***: ameise has joined #arpnetworks
mkb_ has joined #arpnetworks
milki_ has joined #arpnetworks
mhoran_ has joined #arpnetworks
ChanServ sets mode: +o mhoran_
karstensrage_ has joined #arpnetworks
mhoran has quit IRC (*.net *.split)
ant has quit IRC (*.net *.split)
karstensrage has quit IRC (*.net *.split)
mkb has quit IRC (*.net *.split)
milki has quit IRC (*.net *.split)
ameise is now known as ant
mkb_ is now known as mkb
karstensrage_ is now known as karstensrage
karstensrage has quit IRC (Changing host)
karstensrage has joined #arpnetworks nathani: ix33: can you ping6 anything? ix33: nathani: garry got it
it was a setting on a new thunder instance
VLAN
now, if someone would just fix vioblk on openbsd... mercutio: ix33: openbsd's disk subsystem in general isn't great :(
it really needs an overhaul but i haven't seen anyone looking to spend a lot of time on it ix33: no kidding
untar of src.tar.gz on a physical 2-core box w/ssd in openbsd is ~100s mercutio: it's kind of embarassing, like when openbsd had no smp :) ix33: which is awful mercutio: with or without softdep? ix33: good question mercutio: i think slow untar with openbsd is to do with poor write behind ix33: nosoftdep is default, so that mercutio: i'd try comparing with sofdep then ix33: well that's a physical host
same default setup on a VPS instance is ~600s mercutio: ssd's aren't instant for acknowledging writes normally
they're faster than hard-disk, but still..
if there are a lot of files it can add up
so if openbsd can keep writing stuff without waiting then performance can e higher
openbsd 6.0 src.tar.gz ? ix33: mercutio: yes mercutio: i want to compare ix33: i know that's a ton of little, highly compressed files... mercutio: yeah ix33: going to put freebsd on this thing just to compare & feel depressed about it
i mean, 100s on the host i'm referring to is absolutely abysmal mercutio: i got 1 minute 2 seconds on a vps
with softdep though ix33: yeah rebooting to try that
(/)
theo doesn't make fun of people for using softdep, does he? mercutio: no idea
i remounted with softdep
it seems you can do it live, at least it doesn't complain ix33: -o remount works? mercutio: mount -u -o softdep /
1 minute 50 seconds withotu softdep on vps ix33: kvm/vioblk? mercutio: yeah ix33: which version of openbsd? mercutio: 6.0 ix33: something doesn't add up mercutio: how so?
you weren't using softdep
i got 1 minute 50 without softdep, 1 minute 02 with softdep ix33: <@mercutio> 1 minute 50 seconds withotu softdep on vps
wow 91s on the slow FS with softdep mercutio: softdep slowed it down for you?
hmm, i have noatime too it seemms ix33: no softdep sped up
from 800s on that guy i think mercutio: i thought you got a minute without softdep ix33: no that was the physical host
i'm kicking a VPS at the moment
90s on the VPS with softdep is <3 <3 <3
testing on metal too now mercutio: what is sad is how much faster linux is :( ix33: no kidding
like 5s on a haswell w/ssd
btrfs even :/ mercutio: i got 5.2 seconds
dman you are faster than me ix33: lol mercutio: hey it used more than 100% cpu
tar vxzf ../src.tar.gz 2.94s user 3.04s system 114% cpu 5.222 total ix33: 13s on a big zfs fbsd machine too mercutio: oh this was zfs on linux ix33: ^^^^opterons mercutio: i think it's haswell too actually
yeah i7-4770 ix33: haswell i5 here mercutio: same diff basically
i have lz4 enabled too
but i doubt it makes any significant diff ix33: you know i saw the damnedest thing, speaking of this opteron mercutio: but yeah i wish openbsd's disk subsystem would be fixed up ix33: the default cipher on ssh/scp is 70% the speed on the opteron as on a haswel-era pentium d
both machines capable of gig-E wire speed (verified with ioperf)
both freebsd 10.3 machines testing to the haswell linux box mercutio: cypher can make a huge diff on some machines
i don't think pentium d's have aesni
but on such capable machines that can make cpu utilisation of such vary a lot
except openbsd doesn't trust it hah
i think there's a new djb one now
