slabtop and /proc/spl/kmem/slab difference

[antry]

slabtop show

Active / Total Objects (% used) : 4852012 / 4982054 (97.4%)
Active / Total Slabs (% used) : 54282 / 54282 (100.0%)
Active / Total Caches (% used) : 72 / 105 (68.6%)
Active / Total Size (% used) : 386092.82K / 401266.20K (96.2%)
Minimum / Average / Maximum Object : 0.01K / 0.08K / 8.00K

arc_summary from /proc/spl/kmem/slab show
SPL Memory Usage:
SPL slab allocated: 90,346 MB
SPL slab used: 58,770 MB

how those counters coexists?

[AndCycle]

as I know currently ZoL use it's own SLAB, which is not handle by Linux kernel,
so you are looking at two different SLAB,

Add KMC_SLAB cache type
a073aeb
there are some work trying to leverage some work to linux kernel,
currently I have lot's problem with it so I don't recommend using that

[behlendorf]

@AndCycle Exactly which patch stack are you having problems with, it would be helpful to know.

[AndCycle]

@behlendorf sorry, I don't have a confirmed way to reproduce my problem,
tried spl_kmem_cache_slab_limit=16384 at 0.6.3,
sometimes arc_adapt will raise to 100% cpu usage for a while which I have no idea about what happened,
that leads to certain performance degradation so I go back to the default,

here comes the personal part,

after I switch to ZoL few months ago that I break my raid5,
my system is still struggle to survive,
currently longest uptime is 19 days,
there are some issue that makes it won't respond,
won't trigger panic for automatic reboot, won't respond to magic key, softdog won't bark either,
doesn't looks like easy to get into,
my workload is total mix,
I do have crashplan installed which do massive metadata related work about every week,
my daily laptop backup, even got few VM on it,
I don't have zvol/xattr or quota setting,

I will report back once I have good one, thanks for the concern.

oh, by the way, 0.6.3 function normally with PREEMPT_NONE :)

[AndCycle]

just upgrade to Gentoo ZFS 0.6.3-r1 patchset,
which include spl_kmem_cache_slab_limit=16384 as default,

still make arc_adapt hanging on top of cpu usage, so do kswapd0 working on 100% too,

although I have a 32GB machine,
my system do have heavy metadata workload bias,
which make it use up to 12GB arc_meta_used to function properly,
include the overhead, that means arc takes most of my memory,

if I set some reasonable zfs_arc_max like 8GB,
soon the system will go into unusable status due to arc doesn't response well when meta_limit been exceeded.

[kernelOfTruth]

@AndCycle sounds like that you're suffering from the problem that ARC doesn't reclaim/prune objects fast enough - please take a look at: openzfs/zfs#3181

[AndCycle]

@kernelOfTruth thanks for the reminder

[behlendorf]

Closing. Many improvements have been made to improve the ARC behavior, if issues still remain let's open a issue on the zfs tracker.