IRC, freenode, #hurd, 2011-09-01:

<youpi> hum, f951 does myriads of   71->io_seek_request (32768 0) = 0 32768
<youpi> no wonder it's slow
<youpi> unfortunately that's also what it does on linux, the system call is
  just less costly
<youpi> apparently gfortran calls io_seek for, like, every token of the
  sourced file
<youpi> (fgetpos actually, but that's the same)
<youpi> and it is indeed about 10 times slower under Xen for some reason

IRC, freenode, #hurd, 2011-11-02:

<youpi> btw, we have a performance issue with xen
<youpi> an lseek() call costs a huge lot
<youpi> like 1ms
<youpi> while the same costs just a few dozens µs with kvm
<youpi> there's of course the cost of switching between ring3, ring0,
  ring1, ring0, ring3, but still
<gianluca> oh, nice.
<youpi> lseek is supposed to perform only a back&forth
<youpi> and I don't observe disk activity, so it's not waiting for the disk
  to complete whatever atime change & such :)
<youpi> it was mentioned that perhaps xen in hvm mode with pv drivers would
  be faster
<youpi> thanks to the ring3/"1" switching  being done by the processor
<youpi> (and assuming npt)
<gianluca> hm
<gianluca> i'll look into that, sounds fun.
<gianluca> :)
<tschwinge> Here is a testcase:
  http://www.gnu.org/software/hurd/open_issues/performance/io_system/binutils_ld_64ksec.html

binutils ld 64ksec.

Also see the simple testcases test-lseek.c and test-mach.c.

IRC, freenode, #hurd, 2011-11-05:

<youpi> [test-mach.c is] mostly as a reference for the trap overhead
<youpi> 0.56µs (xen) vs 0.48µs(kvm) on test-mach
<youpi> 455µs(xen) vs 16µs(kvm) on test-lseek
<youpi> that might simply be an issue in the RPC  mechanism, which behaves
  badly with the xen memory management
<youpi> yes, about 0.5ms for an lseek, that's quite high :)