On Fri, 17 Jun 2005, Alexander Leidinger wrote: > Robert Watson <rwatson_at_freebsd.org> wrote: > >> Looks like what basically happened is this these kern_malloc.c changes >> increase the memory burden on UMA as statistics structures for malloc types >> now get allocated from UMA. It looks like, from your dmesg, you have a >> fair number of modules loaded, so the storage for the statistics comes out >> of the early UMA page pool, whereas before it came out of BSS. We'll see if >> further tuning is required or not with large numbers of modules. > > I try to load as much as possible as modules. Can you quantify "large > number of modules"? I could load some more modules for testing purposes > at the weekend. Well, it looked like 30 was enough to exceed the 40 page UMA threshold, but it's now been bumped to 48 in HEAD. However, what actually matters is malloc types, not modules, so I think two routes would be productive: to add a debugging printf to UMA to show how much of the boot page space is used at the time it transitions to non-boot pages, and to try creating a module that creates various numbers of malloc types. Robert N M WatsonReceived on Fri Jun 17 2005 - 08:00:25 UTC
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:38:36 UTC