IRC, freenode, #hurd, 2011-11-28
<braunr> youpi: would you find it reasonable to completely disable the page
cache in gnumach ?
<braunr> i'm wondering if it wouldn't help make the system more stable
under memory pressure
<youpi> assuming cache=writeback in gnumach?
<youpi> because disabling the page cache will horribly hit performance
<braunr> no, it doesn't have anything to do with the host
<braunr> i'm not so sure
<braunr> while observing the slab allocator, i noticed our page cache is
not used that often
<youpi> eeh?
<youpi> apart from the damn 4000 limitation, I've seen it used
<youpi> (and I don't why it wouldn't be used)
<youpi> (e.g. for all parts of libc)
<youpi> ah, no, libc would be kept open by ext2fs
<braunr> taht's precisely because of the 4k limit
<youpi> but e.g. .o file emitted during make
<braunr> well, no
<youpi> well, see the summary I had posted some time ago, the 4k limit
makes it completely randomized
<youpi> and thus you lose locality
<braunr> yes
<youpi> but dropping the limit would just fix it
<braunr> that's my point
<youpi> which I had tried to do, and there were issues, you mentioned why
<youpi> and (as usual), I haven't had anyu time to have a look at the issue
again
<braunr> i'm just trying to figure out the pros and cons for having teh
current page cache implementation
<braunr> but are you saying you tried with a strict limit of 0 ?
<youpi> non, I'm saying I tried with no limit
<youpi> but then memory fills up
<braunr> yes
<youpi> so trying to garbage collect
<braunr> i tried that too, the system became unstable very quickly
<youpi> but refs don't falldown to 0, you said
<braunr> did i ?
<youpi> or maybe somebody else
<youpi> see the list archives
<braunr> that's possible
<braunr> i'd imagine someone like sergio lopez
<youpi> possibly
<youpi> somebody that knows memory stuff way better than me in any case
<braunr> youpi: i'm just wondering how much we'd loose by disabling the
page cache, and if we actually gain more stability (and ofc, if it's
worth it)
<youpi> no idea, measures will tell
<youpi> fixing the page cache shouldn't be too hard I believe, however
<youpi> you just need to know what you are doing, which I don't
<youpi> I do believe the cache is still at least a bit useful
<youpi> even if dumb because of randomness
<youpi> e.g. running make lib in the glibc tree gets faster on second time
<youpi> because the cache wouldbe filled at least randomly with glibc tree
stuff
<braunr> yes, i agree on that
<youpi> braunr: btw, the current stability is fine for the buildds
<youpi> restarting them every few days is ok
<youpi> so I'd rather keep the performance :)
<braunr> ok
