On Mon, 12 Jun 2006, Dave Hansen wrote: > On Mon, 2006-06-12 at 18:00 -0700, jnorman wrote: >> Where can I find the patches and instructions for hot-remove for memory >> hotplug? > > If you'd like to document all of this and post it somewhere, I'm sure it > would be appreciated. We could probably use a wiki or something. > > Mel Gorman (cc'd) probably has the most recent set of patches handy. > I don't have a recent set of patches any more. It has been some months since I was last working with hot-remove and then I was taking patches from http://sr71.net/patches/ and rebasing them to a current kernel. At the time, I was using the mhp trees from http://sr71.net/patches and rebasing what was needed to the latest kernel. I lost the machine with those patches since but at a glance, I needed a number of the K* and P* patches from that tree. I don't remember it being particularly difficult to get hotremove working but it was not rock solid either. >> For the test cases I need to be able to determine how much physical >> memory the system has without the memory being onlined, is there a way >> to do this? > > All memory known to the kernel is presented in the same place as the > online memory: /sys/devices/system/memory. The only difference is that > the state shows 'offline'. > What I used to do was let the system boot as normal, then do something like 1. Record what memory looked like 2. echo offline to /sys/devices/system/memory on a number of banks 3. online the banks again using the probe file 4. Run a number of tests and benchmarks 5. Try and offline all memory and measure the success rate The tests were run as part of a program called bench-hotremovecapability that used a very primitive helper script called hotmemory_onoff. They are both part of VMRegress http://www.csn.ul.ie/~mel/projects/vmregress/vmregress-0.41.tar.gz but I stress that they are very primitive. >> I will also need a way to migrate pages in memory and then be able to >> verify that they have been successfully migrated, how would I go about >> that? > > This is actually a completely separate topic. There is some NUMA page > migration functionality in the kernel, but it isn't tied in to hotplug > as far as I know. > No, it hasn't. It's really a NUMA thing and there are a number of important restrictions as far as hotremove is concerned. -- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab
