RE: Any serious production servers yet?
I couldn't have put it better myself. Vis-a-vie network performance, my goal for DragonFly is to have 'good' performance. But I think it is a complete waste of time to try to squeeze every last erg out of the network subsystem like FreeBSD has. We aren't trying to compete with Cisco, and nobody in their right mind would take a turnkey BSD or linux-based system over a Cisco (or other piece of high-end networking gear) to route multi-gigabits/sec of traffic. I still think we can get close to FreeBSD's rated performance, eventually, but I am not willing to create a mess of hacks and crazy configuration options to turn DragonFly into the ultimate ether switch when I can purchase one off the shelf for a few hundred bucks. I think the last time I tried to use a general purpose UNIX OS as an actual 'router' was in 1994. We used two BSDi boxes (and later FreeBSD boxes) to route the two T1's that BEST Internet had when we had just started up. It was a horror, frankly. Hardware bugs in the ethernet cards and even in the T1 card required a lot of hacking to work around, and trying to run BGP with gated was even worse. Back then 'real' networking hardware was bulky and expensive. Today, though, there is no excuse. It's cheap (and even cheaper on E-Bay), and far more reliable then a general purpose PC. If someone is trying to route multi-gigabits worth of traffic then the infrastructure is clearly important enough to warrent purchasing dedicated networking gear. If someone isn't trying to go all out, then a general purpose OS might be adequate, if still not as reliable. So all I can say to Mr Thom in that regard is: Stop trying to fit a square peg into a round hole and just buy the appropriate gear for your network infrastructure needs. -Matt
Re: Any serious production servers yet?
On 6/3/06, Danial Thom [EMAIL PROTECTED] wrote: --- Matthew Dillon [EMAIL PROTECTED] wrote: I couldn't have put it better myself. Vis-a-vie network performance, my goal for DragonFly is to have 'good' performance. But I think it is a complete waste of time to try to squeeze every last erg out of the network subsystem like FreeBSD has. We aren't trying to compete with Cisco, and nobody in their right mind would take a turnkey BSD or linux-based system over a Cisco (or other piece of high-end networking gear) to route multi-gigabits/sec of traffic. I still think we can get close to FreeBSD's rated performance, eventually, but I am not willing to create a mess of hacks and crazy configuration options to turn DragonFly into the ultimate ether switch when I can purchase one off the shelf for a few hundred bucks. I think the last time I tried to use a general purpose UNIX OS as an actual 'router' was in 1994. We used two BSDi boxes (and later FreeBSD boxes) to route the two T1's that BEST Internet had when we had just started up. It was a horror, frankly. Hardware bugs in the ethernet cards and even in the T1 card required a lot of hacking to work around, and trying to run BGP with gated was even worse. Back then 'real' networking hardware was bulky and expensive. Today, though, there is no excuse. It's cheap (and even cheaper on E-Bay), and far more reliable then a general purpose PC. If someone is trying to route multi-gigabits worth of traffic then the infrastructure is clearly important enough to warrent purchasing dedicated networking gear. If someone isn't trying to go all out, then a general purpose OS might be adequate, if still not as reliable. So all I can say to Mr Thom in that regard is: Stop trying to fit a square peg into a round hole and just buy the appropriate gear for your network infrastructure needs. -Matt Your caveman-like views are as troubling as they are entertaining. You seem to have no grasp of the modern world and no understanding of 'BSDs niche. Everything was buggy in '94, but with you and clowns like Paul Borman trying to do networking, what the hell would you expect no matter what you had to work with? :))) Many, many large network appliances (load balancers, bandwidth managers, firewalls, security filters) are based on linux or BSD. The reason is that CISCOs and mega-gigabit routers have no extra CPU power to do things like filtering and shaping at a very high level. I've made myself many millons of $$ selling a few thousand network devices, which is more than you'll ever make having a really cool desktop OS, even if its better than anything else out there. Designing a product for fun is one thing, but if you want to get funding you have to produce something that's useful for the corporate world, not for a bunch of pimply-faced college kids. The reality of the corporate world is that even if DFLY is the best damned OS ever written, they will use windows or linux, because you can't staff a support center with DFLY experts. Its simply never going to happen. You can however get in as a server platform, because only a couple of guys have to know what they're doing. Unix as a desktop box is not even an afterthought. 