1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
|
[[!meta copyright="Copyright © 2011 Free Software Foundation, Inc."]]
[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable
id="license" text="Permission is granted to copy, distribute and/or modify this
document under the terms of the GNU Free Documentation License, Version 1.2 or
any later version published by the Free Software Foundation; with no Invariant
Sections, no Front-Cover Texts, and no Back-Cover Texts. A copy of the license
is included in the section entitled [[GNU Free Documentation
License|/fdl]]."]]"""]]
[[!tag open_issue_xen]]
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
[[performance/io_system/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 :)
|
