it is.

My bet is that you have code that makes lots of fetches for files that do not exist and you get a ton of negative caching. Linux will remove these entries if it's under memory pressure, so it may not be anything to worry about. For example, like in this NSS issue.

  • 5
    Ah, that could explain that. Please forgive my ignorance, I thought that free -m alone would indicate if memory was eaten because of disk caching. I did not know about "negative caching" either, it would be really interesting to understand what's trying to find files that do not exist. Any idea how I could investigate this, out of curiosity? This is a recent install with just a database + web application, so I am really surprised to see this behaviour. – Benjamin Jan 5 '15 at 12:51
  • @Benjamin Follow my link for one example of how this can happen. Also, look at what the web application is actually doing. – David Schwartz Jan 5 '15 at 12:57
  • 10
    Actually, I do use libcurl a lot, more precisely its PHP bindings. My web server downloads and processes many images: they're downloaded into memory, processed, then uploaded to S3, never hitting the filesystem. Hence I have to admit that I overlooked your link, considering it too localized at first glance to impact me, but it might very well be this exact issue: I ran the same strace curl as the guy did, and found that it generated 7421 ENOENT (No such file or directory) for a single query! That must be it. So thanks so much! – Benjamin Jan 5 '15 at 13:24
  • 2
    You might want to see linux-mm.org/Drop_Caches and do echo 2 > /proc/sys/vm/drop_caches to non-destructively drop the dentry and inode caches, and see if your RAM returns. Obviously best done at a quiet time. – abligh Jan 5 '15 at 19:44
  • 1
    An extra confirmation that Linux will drop those entries if under pressure comes from the fact that the bulk of the active slabs (Slab in /proc/meminfo) is under SReclaimable (that, as the name explains, can be reclaimed as needed). – Marco Leogrande Jan 5 '15 at 20:27
4

It might be some internal kernel structures and filesystem/directory related stuff, still. This is also perfectly normal even though confusing; try to see what is the output from slabtopand cat /proc/meminfo.

  • I added the output of these commands to my question, could you please have a look? – Benjamin Jan 5 '15 at 10:59
  • 1
    Well, about 27 gigabytes of your RAM seems to be consumed by the kernel, and most of that goes to dentry, directory entries. I guess you have lots and lots of files around in your filesystem. Those are practically just cached stuff, in case some processes really do need the memory, it will be released in the same way than with the normal buffer/cache you tend to see in top output. – Janne Pikkarainen Jan 5 '15 at 11:08
  • That's good to know, thank you. I'm just surprised though as I don't have many files around on the file system: find / | wc -l returns 134578, not millions as dentry could suggest. Do you have any idea what could do this? – Benjamin Jan 5 '15 at 11:53
  • Likely found what's generating these junk directory entries, thanks to David Schwartz's answer above. Thanks for your help anyway! – Benjamin Jan 5 '15 at 13:26

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.