'BSDs niche is as a network server. Period. You might think its a waste of time to optimize networking, but it seems to me you're wasting your time entirely if your goal is to be a little faster than LINUX as a desktop box. Who cares? FreeBSD with 1 processor is faster than linux with 2, but no-one used FreeBSD anyway. Nobody wants to use 'BSD as a desktop machine, except for a handful of people with a lot more time on their hands than the rest of us. People want to use 'BSD as network servers. People in the real world that is. Maybe thats why your not with FreeBSD anymore; your refusal to modernize your ideas to what's going on in the real world, and your complete lack of understanding where the dollars are to fund your efforts? I should probably be moving on the same trend the other subscribers follow and give you a very diplomatic pat on the shoulder, but your bluntness simply calls for more. Shouldn't you be out, making some millions ? You seem to be better at it than at implanting your ideas into other people's minds. Everything they do, and especially Matt, is pro-bono. For fun. While their idea of having fun consists of spending a considerable amount of hours each day writing code, yours seem to be polishing your typing skills. Do all of us and especially yourself a favor and reconsider your schedule. DT Dumb Troll ? __ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com -- If it's there, and you can see it, it's real.
Re: Any serious production servers yet?
James Mansion wrote: [...] Actually I work in a rather large bank and I write trading systems... The most important thing I've learned from reading this thread is that DragonFly continues to attract attention from an amazing variety of bright people all around the world. Even though I don't understand a lot of the technical talk I read in these groups, I think I know enough about people to be able to spot competent ones when I see them. I see no reason to doubt a bright future for DragonFly. -- Matt: just keep on coding -- you are clearly making progress :o)
RE: Any serious production servers yet?
jemalloc scales great on SMP, see Jason's BSDCan paper. I'll search for it. I note that nedmalloc does well, there's a new ptmalloc, and the latest nedmalloc seems to be based on a new DLMalloc that I think's not public yet. And the google malloc seems to work pretty well, though I've not looked closely at it since its *Nix only and I favour things that will work on WIn32 too. Thanks for the pointer though. James
Re: Any serious production servers yet?
--- Vlad GALU [EMAIL PROTECTED] wrote: On 6/3/06, Danial Thom [EMAIL PROTECTED] wrote: --- Matthew Dillon [EMAIL PROTECTED] wrote: I couldn't have put it better myself. Vis-a-vie network performance, my goal for DragonFly is to have 'good' performance. But I think it is a complete waste of time to try to squeeze every last erg out of the network subsystem like FreeBSD has. We aren't trying to compete with Cisco, and nobody in their right mind would take a turnkey BSD or linux-based system over a Cisco (or other piece of high-end networking gear) to route multi-gigabits/sec of traffic. I still think we can get close to FreeBSD's rated performance, eventually, but I am not willing to create a mess of hacks and crazy configuration options to turn DragonFly into the ultimate ether switch when I can purchase one off the shelf for a few hundred bucks. I think the last time I tried to use a general purpose UNIX OS as an actual 'router' was in 1994. We used two BSDi boxes (and later FreeBSD boxes) to route the two T1's that BEST Internet had when we had just started up. It was a horror, frankly. Hardware bugs in the ethernet cards and even in the T1 card required a lot of hacking to work around, and trying to run BGP with gated was even worse. Back then 'real' networking hardware was bulky and expensive. Today, though, there is no excuse. It's cheap (and even cheaper on E-Bay), and far more reliable then