milki: i'm about that too i think :) 7.1 gb in the last month beta VM guys should re-test disk I/O. I've disabled my munin graphing temporarily (turns out it was munin going crazy, not cacti) munin will wind down in about 5 minutes and so now there shouldn't be big load spikes w00t, sponsored reddit link: http://www.reddit.com/comments/11axe2/arp_networks_offers_freebsd_vps_services_with/ i can haz upvotes?:) toddf: are you compiling something or doing something heavy on your beta VM? (not a problem, just wondering cuz i saw it taking a bit of cpu) anyone know if a munin master at 2.x can be used with munin nodes that are still just 1.x ? i want to move to 2.x for performance on the master, but i'd hate to have to also update all my nodes up_the_irons: I've always had master and nodes on same train. It should work though. josephb: roger up_the_irons: tun/tap support, anything special needed? josephb: nope josephb: pretend it is a bare metal machine that's all yours ta, just trying to get ipip or gre tunnel going back to home so my vps can monitor internal things josephb: shouldn't be an issue. i see gre traffic here and there, so i know other people are doing it too ;) unfortunately, i must dual boot into Windows now (radio programming)... upvoted up_the_irons up_the_irons: abusing it to test opensmtpd .. I think its wedged .. resetting I started a test on friday, I guess it kinda went bezerk. UVM: pid 30715 (procmail), uid 547 killed: out of swap doh yeah, until 'recent' commits this weekend, if you did 'echo "|sleep 60" > .forward' and delivered a bunch of tiny mails you'd end up with an out of fd exhaustion scenario where it falls flat on its face ... local deliveries happened 'quickly' in any prior tests, but since I'm delivering mail via procmail and piping it through a dspam filter .. it takes a few seconds and the more that pile up, the slower it goes, circle of death ensues, etc ;-) now there is a std default MDA limit per user and global for the system, to be tunable after the upcoming release of opensmtpd wew i am still on a mission to find a usable android emial client that isn't gmail stupid conversationviewbeingawesome qbit: 'usable' is in the eye of the beholder. ever consider writing one if you're searching in vain? toddf: no doubt.. but building a java app of that scale is out side of my knowledge-time threshold. even hacking on existing open source mail clients :P toddf: There is a reason why the traditional MTA design was to queue emails at all stages and have a queue runner to process them later plett: are you suggesting smtpd does not do it that way? plett: it does, infact, queue all mails and then process it later, and there is a mda (mail delivery agent) process that you might consider a queue runner to process them later. it simply did a 'forkmda()' call without bounds. that is now fixed ;-) toddf: Ahh. I didn't know opensmtpd did work that way. I read your initial comment to be that it spawned deliver processes for each email without considering how many other deliveries were already happening apparently I was the first one to test with a procmailrc that took more than <1s to deliver ;-) ol l Is there a link on how to configure the parameters of the network device? (I'm manually installing OpenBSD) man hostname.if(5) and if you don't have an openbsd host handy see http://www.openbsd.org/cgi-bin/man.cgi?query=hostname.if *everyone* should have an openbsd host handy. :) bah, just use FreeBSD and get the best bits of OpenBSD :) there's also the http://www.freebsd.org/cgi/man.cgi which has *all* the manpages nice resource I'm sorry, I mean the parameters from arpnetworks to configure my network device (ipv6address, ipv4address and else) warex: all of that's in your portal Notably, under "IP" oh... sorry! thanks! is it needed to disable mpbios so OpenBSD can start its em0 networkservice? I'm trying a fresh install but em0 sends a watchdog watchdog timedout and freezes just when trying to start the network services yes mikeputnam: tnx! toddf: ah ok qbit: what about k-9 mail? that's what i use I also use K-9. It's certainly no mutt, but it's the best I've found. up_the_irons: i wish i could set colors for specific headers and thread things by converstaion qbit: You mean the three long bolts? conversation too CaZe: ya I think I tighten them quite a bit tighter than he recommends. I never did describe the actual alignment process, did I? nope :D Well, like I started saying, you have to first think about why being out of alignment causes the axle to bob up and down. Do you know why that happens? lol, because it's not aligned. :P - same concept as a bent axle No, the axle is straight. At least, it should be. If it isn't, then you need a new axle. i just mean in concept - not sure how else to explain it other than that And a bent axle doesn't bob up and down parallel to its shaft anyway. oh then i don't know :D Well, as Doug describes, the inner burr isn't perfectly round. It's slightly oval. Do you have your Pharos in front of you right now? yep and if i spin the axle when it's upside-down the axle bobs Well that wouldn't happen if it were