It has been my experience that doing what you suggest is a bad idea, I tried it once and had very undesirable results.
It's been running smoothly as a 64-bit kernel on a Debian 32-bit userland for 3 years now. I've been running KVM based virtuals on the same system for almost 2 years now. I was suggesting the 64-bit kernel as it would help you make use of your extra memory, which is one of the reasons I upgraded to 64-bit myself. The trick is to use the 64-bit packaged kernel from the 32-bit architecture stream, which Debian makes it easy to do.
Now that I've dug a little deeper in apt on my core, I'm seeing that Ubuntu (10.04) only packages i386 and generic (which equates to i686), plus a generic-pae. They also have their server kernel, but I'm not sure it's 64-bit (I would presume not), and you'd lose drivers for consumer hardware. So, to get a 64-bit kernel on 32-bit architecture stream, it would take getting the 64-bit streams packages and installing on 32-bit, which would likely have serious breakage.
So, the easiest approach on Ubuntu would be to install the generic-pae kernel if it exists in 8.10 (apt-cache search linux-image or linux-kernel |grep pae). Don't install the server kernel; it'll be missing a lot of drivers for commodity hardware. Whether or not you lose hardware acceleration would depend on your graphics card hardware and the quality of the drivers. You will have a minor performance hit for the memory paging, but it should be negligible.
One thing to consider is what the box is doing when the memory is being used. 'free' and 'top' will tell you a lot there. If you have lots cached and you're not swapping, you're in a good place. Modern kernels will use free memory for the file system cache, and release it to apps as needed. Here's a funny one for ya; Windows can run faster virtualized on Linux than directly on the bare metal, because of the Linux FS Cache. It helps crank up the I/O speeds, where Windows typically has some trouble.
My comments are given as someone who administers Linux systems on many different hardware platforms, architectures and hypervisors at work. Hope it helps!