a general purpose PC. If someone is trying to route multi-gigabits worth of traffic then the infrastructure is clearly important enough to warrent purchasing dedicated networking gear. If someone isn't trying to go all out, then a general purpose OS might be adequate, if still not as reliable. So all I can say to Mr Thom in that regard is: Stop trying to fit a square peg into a round hole and just buy the appropriate gear for your network infrastructure needs. -Matt Your caveman-like views are as troubling as they are entertaining. You seem to have no grasp of the modern world and no understanding of 'BSDs niche. Everything was buggy in '94, but with you and clowns like Paul Borman trying to do networking, what the hell would you expect no matter what you had to work with? :))) Many, many large network appliances (load balancers, bandwidth managers, firewalls, security filters) are based on linux or BSD. The reason is that CISCOs and mega-gigabit routers have no extra CPU power to do things like filtering and shaping at a very high level. I've made myself many millons of $$ selling a few thousand network devices, which is more than you'll ever make having a really cool desktop OS, even if its better than anything else out there. Designing a product for fun is one thing, but if you want to get funding you have to produce something that's useful for the corporate world, not for a bunch of pimply-faced college kids. The reality of the corporate world is that even if DFLY is the best damned OS ever written, they will use windows or linux, because you can't staff a support center with DFLY experts. Its simply never going to happen. You can however get in as a server platform, because only a couple of guys have to know what they're doing. Unix as a desktop box is not even an afterthought. 'BSDs niche is as a network server. Period. You might think its a waste of time to optimize networking, but it seems to me you're wasting your time entirely if your goal is to be a little faster than LINUX as a desktop box. Who cares? FreeBSD with 1 processor is faster than linux with 2, but no-one used FreeBSD anyway. Nobody wants to use 'BSD as a desktop machine, except for a handful of people with a lot more time on their hands than the rest of us. People want to use 'BSD as network servers. People in the real world that is. Maybe thats why your not with FreeBSD anymore; your refusal to modernize your ideas to what's going on in the real world, and your complete lack of understanding where the dollars are to fund your efforts? I should probably be moving on the same trend the other subscribers follow and give you a very diplomatic pat on the shoulder, but your bluntness simply calls for more. Shouldn't you be out, making some millions ? You seem to be better at it than at implanting your ideas into other people's minds. Everything they do, and especially Matt, is pro-bono. For fun. While their idea of having fun consists of spending a considerable amount of hours each day writing code, yours seem to be polishing
Re: Any serious production servers yet?
On 6/4/06, Danial Thom [EMAIL PROTECTED] wrote: --- Vlad GALU [EMAIL PROTECTED] wrote: On 6/3/06, Danial Thom [EMAIL PROTECTED] wrote: --- Matthew Dillon [EMAIL PROTECTED] wrote: I couldn't have put it better myself. Vis-a-vie network performance, my goal for DragonFly is to have 'good' performance. But I think it is a complete waste of time to try to squeeze every last erg out of the network subsystem like FreeBSD has. We aren't trying to compete with Cisco, and nobody in their right mind would take a turnkey BSD or linux-based system over a Cisco (or other piece of high-end networking gear) to route multi-gigabits/sec of traffic. I still think we can get close to FreeBSD's rated performance, eventually, but I am not willing to create a mess of hacks and crazy configuration options to turn DragonFly into the ultimate ether switch when I can purchase one off the shelf for a few hundred bucks. I think the last time I tried to use a general purpose UNIX OS as an actual 'router' was in 1994. We used two BSDi boxes (and later FreeBSD boxes) to route the two T1's that BEST Internet had when we had just started up. It was a horror, frankly. Hardware bugs in the ethernet cards and even in the T1 card required a lot of hacking to work around, and trying to run BGP with gated was even worse. Back then 'real' networking hardware was bulky and expensive. Today, though, there is no excuse. It's cheap (and even cheaper on E-Bay), and far more reliable then a general purpose PC. If someone is trying to route multi-gigabits worth of traffic then the infrastructure is clearly important enough to warrent purchasing dedicated networking gear. If someone isn't trying to go all out, then a general purpose OS might be adequate, if still not as reliable. So all I can say to Mr Thom in that regard is: Stop trying to fit a square peg into a round hole and just buy the appropriate gear for your network infrastructure needs. -Matt Your caveman-like views are as troubling as they are entertaining. You seem to have no grasp of the modern world and no understanding of 'BSDs niche. Everything was buggy in '94, but with you and clowns like Paul Borman trying to do networking, what the hell would you expect no matter what you had to work with? :))) Many, many large network appliances (load balancers, bandwidth managers, firewalls, security filters) are based on linux or BSD. The reason is that CISCOs and mega-gigabit routers have no extra CPU power to do things like filtering and shaping at a very high level. I've made myself many millons of $$ selling a few thousand network devices, which is more than you'll ever make having a really cool desktop OS, even if its better than anything else out there. Designing a product for fun is one thing, but if you want to get funding you have to produce something that's useful for the corporate world, not for a bunch of pimply-faced college kids. The reality of the corporate world is that even if DFLY is the best damned OS ever written, they will use windows or linux, because you can't staff a support center with DFLY experts. Its simply never going to happen. You can however get in as a server platform, because only a couple of guys have to know what they're doing. Unix as a desktop box is not even an afterthought. 'BSDs niche is as a network server. Period. You might think its a waste of time to optimize networking, but it seems to me you're wasting your time entirely if your goal is to be a little faster than LINUX as a desktop box. Who cares? FreeBSD with 1 processor is faster than linux with 2, but no-one used FreeBSD anyway. Nobody wants to use 'BSD as a desktop machine, except for a handful of people with a lot more time on their hands than the rest of us. People want to use 'BSD as network servers. People in the real world that is. Maybe thats why your not with FreeBSD anymore; your refusal to modernize your ideas to what's going on in the real world, and your complete lack of understanding where the dollars are to fund your efforts? I should probably be moving on the same trend the other subscribers follow and give you a very diplomatic pat on the shoulder, but your bluntness simply calls for more. Shouldn't you be out, making some millions ? You seem to be better at it than at implanting your ideas into other people's minds. Everything they do, and especially Matt, is pro-bono. For fun. While their idea of having fun consists of spending a considerable amount of hours each day writing code, yours seem to be
Re: Any serious production servers yet?
--- walt [EMAIL PROTECTED] wrote: James Mansion wrote: [...] Actually I work in a rather large bank and I write trading systems... The most important thing I've learned from reading this thread is that DragonFly continues to attract attention from an amazing variety of bright people all around the world. Even though I don't understand a lot of the technical talk I read in these groups, I think I know enough about people to be able to spot competent ones when I see them. I see no reason to doubt a bright future for DragonFly. The reason DFLY is getting attention is because there's nothing else out there and we've about had it with the FreeBSD camp. FreeBSD has the right idea, but not the talent to get the job done. Matt unfortunately doesn't understand or care what the market wants; he's more on a personal mission of some sort that has nothing to do with providing the marketplace with the next generation 'BSD performer that it wants. Someone eventually will wrestle either FreeBSD or DFLY away, or branch off and do what needs to be done. Unfortunately it may take longer than a lot of us were hoping for. On the other hand, we may never see a team like the original BSD team talent-wise. The Beatles have broken up, and things may never be the same again. DT __ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com
Re: Any serious production servers yet?
