Re: Dom0 memory allocation?

[Manuel Bouyer <bouyer%antioche.eu.org@localhost>]

On Thu, Mar 12, 2015 at 07:35:14PM +0100, tg%gmplib.org@localhost wrote:
>   > The system has 96 GiB RAM and 20 DomU's.  I had assigned 1 GiB for Dom0
>   > when this happened, and no swap partition.  I have never seen any
>   > recommendations on how to make an informed choice of Dom0 RAM
>   > allocation.  Could you help?
>   
>   It really depends on what you're doing in the dom0, and how much memory
>   you allocated to tmpfs. Each guest probably consumes a bit of memory in
>   dom0; but I don't know how much. HVM do probably need more than PV because
>   of the qemu-dm process and associated services (especially the framebuffer
>   and the VNC server - which of course also depends on the framebuffer
>   resolution).
>   How much is "a few" HVMs ?
> 
> Currently 6.  (2 cnfigured with vnc=0 and sdl=0, which I hope means no
> frame buffer.)

An advisory published today says that despite this config, a vnc or x11 backend
is started anyway. But it should not matter much for 2 domUs.

> 
> The corresponding qemu-dm processes take varying amount of memory.  This
> is 'top -o size 10':
> 
>   PID USERNAME PRI NICE   SIZE   RES STATE      TIME   WCPU    CPU COMMAND
>  3816 root      85    0   261M  240M ttyraw     0:08  0.00%  0.00% xentop
>  2724 root      43    0   130M   66M parked     0:12  0.00%  0.00% qemu-dm
>  1619 root      85    0   109M   26M netio      0:26  0.00%  0.00% qemu-dm
>  3909 root      85    0   107M   29M netio      0:25  0.00%  0.00% qemu-dm
>   724 root      43    0    93M   30M parked     0:00  0.00%  0.00% named
>   450 root      85    0    88M   22M netio      0:17  0.00%  0.00% qemu-dm
>  1604 root      85    0    81M   20M kqueue     0:00  0.00%  0.00% dhcpd
>  1307 root      85    0    65M   15M netio      0:22  0.00%  0.00% qemu-dm
>  1068 root      85    0    59M 9164K netio      0:16  0.15%  0.15% qemu-dm
>  1841 postfix   85    0    48M 4008K kqueue     0:00  0.00%  0.00% qmgr
> 
> The big hog is xentop.
> Oh, wait a minute.  When I now restarted xentop, it disappeared from the
> list.  But the xentop from above had not run more than a few hours.
> When the system died, I think a xentop had run for two or three days.
> 
> Running
>   while true; do ps agux | grep xento[p]; sleep 60; done
> for a few minutes reveals tell-tale data:
> 
> root    3096  0.0  0.3  28044  6452 pts/4  S+    6:32PM 0:00.15 xentop 
> root    3096  0.0  0.3  29068  7320 pts/4  S+    6:32PM 0:00.18 xentop 
> root    3096  0.0  0.4  30092  8184 pts/4  S+    6:32PM 0:00.21 xentop 
> root    3096  0.0  0.4  31116  9048 pts/4  S+    6:32PM 0:00.24 xentop 
> root    3096  0.0  0.5  31116  9904 pts/4  S+    6:32PM 0:00.27 xentop 
> root    3096  0.0  0.5  32140 10776 pts/4  S+    6:32PM 0:00.30 xentop 
> root    3096  0.0  0.6  33164 11640 pts/4  S+    6:32PM 0:00.33 xentop 
> root    3096  0.0  0.6  34188 12512 pts/4  S+    6:32PM 0:00.36 xentop 
> root    3096  0.0  0.6  35212 13376 pts/4  S+    6:32PM 0:00.39 xentop 
> root    3096  0.0  0.7  36236 14248 pts/4  S+    6:32PM 0:00.42 xentop 
> 
> I think I'm starting to understand why 1 GiB Dom0 allocation was not
> enough.  :-)
> 
> It is possible to speed up the leak by passing -d 0.5 to xentop.

OK, I guess that explains it :)
that's good to know; I never noticed such problem with the older Xens.

-- 
Manuel Bouyer <bouyer%antioche.eu.org@localhost>
     NetBSD: 26 ans d'experience feront toujours la difference
--