chacha20 ix33: yeah
i was about to tear that machine apart and test new NICs
anyway, thanks for the pointer mercutio mercutio: you're welcome ix33: the stuff i've been googling for has been virt* specific
i didn't look for more general tunings because i was OK with the base-case suckage mercutio: yeah i don't think it's acutally a virtio issue ix33: garry would be happy to hear that mercutio: openbsd is just quite far behind in disk performance ix33: since i opened a ticket on it mercutio: there used to be a huge deficiency with smp too
that's getting better ix33: $ doas /usr/bin/time tar xzpf src.tar.gz 501.06 real 3.78 user 11.95 sys mercutio: like the global lock was around basiclaly everything, where smp was only userland
it's getting a little finer now ix33: <3 openbsd
i'm ok with what it is
but some times it's a head scratcher mercutio: heh ix33: ^ so that was with a remount of /, which didn't appear to honor the -o softdep mercutio: the time seems to work but mount doesn't seem to update ix33: ditto the physical box too mercutio: you could edit /etc/fstab and reboot and check it's still good ix33: right
i shall
can't reboot that physical box till after hours tho mercutio: you have an arp vps right? ix33: yes
a few mercutio: was it migrated/recent? ix33: this was me testing the new thunder setup
very recent mercutio: ahh so that's on arp thunder ix33: yes
i'm ok with crap performance on onesy-twosy small VPSes
i was actually considering freebsd on the thunder instance if i couldn't make it behave better
:/ mercutio: and 90 seconds on softdep? ix33: it's running now (after reboot)
6.0-current even mercutio: cool
6.1 seems to be delayed.
but i don't think it'll improve performance ix33: well i have just upgraded from 5.7 on everything, so i wasn't current on what's what
$ doas /usr/bin/time tar xzpf src.tar.gz
115.48 rel 2.74 user 5.91 sys
daaaaaaang
i'll take it
if i'm untarring with -p, does it modify atime too? mercutio: maybe
i didn't use p ix33: that's smoking fast considering where i came from
thanks! mercutio: and i have noatime on
heh
it's not too bad with cvs updates and the like tbh
like openbsd's disk performance used to bug me a bit on synthetic things like that
but for most things that matter like waiting for a cvs update it doesn't seem that bad
also openbsd kernel compiles are much faster than linux kernel compiels :) HAS_A_BANANA: I dunno, I compiled a Linux kernel (defconfig) on an SSD with -j12 in 19 seconds...
Hard to beat that. mercutio: i think openbsd is less than 2 minutes for default with one core ix33: HAS_A_BANANA: holy cow, with what?
(CPU)
i admit i had an amd thunderbird last time i compiled a kernel for fun... HAS_A_BANANA: I don't really remember, it was a couple years ago... I think an i5-2400. mercutio: are you sure it didn't fail ix33: lol mercutio: with -j12 you might not see the failure so easily
i'm compiling openbsd kernel right now :) HAS_A_BANANA: Yes I was sure, mercutio ix33: make defconf; make bzImage should pretty much always work assuming your env is sane, right? mercutio: well you need as86 etc ix33: well i feel better now. it's champagne friday. thanks all! mercutio: but yeah you'd think it'd be a tested option ix33: (even commented on my arpnetworks ticket to that effect) mercutio: damn openbsd kernel compile is slower than it used to be
make 213.31s user 35.44s system 96% cpu 4:17.53 total
trying with make -j 3 now
actually i think gcc has been getting slower over time
make -j3 211.87s user 39.95s system 186% cpu 2:15.26 total dne: openbsd uses quite old gcc though mercutio: it's recent compared to 2.95? dne: I guess :)
4.2.1 - only 10 years old mercutio: oh wow time flies
how old is 2.95 then? :)
16 years
wow dne: 1999 mercutio: oh hangon, july 31 1999 for first release
it looks like it was 10 years ago they abanonded gcc 2.95
looks like openbsd 5.5 was the first release to not include gcc 2.95.3 though
may 1 2014 ***: Nahual has quit IRC (Ping timeout: 258 seconds)