On Jun 3, 2006, at 1:49 PM, Danial Thom wrote: Many, many large network appliances (load balancers, bandwidth managers, firewalls, security filters) are based on linux or BSD. The reason is that CISCOs and mega-gigabit routers have no extra CPU power to do things like filtering and shaping at a very high level. I've made myself many millons of $$ selling a few thousand network devices, which is more than you'll ever make having a really cool desktop OS, even if its better than anything else out there. Designing a product for fun is one thing, but if you want to get funding you have to produce something that's useful for the corporate world, not for a bunch of pimply-faced college kids. The reality of the corporate world is that even if DFLY is the best damned OS ever written, they will use windows or linux, because you can't staff a support center with DFLY experts. Its simply never going to happen. You can however get in as a server platform, because only a couple of guys have to know what they're doing. When Cisco wants to move billions and billions of packets, they use IOS, the latest version of which is based on QNX... not Linux, Not BSD. Maybe you just missed the news. http://www.qnx.com/news/pr_1074_4.html Unix as a desktop box is not even an afterthought. 'BSDs niche is as a network server. Period. I have used a Mac, and I MUST say, it makes a descent workstation. Darwin ports is much better than that apt/fink thing if you ask me. I would also go so far as to say that Mac OSX doesn't make a very good server out of the box, unless you buy the server edition, and then things start to get distinctly confusing when you try to move out side of apple's box. 'BSDs niche is where ever the person leading the fork wants to take it. There are enough projects out there that people could, say, not work on one where they disagree with the leader. You might think its a waste of time to optimize networking, but it seems to me you're wasting your time entirely if your goal is to be a little faster than LINUX as a desktop box. Who cares? FreeBSD with 1 processor is faster than linux with 2, but no-one used FreeBSD anyway. Nobody wants to use 'BSD as a desktop machine, except for a handful of people with a lot more time on their hands than the rest of us. People want to use 'BSD as network servers. People in the real world that is. Maybe thats why your not with FreeBSD anymore; your refusal to modernize your ideas to what's going on in the real world, and your complete lack of understanding where the dollars are to fund your efforts? It is our belief that the correct choice of features and algorithms can yield the potential for excellent scalability, robustness, and debuggability in a number of broad system categories. Not just for SMP or NUMA, but for everything from a single-node UP system to a massively clustered system. We believe that a fairly simple but wide- ranging set of goals will lay the groundwork for future growth. (http://www.dragonflybsd.org/) no where does it say that the target for DragonFly BSD is aimed only at Workstations or desktops. No where does it say that DragonFly BSD is going to be the fastest OS on the planet. It doesn't even claim that DragonFly BSD will be the most feature complete OS on the planet. It says that the goal is laying the groundwork for future growth. Matt has specifically said that he doesn't want to try to wring out every last possible drop of performance at the cost of giving up this goal, which is very consistent with the projects stated goals. My real question is: Why are you, Danial Thom, interested in DFBSD at all? Jason Watson. smime.p7s Description: S/MIME cryptographic signature
Re: Any serious production servers yet?
Justin C Sherrill [EMAIL PROTECTED] writes: Dave Hayes wrote: So. What would it take to have a simple and concise set of commands any inexperienced adopter could easily apply to get a basic DFly system with X, gnome or KDE, and some basic applications? Someone writing it is what it takes. Like much open source documentation, the handbook and other things are written in DocBook, I have no problem using DocBook (it's my current documentation system as well) and I think I've even contributed one small patch to the handbook. But I'm not sure the handbook is the answer. The wiki presents a nice alternative; it's much easier to add to. Except it's recently been unstable. At some point I presume it will become stable again. :) If you want to write something up and put it in the wiki, that would be great. If you want to add something to the handbook or other documents in CVS under doc/, you can write the plain text and I'll happily do the conversion if you don't want to. I'm coming up on some actual work I have to do comparing DFly to FreeBSD, but I lead three lives at the moment only one of which is computer science. If no one else beats me to it, I guess the best place to start would be to bring up a DFly desktop, journal my experiences, and attempt to put something together which can hopefully be maintained. My one reservation is that I really want to know what the best practices are, and I fear that I won't know that until 6 months down the road where I encounter some problem that highlights that I haven't done the best practice. :) Thus, hopefully if I write something like this I can expect to have lots of peer and mentor review. I completely 100% agree with you. We need more docs that are easy and quick. Even if the Handbook was up to date and revised, it'd still be quite the slog to get through. There's a bit of conflicting information as well. Perhaps the real way to document the process of getting to a place where you can just say install this is to write a shell script? I've been using that methodology to document processes for quite some time. -- Dave Hayes - Consultant - Altadena CA, USA - [EMAIL PROTECTED] The opinions expressed above are entirely my own Nasrudin was throwing handfuls of crumbs around his house. What are you doing? someone asked him. Keeping the tigers away, replied Nasrudin. But there are no tigers in these parts. That's right. Effective, isn't it?
