John Klein wrote on Sat, Aug 03, 2002 at 09:12:39AM -0700: > > > --- "Michael J. Ferrador" <[EMAIL PROTECTED]> wrote: > > 2.4 kernel? but not the latest (< .18 ? or so) ? > > > > Are you looking at resident, or allocated pages? > > > > Earlier 2.4's agressively (some said too much) paged out untouched > > pages. So late in the 2.4 series, with much controversy, they changed > > the VM code.
First of all, the Linux version of CMUCl has been changed not to simply overcommit. That was necceassry because some interim Linux kernel versions turned off overcommit by default. The FreeBSD version will always fully overcommit. Then, the VM systems are just not comparable. The different numbers you see are most likely not related to anything CMUCL does. Either VM poolicies are different or the way the numbers presented to you are not computed in compatible ways. Probably both. The 2.4.10+ VM system in Linux is crap, BTW, I highly recommend using the Van Riel VM system for any kind of heavy load of non-interactive application. > > 2.4.9 kernel (you mean Linux, I assume). > > Both top's SIZE and RSS are about 35 meg - SHARE is about 14.4. > %MEM times physical memory roughly matches SIZE, as I would expect. > I don't know whether SHARE is a subset of SIZE, but even if it > isn't cmucl on Linux still seems to use much less memory than > on FreeBSD where the resident size grew to about 120 meg... I think > the allocated size on FreeBSD was vast but I understand that CMUCL > overallocates. Much of this is rather cryptic to me. I do not believe that this is actually reflecting anything real. That doesn't mean the performance wouldn't be different, but that is because of VM systems, not CMUCL's gc. Martin
