optimal kernel options for VMWARE guest system

[Jeff Dickens]

Jeff Dickens wrote:
> Jeff Dickens wrote:
>> John Nielsen wrote:
>>> On Tuesday 03 October 2006 12:58, Jeff Dickens wrote:
>>>  
>>>> I have some Freebsd systems that are running as VMware guests.  I'd 
>>>> like
>>>> to configure their kernels so as to minimize the overhead on the 
>>>> VMware
>>>> host system.  After reading and partially digesting the white paper on
>>>> timekeeping in VMware virtual machines
>>>> (http://www.vmware.com/pdf/vmware_timekeeping.pdf) it appears that I
>>>> might want to make some changes.
>>>>
>>>> Has anyone addressed this issue?
>>>>     
>>>
>>> I haven't read the white paper (yet; thanks for the link), but I've 
>>> had good results with recent -STABLE VM's running under ESX server 
>>> 3. Some thoughts:
>>>
>>> As I do on most of my installs, I trimmed down GENERIC to include 
>>> just the drivers I use. In this case that was mpt for the disk and 
>>> le for the network (although I suspect forcing the VM to present 
>>> e1000 hardware and then using the em driver would work as well if 
>>> not better).
>>>
>>> The VMware tools package that comes with ESX server does a poor job 
>>> of getting itself to run, but it can be made to work without too 
>>> much difficulty. Don't use the port, run the included install script 
>>> to install the files, ignore the custom network driver and compile 
>>> the memory management module from source (included). If using X.org, 
>>> use the built-in vmware display driver, and copy the vmmouse driver 
>>> .o file from the VMware tools dist to the appropriate dir under 
>>> /usr/X11. Even though the included file is for X.org 6.8, it works 
>>> fine with 6.9/7.0 (X.org 7.1 should include the vmmouse driver.) Run 
>>> the VMware tools config script from a non-X terminal (and you can 
>>> ignore the warning about running it remotely if you're using SSH), 
>>> so it won't mess with your X display (it doesn't do anything not 
>>> accomplished above). Then run the rc.d script to start the VMware 
>>> tools.
>>>
>>> I haven't noticed any timekeeping issues so far.
>>>
>>> JN
>>> _______________________________________________
>>>   
>> What is the advantage of using the "e1000 hardware", and is this 
>> documented somewhere?  I got the vxn network driver working without 
>> issues; I just had to edit the .vxn file manually:  I'm using the 
>> free VMware server V1 rather than the ESX server.
>>
>>   ethernet0.virtualDev="vmxnet"
>>
>> I've got timekeeping running stably on these.  I turn on time sync 
>> via vmware tools in the .vmx file:
>>
>>  tools.syncTime = "TRUE"
>>
>> and in the guest file's rc.conf start ntpd with flags "-Aqgx &" so it 
>> just syncs once at boot and exits.
>>
>> I'm not using X on these.  They're supposed to be clean & lean 
>> systems to run such things as djbdns and qmail.  And they do work 
>> well. My main goal is to reduce the background load on the VMware 
>> host system so that it isn't spending more time than it has to 
>> simulating interrupt controllers for the guests.  I'm wondering about 
>> the "disable ACPI" boot option.  I suppose I first should figure out 
>> how to even roughly measure the effect of any changes I might make.
>>
> Well, I've done some pseudo-scientific measurement on this.  I 
> currently have five freebsd virtual systems running, and one Centos 4 
> (linux 2.6),   This command give some info on the background cpu usage:
>
> (The host is a Centos 3 system, linux 2.4)
>
> [root at otter root]# ps auxww | head -1
> USER       PID %CPU %MEM   VSZ  RSS TTY      STAT START   TIME COMMAND
> [root at otter root]# ps auxww | grep vmx
> root     18031 12.7  1.5 175440 39916 ?      S<   Oct09 345:50 
> /usr/lib/vmware/bin/vmware-vmx -C /var/lib/vmware/Virtual 
> Machines/Goose/freebsd-6.1-i386.vmx -@ ""
> root     18058 12.9  1.4 174772 36916 ?      S<   Oct09 351:01 
> /usr/lib/vmware/bin/vmware-vmx -C /var/lib/vmware/Virtual 
> Machines/Duck/freebsd-6.1-i386.vmx -@ ""
> root     18072 16.2  5.5 246372 141776 ?     S<   Oct09 440:16 
> /usr/lib/vmware/bin/vmware-vmx -C /var/lib/vmware/Virtual 
> Machines/BlueJay/freebsd-6.1-i386.vmx -@ ""
> root     18086 12.9  1.4 174688 38464 ?      S<   Oct09 351:47 
> /usr/lib/vmware/bin/vmware-vmx -C /var/lib/vmware/Virtual 
> Machines/Heron/freebsd-6.1-i386.vmx -@ ""
> root     18100  9.4  4.1 385712 107348 ?     S<   Oct09 256:25 
> /usr/lib/vmware/bin/vmware-vmx -C /var/lib/vmware/Virtual 
> Machines/Newt/freebsd-6.1-i386.vmx -@ ""
> root     18139 12.2  2.5 299388 65132 ?      S<   Oct09 330:35 
> /usr/lib/vmware/bin/vmware-vmx -C /var/lib/vmware/Virtual 
> Machines/Centos4/Centos4.vmx -@ ""
> root     28930  0.0  0.0  3680  672 pts/3    S    14:08   0:00 grep vmx
> [root at otter root]#
>
>
> As one can see the one called "Newt" is consistently lower in the 
> "%CPU" column.  Curiously enough, this *is* the one I built a custom 
> kernel for.
> The config file I used is posted below:  Besides commenting out 
> devices I wasn't using & NFS, etc, I commented out the apic and 
> pctimer devices.  Do you think I'm on the right track for reducing 
> interrupt frequency?
>
> Also, if I were to want to move this kernel to other FreeBSD systems, 
> how much has to move, the whole /boot/kernel directory?
>
> Finally I did have to re-run the vmware-config-tools.pl script after 
> rebuilding the kernel.
> <snip>
Could anyone perhaps share their thoughts on just this part of my 
question?  Thanks in advance.

--->Also, if I were to want to move this kernel to other FreeBSD 
systems, how much has to move, the whole /boot/kernel directory?