On Tue, 12 Feb 2013 21:03:31 +0100 Bertrand Jacquin <be...@meleeweb.net> said:
> D'ar yaou 07 a viz C'hwevrer 2013 e 04 eur 06, « Carsten Haitzler » he deus > skrivet : > > On Wed, 6 Feb 2013 20:19:56 +0100 Bertrand Jacquin <be...@meleeweb.net> > > said: > > > > > Hi, > > > > > > D'ar lun 04 a viz C'hwevrer 2013 e 07 eur 29, « Carsten Haitzler » he deus > > > skrivet : > > > > On Wed, 30 Jan 2013 12:32:23 +0100 Bertrand Jacquin <be...@meleeweb.net> > > > > said: > > > > > > > > while this would be nice... we don't even have our "home turf" vm's all > > > > set up yet. > > > > > > > > 1. we need to drop lvm. qcow+file images. it just makes migrating vms > > > > elsehwere a lot harder if you put them on lvm requiring extra levels of > > > > special setup to use them. > > > > > > I do agree that qcow can be more easy, but here is the trick. We have > > > two cases : VM migration, VM duplication for tests purpose. > > > > > > ** VM duplication : > > > > > > Using libvirt to migrate need one of Fibre Channel, NFS, GFS or > > > iSCSI between the two hosts. We don't have it. I asked to OSUOSL what is > > > their internal process so maybe we can use it in the future. > > > > eh? i literally have just copied a qemu disk image from machine a to machine > > b... > > Oh no, I'm talking about live migration here. > > > and brought up qemu on it and it works out-of-the-box, zero changes, > > migrating etc. it was "scp" and presto. thats why "nfs" for homedirs i > > think is bad - it's not self-contained in the image file - you have to > > transprot your entire setup across which is complex and not self-contained. > > that's why i think lvm just complicates things too. with simple disk files > > we can easily do delta snapshot files to transport "current state relative > > to a baseline" and have the baseline backed up and stored too. u can just > > bring it up with qemu cmdline by hand. dont HAVE to use libvirt. it's > > trivial to do. you can use libvirt too. > > > > qemu-img create -f qcow2 -b base.img delta.img > > > > personally i'm perfectly happy with shutting down the kvm/qemu image once > > per day and doing the above to gen a new daily delta img and backing that > > up (eg scp it home - i can do this trivially - i have the bandwidth and > > space. we can also do it via scping over to e3/e4 too etc.). osuosl plan on > > bringing up a big nfs storage system sometime for backups and data and we > > can use that too when it appears. :) > > On not closed on qcow images, this was just an advise for performances > reason and making things the more possible close to hardware. > > I can move things to a fresh new FS. I'll start on this tonight. ool > > and yes - i know qemu can do LIVE snapshots without shutdown. > > > > http://wiki.qemu.org/Features/Snapshots > > > > also possible. :) > > > > > For the moment there is nothing, so here is the process : > > > > > > = Using LVM : > > > > > > pv /dev/vg-data/e5-template1-sys | ssh osuosl dd > > > of=/somewhere/e5-t1-sys.img <repeat for each disk> > > > > > > = Using qcow : > > > > > > scp /somewhere/e5-template1-sys.img osuosl:/somewhere > > > <repeat for each disk> > > > > > > So here it's the same and not so much complicated for one way or one > > > other. > > > > the files are more obvious. if we need bigger base images it's a simple > > re-generate of img file. lvm is more involved to expand image sizes. :) > > > > > ** Dump to test purpose > > > > > > That's already what we did with Tom and Dan needed a fresh VM : > > > > > > pv /dev/vg-data/e5-template1-sys | xz -f | dd > > > of=/root/e5-template1-sys.img.xz > > > > > > In all case, the source file is just different, so it's not higly more > > > complicated, it's the same. > > > > > > Using FS over FS in not really efficace, plus, you need to care of the > > > multiple sync that append on the VM FS, on the host FS, on the host > > > RAID, on the host disks. Using LVM, there is one less layer. > > > > this isnt a "mission critical" system. we will have backups. if we loose > > half a day of stuff because of a disk failure.. so be it.. we lose it. go > > back to last snapshot/backup. with git we have a massively distributed > > backup/copy of all the src which is the most important thing we have, so we > > are good imho. > > > > > Also, all resize feature etc are still available. > > > > > > I do think performance is a criteria as we can observe on e2 for now. We > > > should avoid to reproduce a such situation at the very begenning to not > > > be surprise in some times. > > > > i am sitting here with a totally unoptimized ubuntu 12.04 image in qemu. i > > actually don't NOTICE a slowdown compared the to the 12.04 host when it > > comes to I/O. in fact... its FASTER than the host > > I can agree with one VM with no I/O or high CPU usage. > > This is not the case when (for now) e5-buildbotslave1 is using a lot a > CPU, you can feel the effect on this on other VM. i'm not sure here. i've run efl builds in a simple qemu + kvm vm with qemu disk image and my os didn't even notice. well not beyond it being like a native compile. that was on a desktop with 8g ram and a single i5. :) > > - by a wide margin, when it comes > > to boot for example. once cached it boots in about 7 seconds all the way to > > a gui (e17+xorg+etc. with all the ubuntu overhead behind it too). the host > > takes something like 20-30sec to boot... you currently say that it takes > > about 5 sec