Re: Any serious production servers yet?
On Fri, June 2, 2006 2:09 am, Dave Hayes wrote: The wiki presents a nice alternative; it's much easier to add to. Except it's recently been unstable. At some point I presume it will become stable again. :) I think it's stable now; the issue wasn't stability as much as missing. If no one else beats me to it, I guess the best place to start would be to bring up a DFly desktop, journal my experiences, and attempt to put something together which can hopefully be maintained. I've been doing something similar with the DragonFly BSD Log ( http://www.shiningsilence.com/dbsdlog/ ). Perhaps we should have something similar to freebsddiary.org? Perhaps the real way to document the process of getting to a place where you can just say install this is to write a shell script? I've been using that methodology to document processes for quite some time. You mean a shell script that automates certain common processes? That would be handy. I have been thinking that instead of larger documents like the Handbook, we could work on creating a series of how-to documents, like other systems. This may be more maintainable, and translates more directly to Wiki material, which is so far the easiest way for folks to contribute.
RE: Any serious production servers yet?
--- James Mansion [EMAIL PROTECTED] wrote: A dual-core 2.6 Opteron is about US$1079. whereas a single core is about $460. So for about $200. more I can build 2 2.6Ghz systems that give me a lot more bang for my buck than 1 dual-core system. Well, the bleeding edge is always at a premium. But you mention a wall. A wall doing what? Single threaded monte-carlo? Single postgres query? Almost any real-world load that will stress a modern server box comes from multiple requests. As for pf performance - who the hell cares? Are you routing between two 10GBit LANs? Frankly I don't care about pf performance, but your comments indicate your ignorance as to how it performs under load. pf will barf on many real networks pushing a lot less than you think, depending on the complexity of the ruleset. If you're using that much CPU, then if you care what OS you're using, your app is badly written, cos you should first avoid entering the kernel anyway as much as you can. Ah, from the mouths of babes (or people who live in tiny caves)! If your network is pushing 300-500K pps its nice to have a firewall or security device or router that can handle it. And those filtering/network functions don't benefit much from MP. The wall is what such a box can handle with the fastest processor available. Typically MP doesn't scale well for such APPs, as the overhead associated with threading the kernel slows the raw performance more than is gained by having multiple processors. Getting past the wall would be good, but if the trade off is being able to do some other stuff with the box simultaneously, the capacity can't diminish too much, because once your traffic levels are past the wall you are in trouble. That is the Wall for people who are on real networks. You could always count on the wall advancing as our buddies at AMD and Intel increased the GHZ. Now we're going sideways, and many can't afford to have the wall regress to accommodate smoother audio performance. DT __ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com
RE: Any serious production servers yet?
Danial Thom wrote: That is the Wall for people who are on real networks. You could always count on the wall advancing as our buddies at AMD and Intel increased the GHZ. Now we're going sideways, and many can't afford to have the wall regress to accommodate smoother audio performance. Danial this is the wall for *you* and only for you. Do you want to know the truth for me? I could not care less if you can route gigabit links with BSD and filter them with pf. If i was in the situation to do that i would buy dedicated hardware. But i care very much that the software i am using runs smoothly, and for that several processors give a very considerable bonus. With *my* present needs FreeBSD, DragonFly and Linux give a very good experience. There are certainly far more machines running tomcat or jboss servers with a lot of threads which greatly and immediately benefit from dual cores or more, than commodity machines used to do the job of dedicated hardware. And yes, as Kris said, jemalloc works well at present on FreeBSD. -- Michel Talon
Re: Any serious production servers yet?