aligned, because even if the inner burr is oval, if it were perfectly centered, then the two "far" points of the oval would be scraping the outer burr, counterbalancing one another. so it bobs because of the oval-ness of the inner burr? It bobs because it's out of alignment. If it's out of alignment, then that means the inner burr isn't centered. so the burrs are hitting when it's bobbing The burrs are always hitting when you have it upside down. The burrs are the only thing that keeps the axles from sliding out of the apparatus and falling on the floor. s/axles/axle/ makes sense Having it upside down puts the burrs at "zero", as Doug calls it. k so you loosen it and spin the axle to get minimal bob So when the inner burr is off center, one of the "far" points of the oval is scraping the outer burr, without being balanced by the other "far" point of he oval. Keep in mind that the outer burr doesn't have straight walls. It's a cone. right So as the far point of the inner burr comes into contact with the outer burr, it wants to try and push into it. But since it's a cone, to the inner burr, it looks like a ramp. So the inner burr basically creeps up the "ramp", bringing the entire shaft assembly with it. so shift the outer burr to minimize "ramp" It's basically the way that the far point of the inner burr pushes through the wall of the outer burr. It seeks the path of least resistence, and that means creeping up along hte wall of the cone, where the diameter of ther outer burr is larger. Well the ramp is always there, you just want the far ends of the oval to be balanced against one another. up_the_irons: you around? qbit: It might be easier to visualize if you take out the burrs and hold them relative to one another in space. think i have a pretty good visualization Actually, let me revise that. I just found my notes. I was describing to you an earlier version of my theory. There doesn't have to be any balancing. In fact, the inner burr doesn't have to be an oval. It could be (and probably is) an ellipse. Where one of the far points never touches the outer burr. As far as our model is concerned, the inner burr can be modeled as just a single stick protruding perpendicularly from the shaft. If you imagine the stick to be a piece of chalk, then it draws its path on the surface of the outer burr. k When it's perfectly centered, the path is a level circle. right When it's off center, the path it traces is an ellipse. It climbs higher along the cone, then drops down to a lower point. http://csep10.phys.utk.edu/astr161/lect/history/conic.gif Compare circle and ellipse in that diagram. The cone is the outer burr, of course. aye The motion of climbing higher on the wall of the cone and then dropping down again is what causes the bobbing motion of the axle. yep When it's centered, the path is a level circle that has a constant height along the cone, so there is no bobbing. yerp If you do as Doug says, and leave the three long bolts loose, and just rotate the handle, then it self-aligns. that's how i did it last time tightened ~ to his spec.. but it's wobbled loose Since the outer burr is free to move around, the inner burr just sort of pushes the outer burr into place, since that's easier than climbing the wall and moving the entire shaft up along with it. Well you probably want to tighten it more. But the problem when you do that is that overtightening actually puts it out of alignment. And I think that's why he tells people not to overtighten. Now, we have to investigate *why* overtighening puts it out of alignment. because the middle plate moves the outer burr * moves it's angle Yeah, sort of. The appartatus isn't a perfectly rigid structure. As you tighten, things bend slightly. Enough to put it out of alignment. k But if you only tighten one bolt at a time, and check the alignment before you do anything else, then you *know* for sure how it was distorted. And fortunately, he only used three bolts, which makes this process easier. If he had use four, then it would not only be a longer process, but more complicated. So when you tighten a bolt, it pushes the steel plates closer to one another on that side. It has the effect of tilting that side of the outer burr towards the handle. which may or may not be a good thing.. depending on the ellipsieness Well, as far as alignment goes, what effect does tilting the burr have? And how could you correct it? seems if you loosen the bolt by the "ramp" and tighten the others .. it should reduce the ramping Yes, but then you have one bolt loose. right - but i don't mean all the way And how do you know that tightening the other two bolts won't put it out of alignment in another direction? by respinning But that won't work once it's torqued down. That's the whole point of overtightening - so that *won't* happen. Otherwise you'll be realigning every week. Tightening a bolt has the effect of lifting the side of the outer burr that the bolt is on upwards, towards the handle. As far as the "ramp" goes, it's the same as moving the outer burr closer to the inner burr. Since the ramp is being lifted, the "stick" representing the inner