IRC, freenode, #hurd, 2011-06-07

<braunr> hm, i guess i should have used cache=writeback with kvm before
  starting the debian installer :/
<braunr> ah yes, much better
<braunr> this shows how poor the state of our I/O drivers and subsystem is
  :/
<antrik> indeed... still no clustered pageout :-(
<braunr> and no I/O scheduler either
<braunr> although an I/O scheduler has limited value without clustered
  pageouts
<braunr> since one of its goals is to pack related I/O requests together eh
<braunr> i wonder if the wiki mentions using cache=writeback to speed up
  qemu performances
<braunr> it would help those unable to use kvm a lot
<braunr> and even those running kvm too
<braunr> kvm -m $RAM \ -monitor stdio \ -drive
  cache=writeback,index=0,media=disk,file=hd0.img \
<braunr> etc..
<braunr> the idea is that qemu doesn't open its disk file synchronously
<braunr> changes are queued in the host page cache before being flushed to
  the disk image
<braunr> but if you brutally close your qemu instance, you're likely to
  loose file system consistency
<braunr> ext2fs will think it has committed its metadata to the disk, but
  the disk image won't be updated synchronously
<braunr> on my machine (which is quite fast), my kvm has installed debian
  like 10 times faster than without the option
<antrik> braunr: I don't think killing qemu should hurt in this
  case... probably only matters when the host machine dies
<braunr> antrik: ah yes, right
<braunr> it really makes everything faster, even downloading, since I/O
  requests aren't interleaved between networking RPCs
<antrik> regarding I/O sheduler... this discussion came up before, but I
  don't remember the outcome -- doesn't the glued Linux driver actually
  come with one?
<braunr> i don't remember either
<antrik> braunr: err... I don't think interleaving has anything to do with
  it... I guess it's simply the fact that downloading writes the result to
  disk, which suffers from lacking clustered pageout like everything else
<antrik> (my internet connection is too slow though to notice :-) )
<braunr> well, if there is no I/O during downloading, downloading is faster
  :)

IRC, freenode, #hurd, 2011-06-08

<braunr> youpi: does xen provide disk caching options ?
<youpi> through a blktap, probably
<braunr> ok

(Xen, Host-side Writeback Caching.)

<braunr> we should find the pages mentioning qemu on the wiki and add the
  options to enable disk image caching
<braunr> it really makes the hurd run a lot faster
<braunr> as a workaround for emulators until I/O is reworked, ofc

IRC, freenode, #hurd, 2011-06-09

<gnu_srs> braunr recommends to use writeback caching with kvm. Is this
  really recommended with the frequent crashes I experience?
<youpi> provided that you terminate your kvm normaly (i.e. quitting it, not
  killing it), there should be no difference
<jkoenig> I think the host's stability is what matters
<jkoenig> the data presumably sits in linux's cache even if qemu dies
  violently
<gnu_srs> But the freezes I see force me to kill kvm :-(
<youpi> maybe kvm doesn't even do caching indeed, I don't know
<youpi> gnu_srs: you can quit even when frozen
<youpi> use the console
<youpi> (the kvm console)
<jkoenig> gnu_srs, "Writeback caching will report data writes as completed
  as soon as the data is present in the host page cache.  This is safe as
  long as you trust your host.  If your host crashes or loses power, then
  the guest may experience data corruption." (from the qemu manpage)

IRC, freenode, #hurd, 2012-12-30

<gg0> http://bugs.debian.org/622319#51
<gg0> braunr: just pointing out writeback is default since qemu 1.3.0
<gg0> that makes hurd VMs faster by default :p
<braunr> ahh :)
<gg0> about data loss I had read (qemu man pages?) it may occur in case of
  loss of power
<braunr> well yes
<gg0> probably on hurd may occur even with writethrough due to non
  journaled ext2
<braunr> yes but the hurd flushes everything (and i mean everything) every
  5 seconds
<gg0> http://lists.gnu.org/archive/html/qemu-devel/2012-02/msg02682.html
<braunr> so it's actually less likely to lose data because of the hurd than
  because of power loss
<braunr> what i meant earlier is that we've never experienced unexpected
  data loss because of qemu writeback policy
<braunr> if data is in the host cache and you lose power, qemu or not, you
  lose data