[OmniOS-discuss] Omnios: can't start as many kvm's as I thought due to memory pressure
pieter van puymbroeck
pietervanpuymbroeck at hotmail.com
Wed Mar 8 19:32:23 UTC 2017
Hi,
I want to start 12 kvm machines in my global zone and I end up in an memlock issue after the 6th machine:
qemu_mlock: have only locked 0 of 7516192768 bytes; still trying…
The details:
# echo "::memstat" | mdb -k
Page Summary Pages MB %Tot
------------ ---------------- ---------------- ----
Kernel 1462114 5711 5%
Boot pages 166 0 0%
ZFS File Data 143529 560 0%
Anon 27477 107 0%
Exec and libs 1244 4 0%
Page cache 4911 19 0%
Free (cachelist) 3597 14 0%
Free (freelist) 29809914 116444 95%
Total 31452952 122863
Physical 31452951 122863
#
My pagesize is:
# pagesize
4096
#
so 31452952 pages = (122863*1024*1024) bytes / 4096 (bytes)
thus: 31452952 / 256 = 122863MB
That's correct. So pages to MB we can divide by 256. MB to pages, multiply with 256.
I want 8GB per vm and I want 12vm's. This is 96GB (98304MB)
-> needs 98304 * 256 = 25165824 pages.
31452952 - 25165824 = 6287128 left (or 24559 MB ~ 23.9GB )
I reserved (in /etc/system ) 15 GB for zfs arc
15GB = 15360 MB = 3932160 pages
So this gives 6287128 - 3932160 = 2354968 pages left (or 9199MB ~ 8.98 GB)
-> this should be sufficient for omnios to run. Unless I'm not aware of other things, which is very well possible.
So remember
Page Summary Pages MB %Tot
------------ ---------------- ---------------- ----
...
Free (cachelist) 3597 14 0%
Free (freelist) 29809914 116444 95%
Total 31452952 122863
Physical 31452951 122863
#
Let's boot one kvm with 8GB ram for the vm. We expect to see (8GB *1024) * 256 = 2097152 pages be eaten up. So we should end up with 31452952 - 2097152 = 29355800 pages.
OR if we are eating from the "Free" part, we should end up with 29809914 - 2097152 = 27712762 pages (or 108252MB ~105.71 GB)
Let's confirm:
Before boot:
# mdb -ke 'availrmem/D ; pages_pp_maximum/D'
availrmem:
availrmem: 29845807
pages_pp_maximum:
pages_pp_maximum: 1216998
#
Boot the machine.
# echo "::memstat" | mdb -k
Page Summary Pages MB %Tot
------------ ---------------- ---------------- ----
Kernel 1472257 5751 5%
Boot pages 166 0 0%
ZFS File Data 149321 583 0%
Anon 2139596 8357 7%
Exec and libs 1549 6 0%
Page cache 4955 19 0%
Free (cachelist) 3445 13 0%
Free (freelist) 27681663 108131 88% <<<----- 27712762 pages
Total 31452952 122863
Physical 31452951 122863
# mdb -ke 'availrmem/D ; pages_pp_maximum/D'
availrmem:
availrmem: 26621718
pages_pp_maximum:
pages_pp_maximum: 1216998
#
look there. We expected to end up with 29355800 pages or 27712762 free pages.
If we check in the available memory 29845807 - 26621718 = 3224089 pages were consumed (12594.09MB ~ 12.29GB )
Double check: 29809914 - 27681663 = 2128251 pages consumed ( 8313.48MB ~ 8.11GB )
And this for a machine of 8GB.
Let's do a test with a vm with 4GB allocated ( 1048576 pages )
Before boot:
# echo "::memstat" | mdb -k
Page Summary Pages MB %Tot
------------ ---------------- ---------------- ----
Kernel 1476857 5768 5%
Boot pages 166 0 0%
ZFS File Data 176243 688 1%
Anon 2139416 8357 7%
Exec and libs 1549 6 0%
Page cache 4961 19 0%
Free (cachelist) 3440 13 0%
Free (freelist) 27650320 108009 88%
Total 31452952 122863
Physical 31452951 122863
# mdb -ke 'availrmem/D ; pages_pp_maximum/D'
availrmem:
availrmem: 26596813
pages_pp_maximum:
pages_pp_maximum: 1216998
#
So we expect
27650320 - 1048576 = 26601744 free pages (103913.06 MB ~ 101.47 GB free)
or
26596813 - 1048576 = 25548237 free pages (99797.80 MB ~ 97.45 GB free)
Booting the vm.
And the aftermath:
# ./start_kvm_pogo.sh
qemu-system-x86_64: -net vnic,vlan=40,name=net0,ifname=vnic_kvm_pogo0,macaddr=2:8:20:87:2f:69: vnic dhcp disabled
Started VM: pogo
VNC available at: host IP 127.0.0.1
10.0.1.171
::1/128
::/0 port 5900
QEMU Monitor, do: # telnet localhost . Note: use Control ] to exit monitor before quit!
#
Check the values
# echo "::memstat" | mdb -k
Page Summary Pages MB %Tot
------------ ---------------- ---------------- ----
Kernel 1482301 5790 5%
Boot pages 166 0 0%
ZFS File Data 180958 706 1%
Anon 3198541 12494 10%
Exec and libs 1549 6 0%
Page cache 4961 19 0%
Free (cachelist) 3440 13 0%
Free (freelist) 26581036 103832 85%
Total 31452952 122863
Physical 31452951 122863
#
# mdb -ke 'availrmem/D ; pages_pp_maximum/D'
availrmem:
availrmem: 24472324
pages_pp_maximum:
pages_pp_maximum: 1216998
#
let's verify.
Test1:
27650320 - 26581036 = 1069284 pages consumed ( 4176.89 MB ~ 4GB )
Test 2:
26596813 - 24472324 = 2124489 pages consumed ( 8298.78 MB ~ 8.10GB )
3 questions remain.
- Why does this happen. If we check using the memstat it's fairly ok, but appearently this is not correct as I can't start the 12 machines.
- How to avoid this. If I say to kvm "use 8GB" I want it to use 8GB not 12GB. Or am I missing something crucial?
- If it's a (kernel) parameter ... which one and how to set it?
Thanks and best regards,
Pieter
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://omniosce.org/ml-archive/attachments/20170308/141acc32/attachment-0001.html>
More information about the OmniOS-discuss
mailing list