>>> On Sat, Sep 1, 2007 at 1:04 PM, in message <[EMAIL PROTECTED]>, Tom Marchant <[EMAIL PROTECTED]> wrote: -snip- > My comments were in response to the creation of Linux images that > did not include swap space, and that were therefore not going to do any > paging on their own. It is not obvious to me under those conditions that it > is a bad idea to provide a sufficiently sized virtual guest is a bad idea.
It is a very bad idea, for all the reasons we've been talking about, and the one I talk about below. It would be a bad idea on VMWare or Xen as well. The only place it would make sense is on discrete servers that aren't sharing any resources with other system images. > Indeed, I wonder if it might be better to define Linux guests that way and > let z/VM handle all the paging for all of them. As Lynn pointed out, > strange things can happen when you run a paging guest in a virtual machine No, that would not be better, at all. It would be much, much, worse, and cause you to spend far too much money on real storage, and still not utilize as much of the rest of the hardware as you should. -snip- > The potential problem is not with the virtual machine's memory size, but > with the working set size. A 64 GB z/VM guest is only a problem if it has a > large working set. Linux will use what it needs, not necessarily all that > is available. And if it has a 16 GB working set, what happens to it if it > is in a 4 GB virtual machine? If you give a Linux guest a 4GB virtual machine, it will have very close to a 4GB working set. If you give that same Linux guest 64GB, it will have very close to a 64GB working set. The fact that you say "Linux will use what it needs" tells me that you have little or no experience running Linux, either on midrange systems or on the mainframe. Linux will _always_ use everything you give it, if for nothing else than buffers and cache. Hence the constant battle we have with midrange Linux sysadmins, DBAs, etc., regarding this topic. It's also a recurring topic with the midrange performance/capacity folks, since we keep getting concerned phone calls and emails about how we have to add more RAM to a Linux system because it's running out. It hasn't run out, it is just using all the otherwise unused storage for buffers and cache, but that's not the behavior they're used to seeing from AIX, Solaris, HPUX, etc. >>As someone who is very heavy into z/VM performance once told me, >>"z/VM is very good at managing large numbers of small things. It's >>not so good at managing a smaller number of very large things." >>I tend to agree. The z/VM scheduler isn't too happy about guests >>with large working sets. > > I'm not impressed with any quote of an anonymous "authority." That's OK, I wasn't expecting anyone to be impressed, just take the point at face value. It has been everyone's experience that this is the case, including mine. Major surgery had to be performed on z/VM 5.x to get it to work reasonably well with some number of large guests. It's still not as good at it as we would like, but I know the z/VM developers are working on that too. Mark Post ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