to boot your gentoo vm's.. thats without e17+xorg+the rest > > ubuntu slaps on top. this is a simple core i7 desktop with only 3.5gb of > > available ram (32bit and your regular spinning hdd... i see no issues here. > > > > what we need to avoid is things like mysql that hammer the disk with syncs. > > svn +apache are a lethal combination easily consuming 2gb of ram when > > someone checks out our svn tree via http. and trac in general just loves to > > consume cycles. > > Yep, this is why the two 'applications' are splitted on 2 differents > VMs. actually trac+svn and git+phab are split because we want to be able to just kill off trac+svn once everything is moved over (though trac will likely hang around due to its bugtracker having content we wont move over and wiki etc. for a long time - but it'll be shut down and "read only" at some point). > > remember e2 is a 6+ year old machine, with a raid controller on the > > blink, 2g of ram and a fairly old cpu... e5 is beefy on a bad day. 48g of > > ram is one really nice disk cache. > > That also make more differences in case of crash, but yes we can re > import a fresh backup. Also, we can use qemu live snapshot. yup. at the most primitve it's simple to deal with, we can still also use live snapshots etc. too. ;) > > > Also, yes we have a lot of RAM and CPU on e5, that's really fine. But > > > I/O are absolutely not better with a lot of RAM or CPU. > > > > that's why i'd rather not FORCE it to be I/O - let it go to ram. this is not > > "mission critical". if we "lose a transaction" - too bad. someone's bug > > report disappears. for git - this is not a problem as its a distributed > > system with every client having backups. the person who did the push has > > the data - we didn't lose it. we can push it again in this rare event. we > > can actually minimize this simply by running fsync on the host every now > > and again to ensure syncs to real media - but dont force this per > > transaction. > > > > i understand your point of view. it's one of "this is a mission critical > > system and we must not lose anything ever". but we can afford that - thus > > we can make different trade-offs. ease of administration. obviousness. ease > > of migration of vm's, ease of setting up new things ad-hoc when needed and > > simplicity in general are worth more imho. the experience of 6+ years of > > e.org tell me to "keep it simple". > > Things are 'simple', this don't need any magic stuff, they are basic > (well maybe many basic stuff) accumulated to make a complete solution, > but it's not complex. Making things different from what you are use to > does not make this 'complex'. nfs raises complexity - you can't move vm's around as they are tied to an nfs service that then also needs moving. we need to run extra tools to "Extract" from lvm to tar-up/package/ship vm info, as opposed to just tar it up and send it off. :) > > what happens when you get run over by a bus? we don't have > > full-time admins following your every setup and reading manuals. they look > > at it once and then use it. > > I absolutely agree, I don't want to have all this stuff on my shoulders. > I try to comment all the config files and script to let people the > knownledge and the philosophy. indeed. and much appreciated. :) > At the begenning I wanted to write some docs but it seems people doesn't > really want to read things ;) something i learned over the years with e and efl. write docs all u want.. people dont want to read. and to be honest - i see why. it's just extra time needed to get something - anything done. make things simple and obvious is best. if we had to deploy 1000+ e.orgs and manage them day to day, scale them up and make this installable and so on by 10000's of people around the word, then it'd be different. but we really have enough hw now to live on for a long time, and don't need to scale out or up really. :) > And you can be sure I'm not ready to handle all this for the next 10 > years. :) > Spending time to make a simple solution (this is the case for me) take > really much more time than build a complex one that doesn't have a > global philosophy and/or architecture (differents distro, different web > servers, non common config files ...). > > I think some other can agree on what i'm saying. well having a single base "already installed vm" and then just adding to it the services needed works well. it saves time. lvm imho just makes things hidden behind another layer of tools (i have an lvm raid5 array of my own at home on my media server - so i have dealt with it before. and my experience is, when i have to deal with lvm once every few years, getting it to "work" again means reading up manuals every single time to try get it to scan and recover my arrays. for people who dont touch lvm regularly, it's a "wtf" step of "wehere the frig are the disk images?" and then having to find out 3 years form now what has been done given a toolset they are not familiar with and have to read up manuals on. :) until you sent that mail i was going: df -h eh? wtf? where did the raid10 array go? where is all the data? eh? wtf? it was non-obvious where all this stuff had gone to. in 3 years i have to try remember this and dig out documentation from somewhere and then read specific kvm configs to hunt down which lvm volumes belong to what with lvm. as opposed to something simple like: df -h and finding 500gb has been used by /media and i