On Thu, 2005-03-10 at 11:48 -0500, Stephen Smalley wrote: > On Wed, 2005-03-09 at 09:58 -0600, Serge Hallyn wrote: > > The modified stacker creates a list_head for each LSM hook and adds a > > module's hook to the list only if it is defined. This way we don't have > > to check for (module->operation) for each module on each hook call. > > This results in performance on macrobenchmarks (kernel compile) which > > actually seems on par with or slightly better than non-stacker. Lmbench > > results (attached) for some reason do not back this up. I can't explain > > those results. Might just try a whole new set of lmbench tests, in case > > I left some service enabled. > > Some of the data seems to be missing, or I'm blind. hlist and stack are > missing their UDP data, while nostack is missing TCP conn data. Yes, I don't know what happened with those... I don't seem to have good luck with lmbench. On my rel4/ppc64 system the scripts failed altogether. > Which results did you view as especially problematic? It would help to > have a summary table comparing the means of the different > configurations. Well, the shproc in particular, as well as fork proc. > In the past, I think I've run a lot of trials of lmbench to reduce > noise, not sure you are running enough here. How many would you recommend? Would 10 be enough? I suppose I should just be calculating 95%CI as a guide. I am reinstalling rel4 (or should I be using fedora instead?) on a 4-cpu ppc64 on ext2 for benchmarking. The previous results were done on my main desktop. which kind of put a stop to other work :) thanks, -serge -- Serge Hallyn <serue@private>
This archive was generated by hypermail 2.1.3 : Thu Mar 10 2005 - 09:16:15 PST