about libmount library performance problems
Subsystems (3)
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
I have 1 question which need confirm.
in the RHEL7.0alpha version
nfs-utils-1.2.5-3.el7.x86_64 --> use mount_libmount.c util-linux-2.20.1-2.el7.x86_64
when umount a NFS directory, the mnt_table_parse_mtab() function will merge user options(/run/mount/utab)
into mountinfo(/proc/self/mountinfo)from kernel. if the mount number is large, leading to umount a directory time-consuming. the complexity of the algorithm is O(n(n+1)/2).
but in the old version(RHEL6.3GA/nfs-utils-1.2.3-26.el6.x86_64(use mount.c)), only according to the parameters of umount command,
find the corresponding record in the /etc/mtab file
Question)*:* merg the tow file are necessary for umount a dir?
TEST LOG:
#cat mount_test.sh
#!/bin/bash
mount=/tmp/mount
if [ ! -d $mount ]; then
mkdir -p $mount
fi
for dir in {1..8000}; do
if [ ! -d $mount/$dir ]; then
mkdir -p $mount/$dir
fi
echo mount $dir
mount -t nfs 127.0.0.1:/ $mount/$dir
done
# time umount /tmp/mount/1 --------->RHEL7.0alpha real 0m2.395s user 0m1.640s sys 0m0.466s
# time umount /tmp/mount/1 --------->RHEL6.3GA real 0m0.159s user 0m0.042s sys 0m0.114s
thanks.
Karel Zak
maintainerQuoted 8 lines · chenditang
I have 1 question which need confirm.
in the RHEL7.0alpha version
nfs-utils-1.2.5-3.el7.x86_64 --> use mount_libmount.c util-linux-2.20.1-2.el7.x86_64when umount a NFS directory, the
mnt_table_parse_mtab()function will merge user options(/run/mount/utab)
into mountinfo(/proc/self/mountinfo)from kernel. if the mount number is large, leading to umount a directory time-consuming. the complexity of the algorithm is O(n(n+1)/2).
Yes, maybe we can somehow improve the file format and the algorithm, suggestions?
Quoted 2 lines
but in the old
version(RHEL6.3GA/nfs-utils-1.2.3-26.el6.x86_64(use mount.c)), only according to the parameters of umount command,
find the corresponding record in the/etc/mtabfile
But we don't have mtab file in RHEL7 anymore, so we have to store the userspace specific mount options somewhere... (/run/mount/utab). The solution is to minimize number of information we have to maintain in userspace.
Unfortunately it seems that for example on my system we duplicate the information:
utab: retrans=20,addr=192.168.111.1
mountinfo: rw,vers=3,rsize=131072,wsize=131072,namlen=255,hard,proto=tcp,timeo=600,
retrans=20,sec=sys,mountaddr=192.168.111.1,mountvers=3,mountport=39755,
mountproto=udp,local_lock=none,addr=192.168.111.1(see retrans= and addr=)
Maybe we can detect this situation and don't store duplicate options in userspace. The question is what happen with retrans= and addr= after remount. If I good remember there was any reason to maintain these stuff in userspace. Chuck?
Quoted 1 line
Question)*:* merg the tow file are necessary for umount a dir?
I guess nfs umount needs the addr= info.
Karel
Signature · 2 lines
Karel Zak <[email protected]> http://karelzak.blogspot.com
Karel Zak
maintainerQuoted 2 lines · Karel Zak
Quoted 1 line · chenditang
Question)*:* merg the tow file are necessary for umount a dir?
Ah, do you mean -- why we need to merge all the entries when we want to work with only one mountpoint? That's good question ;-) I'll try to improve the code...
Karel
Signature · 2 lines
Karel Zak <[email protected]> http://karelzak.blogspot.com
Karel Zak
maintainerQuoted 6 lines · Karel Zak
Quoted 2 lines · chenditang
Quoted 1 line
Question)*:* merg the tow file are necessary for umount a dir?
Ah, do you mean -- why we need to merge all the entries when we want to work with only one mountpoint? That's good question ;-) I'll try to improve the code...
Improved:
old version:
time ./umount /tmp/mount/2255
real 0m1.254s
user 0m1.002s
sys 0m0.238s
new version:
time ./umount /tmp/mount/2244
real 0m0.332s
user 0m0.111s
sys 0m0.218sKarel
Signature · 2 lines
Karel Zak <[email protected]> http://karelzak.blogspot.com