Xen users - Jun 2007 - vcpu performance : 1 vcpu for all guets or 4 vpcu ?

Hello all ;)

Tell I have a Xeon server with 4 vcpus

If on this box I have some guets, tell 10, what is the best solutions :
- Set all guests to 1 vcpu ?
- Set all guests to 4 vcpus ?

I well understand that if I set a guest with 4 vcpus and all others only 
with one vpcus, then the one with 4 vcpus will have more "cpu time" 
available than others guests.
But in case I''d like to have an equal cpu "share" between all
guests,
which is the best solution :
- set 1 vcpu for all
- set 4 vcpu for all

Thanks a ton

Pascal  


_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users

Pascal wrote:
> Tell I have a Xeon server with 4 vcpus
> 
> If on this box I have some guets, tell 10, what is the best solutions :
> - Set all guests to 1 vcpu ?
> - Set all guests to 4 vcpus ?
In my understanding, if you set all guests to 1 vcpu, each guest can
utilize one (p)cpu at most, but most likely the credit-scheduler will
put each guest on a separate (p)cpu, thus all guests are able to run at
the same time.

If you set all guest to 4 vcpus, each guest can utilize all four (p)cpus
at most, but all guests will share these (p)cpus.

Which solution is best depends on the workload, and imho there is
no "best way" for all situations.

If all guests are at 100 percent cpu, but the guests are not able to
utilize all four vcpus because there is only one process consuming most
of the cpu time, you won''t gain any profit of using four vcpus and it
may be better to use only on vcpu.

Greetings,
-timo

-- 
Timo Benk - Jabber ID: timo@jabber.de - ICQ ID: #488595326
PGP Public Key: http://www.m28s01.vlinux.de/timo_benk_gpg_key.asc


_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users

On Sat, 2007-06-09 at 11:10 +0200, Pascal wrote:
> Hello all ;)
> 
> Tell I have a Xeon server with 4 vcpus
> 
> If on this box I have some guets, tell 10, what is the best
> solutions :
> - Set all guests to 1 vcpu ?
> - Set all guests to 4 vcpus ?
> 
> I well understand that if I set a guest with 4 vcpus and all others
> only with one vpcus, then the one with 4 vcpus will have more "cpu
> time" available than others guests.
It can be hard to contain and deal with sudden onslaughts of I/O
requests from guests. 

Our recipe has been to reserve at least 1/2 core for dom-0, then give
each guest 2, balancing them over the remaining 3 vcpu''s.

I have a whole CD full of little ''nasties'' written in perl
that show up
in /tmp on shared servers, usually because of weak php scripts. While
you won''t ''see'' the strain a few unruly guests can
put on I/O for all
guests on dom-0, you''ll surely see it in the other guests, especially
if
your using off the shelf cheap drives.
> But in case I''d like to have an equal cpu "share"
between all guests,
> which is the best solution :
> - set 1 vcpu for all 
> - set 4 vcpu for all 
How much ram are you giving your guests on average? There''s only but so
much someone can do with 256 MB regardless of how this value is set.
There *are* exceptions to this, especially if someone is determined to
cause a disruption.

Common disruptions tend not to bother any other guests but the one
running it, the isolation is very good, but things do happen.

If your guests have substantial amounts of memory (512 MB <), There
really isn''t one "good" way to do it. 

Going by the domain name of your e-mail address, I *strongly* suggest
not using all equal credit, or reserving at least one core that the
guests can''t touch. You just never know what someone is bound to
upload.

It may work perfectly for you. It all depends on who your hosting. What
sucks is having to reboot dom-0 uncleanly and watch 30+ guests fsck
themselves, just because of a bot infestation on a few servers.

I don''t see this nearly as much with Xen 3.1 as I did with 3.0 ->
3.0.4,
but it still does happen from time to time. I also stopped using ext3
and switched to jfs for all of my guests, this noticeably reduced
sluggishness and breakage due to many guests demanding I/O at once.

I wish I could give you a magic formula :) 
> Thanks a ton 
> Pascal   
Hope this is of some help.

--Tim
> _______________________________________________
> Xen-users mailing list
> Xen-users@lists.xensource.com
> http://lists.xensource.com/xen-users

_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users

On Mon, 2007-06-11 at 06:59 +0200, Pascal wrote:
> Hi Tim 
> 
> Thanks a lot for this long answer  ;)
> 
> We''ve done some tests and it seems good to have 1 core reserved
for
> dom0, the other balancing through vcpu to all guests.
> 
Or at least 1/2 of one core. It really depends on the I/O needs of your
guests.

I have one server that has nearly 50 load balancing guests running on
it. It uses software that never touches disk once resident and running.
It could easily hold 50 more. 
> Maybe one more question.
Sure :)
> We are working on a solution to count the guest trafic by month and
> maybe to limit the bandwidth used by every guests.
> 
> For the trafic, we are currently trying to play with rrdtool + a
> script that does something like this and write these data in a mysql
> database :
> ifconfig -a vif34.0 | grep ''RX bytes'' | cut -d : -f 2 |
cut -d '' '' -f
> 1
> 
> For the bandwidth limit there is of course TC 
> 
> Just by curiosity, did you implement one of these solutions ?
> Do you have others idea, is there some control panel or existing
> script which do that ?
> 
> Thanks
> 
> 
> Pascal
Try reading from /proc/net/dev directly. That pipe invokes many things
that allocate much more memory than is actually needed.

Line editors have *no* idea how big stdin is going to be, so they need
to allocate a lot of memory that is just wasted for 15 - 30 mins for
such a menial task.

Ideally, back end stuff on dom-0 tries its best not to do that, since
contiguous chunks of RAM can be hard to come by in such small spaces.

Ideally, such a solution would e-mail (or otherwise notify) your clients
when they reached about 80% of that quota. Floats in simple shell
scripts are a pain, I recommend you use C or python. Bash is very
bloated for this particular use. Every invocation of /bin/sh puts a
gazillion things in memory you''ll never use. You could try other
bourne-ish shells, but most of them don''t support arrays which
you''d
want for this.

If you don''t know python or C, try grawk.

http://dev1.netkinetics.net/hg/utils.hg/grawk.c

.. in place of grep / awk / cut

grawk "^\ \ eth0" ''$2,$9'' /proc/net/dev

produces the two values I think that you want with a single invocation. 

Of course, substitute eth0 for the vifname. That repo also contains non
blocking highly abuse-able http and ftp services you may find of use.

Please reply-to-all or CC the list so everyone can benefit from the
answers to your questions :) 

Best,
--Tim



_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users