burr is now touching a lower part of the cone. Actually hold on, let me visualize this to make sure I have the direction right. Maybe I should draw a picture. heh right on Hmm, I have to go make dinner. Anyway, just tap the burr on the opposite side of where you tightened the bolt. Since tightening that bolt reduced the clearance, shifting the outer burr in the direction of that particular bolt will increase the clearance. And just do that for each of the bolts. (You'll have to tap harder as you progress through the bolts.) It's possible to tap it too much, and it's tricky to tell when you've done so. But since the configuration of the three bolts is a triangle, pushing the outer burr too far when you're tighening one of the bolts is the same as putting it out of alignment with respective to the two other bolts. And you can correct for that when you visit the two other bolts. qbit: ^ rad thanks :D have a good dinner! up_the_irons: you there? Webhostbudd: yes, but busy oh alright quick questions, no problem; need me to do something, gotta wait :) I'll just submit a ticket then roger Channel poll: raw backup space (rsync) to be available for $0.20 per GB, good price? A bit cheaper than Linode and 30% cheaper than RootBSD hmmm that sounds pretty good $0.20 per GB per month? yeah yea i like that so if you had a small plan ($10) with 5GB disk, you'd pay only $1 more for enough space to fully rsync all that but im guessing you can rsync just individual files too no one is going to sync their entire fs.... at least, i would hope not yes, that's the idea, you run rsync within your vps i think i'd build that backup host with FreeBSD + ZFS mirror. Initially with just a 2 disk mirror to get it off the ground. I already have a blade ready for it.. chroot home dirs, like ryk suggested yea definitely up_the_irons: you have a blade server? thought you use supermicro twins or super twins or whatever they're called now ryk: i do have a blade server, not the twins food time lol up_the_irons: I'd be up for it. hot up_the_irons: raidz probably makes more sense? up_the_irons: i wonder if it's possible to do some kind of lvm snapshot copy type backup... which you could "recover" from if you wanted to the problem with mirror is that if you want to grow it you'd have to add two more disks, or add a raidz on .. as zfs doesn't allow changing from mirror to raidz raidz would not be a good choice for this, imo 1) it's backup 2) raidz is not easily expandable now, you can keep adding raidz devices to your zpool, but you have to add an entire raidset every time you want to expand it. mirroed you have to add a pair each time raidz you have to add 3 disks every time err 3 or more and of course you're going to want to do more, because eww 3-disk raidz? so that gets expensive. what's wrong with 3 disk raidz? what's wrong with a mirror vdev? 3 disk raidz gives you twice as much storage space as mirror it also gives you faster write speeds but if you know up_the_irons , you know he's a raid 1- guy *10 it's a bit different for backup space to virtual machine storage there's no requirement for speed, here. i'ts backup space. well yaeh, most of the time there isn't but if lots of people cron at the same time their backup.. and someone tries do a read at the same time or noties their backup taking ages raidz gives higher write bandwidth i mean, you're right -- if it had to be parity raid, i would choose raidz over raid5 but yeah, even for normal performance generally speaking, i'd say that 4 raidz3 set's stripped would be competitive with raid10 with 12 disks raidz3? or 3-disk raidz? although when getting to 12 disks the chances of dual disk failure increases i thought raidz3 was super-redundant 3 disk raidz sorry, confusing terms a little :) i don't think there's a raidz3 there's raidz2 there is, now, it's new-ish generally speakign above that i think people say you should have more than one data set or more than one raid volume err more than one pool i'm confusing terms again anyone the rational was if you have 3 2tb hard-disks with raidz you get 4 tb of disk space, if you have 4 2tb hard-disks with 2 pairs of mirrored drives you get 4tb of disk space with 6 disks.. raid-z gives 8 tb .. raid mirrored gives 6 tb i suppose it's not a big diff it'd be better if you could grow storage nicely like just chuck another 2tb in and rebalance so, having never used raidz, i haven't measured write speeds. you're saying write is faster? because that's contrary to conventional wisdom i would associate with parity raid read is usually faster, write is slower due to the calculations calculations are minimal bsaically it can write half the data to one disk half to the other as well as the parity so if you write 100 gig of data you write 50 gig to 3 disks whereas with mirrored raid you write 100 gig to 2 disks an thus you write 150 gig versus 200 gig and read? read performance is similar for sequential the problem with raidz is for lots of small reads you reduce your seek performance, because say you have 128k stripe size per disk then you'd have to read 128k off each disk for even a 4k read limiting seeks to