bhyve uses all available memory during IO-intensive operations

[Allan Jude]

On 12/01/2017 12:42, Dustin Wenz wrote:
> Here's the top -uS output from a test this morning:
> 
> last pid: 57375;  load averages:  8.29,  7.02,  4.05                                                                                                                                                                                                                                                                                                     up 38+22:19:14  11:28:25
> 68 processes:  2 running, 65 sleeping, 1 waiting
> CPU:  0.1% user,  0.0% nice, 40.4% system,  0.4% interrupt, 59.1% idle
> Mem: 2188K Active, 4K Inact, 62G Wired, 449M Free
> ARC: 7947M Total, 58M MFU, 3364M MRU, 1000M Anon, 2620M Header, 904M Other
>      4070M Compressed, 4658M Uncompressed, 1.14:1 Ratio
> Swap: 112G Total, 78M Used, 112G Free, 4K In, 12K Out
> 
>   PID    UID    THR PRI NICE   SIZE    RES STATE   C   TIME    WCPU COMMAND
>    11      0     24 155 ki31     0K   384K RUN     0    ??? 1446.82% idle
>     0      0    644 -16    -     0K 10304K swapin 21 554:59 492.45% kernel
> 57333      0     30  20    0 17445M  1325M kqread  9  16:38 357.42% bhyve
>    15      0     10  -8    -     0K   192K arc_re 20  80:54  81.55% zfskern
>     5      0      6 -16    -     0K    96K -       5  12:35  11.50% cam
>    12      0     53 -60    -     0K   848K WAIT   21  74:35   9.40% intr
> 41094      0     30  20    0 17445M 14587M kqread 17 301:29   0.39% bhyve
> 
> Dec  1 11:29:31 <kern.err> service014 kernel: pid 57333 (bhyve), uid 0, was killed: out of swap space
> Dec  1 11:29:31 <kern.err> service014 kernel: pid 69549 (bhyve), uid 0, was killed: out of swap space
> Dec  1 11:29:31 <kern.err> service014 kernel: pid 41094 (bhyve), uid 0, was killed: out of swap space
> 
> 
> This was with three VMs running, but only one of them was doing any IO. Note that the whole machine hung for about about 60 seconds before the VMs were shut down and memory recovered. That's why the top output is over a minute older than the kill messages (top had stopped refreshing).
> 
> What I'm suspicious of is that almost all of the physical memory is wired. If that is bhyve memory, why did it not page out?
> 
> 
> 	- .Dustin
> 
> 
>> On Nov 30, 2017, at 5:15 PM, Dustin Wenz <dustinwenz at ebureau.com> wrote:
>>
>> I'm using chyves on FreeBSD 11.1 RELEASE to manage a few VMs (guest OS is also FreeBSD 11.1). Their sole purpose is to house some medium-sized Postgres databases (100-200GB). The host system has 64GB of real memory and 112GB of swap. I have configured each guest to only use 16GB of memory, yet while doing my initial database imports in the VMs, bhyve will quickly grow to use all available system memory and then be killed by the kernel:
>>
>> 	kernel: swap_pager: I/O error - pageout failed; blkno 1735,size 4096, error 12
>> 	kernel: swap_pager: I/O error - pageout failed; blkno 1610,size 4096, error 12
>> 	kernel: swap_pager: I/O error - pageout failed; blkno 1763,size 4096, error 12
>> 	kernel: pid 41123 (bhyve), uid 0, was killed: out of swap space
>>
>> The OOM condition seems related to doing moderate IO within the VM, though nothing within the VM itself shows high memory usage. This is the chyves config for one of them:
>>
>> 	bargs                      -A -H -P -S
>> 	bhyve_disk_type            virtio-blk
>> 	bhyve_net_type             virtio-net
>> 	bhyveload_flags
>> 	chyves_guest_version       0300
>> 	cpu                        4
>> 	creation                   Created on Mon Oct 23 16:17:04 CDT 2017 by chyves v0.2.0 2016/09/11 using __create()
>> 	loader                     bhyveload
>> 	net_ifaces                 tap51
>> 	os                         default
>> 	ram                        16G
>> 	rcboot                     0
>> 	revert_to_snapshot
>> 	revert_to_snapshot_method  off
>> 	serial                     nmdm51
>> 	template                   no
>> 	uuid                       8495a130-b837-11e7-b092-0025909a8b56
>>
>>
>> I've also tried using different bhyve_disk_types, with no improvement. How is it that bhyve can use far more memory that I'm specifying?
>>
>> 	- .Dustin
> 

'Wired' memory, specifically means that it cannot be paged out. It is
not bhyve, it is ZFS.

Please lower your vfs.zfs.arc_max to about 10-12 GB instead of 60+ GB it
is at now.

And you might also want to double your: vfs.zfs.arc_free_target to make
ZFS give up memory more easily.

-- 
Allan Jude