look: /media/vms/phab+git/... /media/vms/trac+svn/... /media/vms/www/... and in each dir is a nice simple "rootfs.img" and maybe a "start-vm.sh"... (and a suspend-vm.sh, snapshot-vm.sh, stop-vm.sh, ssh-in.sh)... and then i go "awesome!" and know instantly what to do with no docs. :) it's self-documenting. everything to do with a vm is located in a single named and easy to find dir. if there are qcow deltas they are in there too, nicely named by time/date or version # etc. :) btw - you do not need to do exactly as above. it's an example of "an obvious setup" that is easy to figure out and deal with N years from now. if there is a firewall in addition its yet more stuff to remember exists and read up docs on where its configured, how to fix it etc. > > in 3 years something happens to you (the bus, or you > > get bored of e, or are just on a long holiday for 2 months?) and someone > > else needs to "fix things"... if what they find is incredibly complex, tied > > together deeply and they haven't spent day in and out adminning the box, > > they can't figure out what to do and eventually just ad-hoc make a mess > > trying to get things back up at all. you may be back after your holiday > > only to find things are now messed up, or we spend longer "down" than > > needed etc. SIMPLICITY trumps performance here. GIT solves the most > > important side of data integrity by its very nature. the rest is "nice to > > have, but not critical". if we have to roll back 1/2 a day or a day of > > data... we'll live just fine, unless this is happening every few days.. and > > then we have a major serious problem. > > > > > > 2. nfs homedirs are a no-go. homedirs must/should be local to the vm fs > > > > image. making them nfs makes them non-portable. the whole POINT of using > > > > qemu/kvm in the first place is, that if we get overloaded, or if e5 goes > > > > down, or whatever we can TRIVALLY just move any vm over to osuosl's vm > > > > cluster system and re-point dns and presto... it works. this is WHY we > > > > want a SIMPLE setup. not complex. SIMPLE. > > > > > > Yep ! NFS is here to speed up thing for now, it's temporary and will not > > > stay. Never wanted to. > > > > cool! :) again as above - speed.. don't care. :) the qemu -fsdev/virtfs > > option is more "sane" imho if we want to have arbitrary expansion space > > outside the fs image - this means we can have base fs img + delta + dir of > > "extra files" (tar/rsync that dir as needed) > > I mean setup time, not the final usage. NFS is temporary while the > new users script is not complete. aaah ok. > > > > 3. firewall needs to go - NAT is enough. > > > > > > NAT is not firewalling, hope we agree on this. > > > > i know.. but it provides protection in that "you can't get in through the > > nat unless something explicitly forwards a port". getting in is a reality > > we will have to live with unless we want to ask for an IP for every VM... > > and that's asking a lot of osu to go vie us so many IP's. > > So you want me to ask to OSUOSL one IP per VM ? Is it that ? This is a > rare ressource nowadays. Well, we can have a complete IPv6 solution, not > sure it's the solution here ;) oh hell no! no no! thats why i know we need to rely on PORTS to split services. asking of an IP per VM is just nuts. :) we don't have that kind of luxury. we have to live on very few of them. we would need to rely on osuosl to either provide ip's if we move things to their cluster OR for them to provide some kind of port forwarding and http routing. that's reality. as long as our vm's are simple self-contained things, this is easy. :) > > > > firewall just wastes peoples time > > > > debugging services. > > > > > > We should not have to debug anything when thing are up, when we deploy > > > new since I agree, it's normal. And this does not happend so often. > > > > ask daniel. he already spent hours figuring out a problem that was the fw > > blocking OUTGOING traffic. we JUST need a NAT. we all know the NAT will > > provide incoming connections unless we explicitly route them. we "expect" > > outgoing connections to "just work" (tm) with nothing in the way. > > This is now done. But i'm not really a fan of this. All is logged in all > cases. sure - understood that it disturbs you :) but frankly it's easier. :) > > > > in all the years e1/e2 have been up we have nver needed or > > > > uses a fw. > > > > > > And for 4 years, MySQL has been opened to whole internet. This just > > > prevent mistakes. > > > > we're behind a nat... that already does all the work we need. the host > > system (e5) has nothing running anyway, so doesn't matter. > > and NAT also doesn't append on e5, it's on e5v1. Access to all e5v* is > bridged to e5v1 with no firewall/NAT, all is done on e5v1. ok - so e5v1 bounces to a natted internal network for vm's... why e5v2 then? other than to put ssh "on a normal port"? :) -- ------------- Codito, ergo sum - "I code, therefore I am" -------------- The Rasterman (Carsten Haitzler) ras...@rasterman.com ------------------------------------------------------------------------------ Free Next-Gen Firewall Hardware Offer Buy your Sophos next-gen firewall before the end March 2013 and get the hardware for free! Learn more. http://p.sf.net/sfu/sophos-d2d-feb _______________________________________________ enlightenment-devel mailing list enlightenment-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/enlightenment-devel