that of a single disk with backups though they're by far sequential so it can just read huge chunks off each hard-disk or write big chunks to each hard-disk but y eah when people say about the overheads of raid parity calculations it's way less significant than aes (https, ssh, etc) sure it may have been a significant overhead in 486 days or more so in sparc days err early sparc days but it's not a significant concern. aes is more likely to be a concern if lots of people scp/ssh at once but even then if you have multiple cores without hardware acceleration it should be fine anyway also another thing to keep in mind with raid10 versus basically raid 50 is that most work loads cache the majority of their reads but you can't really cache writes, you can buffer them... but sooner or later you have to write them to the disk whereas reads just can from memory a lot of the time so for things like hosting web sites, it's mostly a "write load" not a "read load" because of appending the log file with updated accesses etc although it's also mostly small writes but if you can't keep your most visited documents in memory, and are frequently going to disk, then either you're starved for memory, or you have a huge dataset and if you have a huge dataset and it's text files, it may make more sense to compress your memory cache apache traffic server is an example of a proxy server that can compress it's memory cache i don't know why it's not used more often, it's faster than varnish or squid or ngnix ngnix apache traffic server? it was a yahoo project that went open source and was made part of apache's greater uhh whatever it's called it's multithreaded, high performance, under active development up_the_irons: Another thing that might be nice is a domain for customers. You know, for those of us who don't care to get our own. Those free dns services are a little annoying because you have to update the ip evey once in awhile, even with a static ip, or else they drop you for inactivity. qbit: I'm investigating another possible reason for why tightening the bolts puts the burr out of alignment. qbit: It has to do with the type of nuts that are on the end, and how they grab the bottom plate as they're tightened. If the long bolt is not perfectly straight through each plate, tightening the nut will straighten it. But if the teeth in the nut have grabbed into the plate, then the position of the plate will shift with the nut. I'm going to see if adding a washer will change anything. But it'll have to wait until tomorrow. right on seems that if you gighten from the top - and keep the bottom in one position it owuld hav ethe sameish effect Maybe, but it still wanders a bit. btw, I got rid of those four smaller screws holding thd outer burr. And the springs, too. interesting whY? cpu hs mounts? Because I think they just get in the way if you're planning on overtightening. They certainly don't add anything. true HighJinx: Coffee grinder. ahh If you're going to tighten to the point that the springs bottom out, then there's no reason to have them. And the only reason the four small bolts are there is to keep the spring in place. has anyone had success with virtio on kvm? errr freebsd virtio modules on kvm I have been working on some web based DNS tools, if anyone would care to provide some constructive criticism: http://dns.winvive.com mnathani: pretty array formatting :) lol gotta love print_r ;) :-) staticsafe I like it tbh, simple and clean Odd... according to your site, mnathani, google.com isn't DNSSEC. That doesn't seem right webhost: i've had openbsd work with virtio on kvm I modeled some of the tools after http://www.dnsstuff.com/ but openbsd takes code from netbsd not freebsd mnathani: +1 for ipv6 Thanks brycec dont think google.com is dnssec yet winvive.com is however damn, really need to read up on dnssec touche, google.com really isn't signed You just figure... IT'S GOOGLE ah indeed you are mnathani http://dnssec-debugger.verisignlabs.com/winvive.com I have a bunch of tools like that over at my dnssec signed blog: http://mnathani.com/blog/dnssec-links/ *links to tools rather What, using your own IP is too hard? Not knocking the idea really, just curious at the gap between "I can't be arsed to get a domain for $1.10/yr" and "I can't be arsed to type an IP address or edit my hosts file" [19:42:48] up_the_irons: Another thing that might be nice is a domain for customers. up_the_irons: The rsync backup space is a neat idea. Critical question time - why would I use that space over Amazon's S3 RRS which is $.09/GB/mo? Besides saving on VPS b/w costs. Aaaand before I'm over this dnssec kick - hey up_the_irons why isn't arpnetworks.com DNSSEC signed? :) here is another question, why are the arpnetworks customer resolvers not validating DNSSEC? Domain names are easier to remember than ip addresses. < CaZe> up_the_irons: Another thing that might be nice is a domain for customers. <-- you mean like the dns records for "your-name.cust.arpnetworks.com" that are created when you sign up? Ah, I didn't know about that. But it'd be nice if we could change the name. That's what name.com is for.