On 2006-06-01 18:46, Danial Thom wrote: --- Sascha Wildner [EMAIL PROTECTED] wrote: Danial Thom wrote: Surely it makes sense to begin developing O/S applications (which is what I need to do), however I need an OS that is production ready, even if its not as good as its going to be, because I can't reasonably test the performance of an application on an OS that can't handle production loads. *sigh* Is this going to be another of those half-yearly Danial vs. the rest threads? How about this: You restrain yourself from stealing people's time with your annoying discussion for discussion's sake and I promise to get back to you in personal email as soon as I think that DragonFly has reached the point where it could be interesting to you? Sascha I don't see that its me vs anything. I have to chose an MP OS for a big project and I just asked if the project is production-ready yet, and instead of getting an answer, I get a lot of pointers to personal web pages and routers that aren't even pushing a T1. A simple answer like No, DFLY isn't ready for prime time yet, and we don't expect it will be until Sept '07. would have avoided wasting your time. Well, the definition of production-ready differs with the needs of the production server. You've got lots of examples of DF being used in production but none of them happens to be the kind of environment that you'll use. Regardless of the type of services one would use an OS for there's only one reliable way to determine if a particular solution is the one you want or not, and that is testing it yourself under the same conditions that it will encounter in production. In short the only way to know is to test it yourself. To ask if someone has done X using Y and Z under conditions A, B and C is often pointless since no setup or requirement is identical to another unless it's very simple. Erik Wikström -- I have always wished for my computer to be as easy to use as my telephone; my wish has come true because I can no longer figure out how to use my telephone -- Bjarne Stroustrup
Re: Any serious production servers yet?
Danial Thom wrote: ...I just asked if the project is production-ready yet, and instead of getting an answer... Well, I'm a hobbyist who doesn't even own a MP machine, so of course I'll be happy to answer ;o) Matt is right in the middle of a major revision of the SMP parts of the kernel even as we speak [see other recent threads in this group]. My prediction: whatever the performance of recent DragonFly releases may be -- it's about to change again. Most likely for the better. But I could be wrong.
Re: Any serious production servers yet?
Danial Thom [EMAIL PROTECTED] writes: You're thinking like an engineer, and not a marketeer. Yes. This is an excellent reason to use DragonFly. :) -- Dave Hayes - Consultant - Altadena CA, USA - [EMAIL PROTECTED] The opinions expressed above are entirely my own Wisdom (n.) - 1. Something you can learn without knowing it.
RE: Any serious production servers yet?
I guess I should have qualified my question. If you're pushing less than 100Kb/s then there's really no reason to spend 3X the dollars on a multi-core system. So the only real value of an As of NOW, the price differential between a single core 2.6ghz Opteron and a dual-core one is about 120%. I can't think of many applications that are going to push a 2.6Ghz opteron that justify spending more than twice as much. Of While I'd agree that in general CPUs today are really pretty fast, I think this '3X' and '120%' pitch suggests borked thinking, at least for the case of whether to buy a dual core socket 939 or 940 chip - because while the cost differential is quite steep, its only the CPU and in effect you get a lot more bang for an incremental change in system bucks - you don't even need a pricey mobo. If you're saving pennies, get a cheap Intel D920 system. Can't argue with the amount of grunt you get for 200 bucks (well, 140 quid including VAT). And the D805 is only 85 quid! Might be last year's FSB speed etc and hot compared to AMD, but if you compare the bang you get from a cheap mobo and one of these things with the breakthrough price performance we got from dual PPro systems in their day, its just laughable. Of course, it would be nice if one of the BSDs could actually have a working pthread system that scaled well and had a decent malloc too. How about it guys? ;-) How many do FreeBSD have now? 3? And do any of them work properly? James
RE: Any serious production servers yet?
--- James Mansion [EMAIL PROTECTED] wrote: I guess I should have qualified my question. If you're pushing less than 100Kb/s then there's really no reason to spend 3X the dollars on a multi-core system. So the only real value of an As of NOW, the price differential between a single core 2.6ghz Opteron and a dual-core one is about 120%. I can't think of many applications that are going to push a 2.6Ghz opteron that justify spending more than twice as much. Of While I'd agree that in general CPUs today are really pretty fast, I think this '3X' and '120%' pitch suggests borked thinking, at least for the case of whether to buy a dual core socket 939 or 940 chip - because while the cost differential is quite steep, its only the CPU and in effect you get a lot more bang for an incremental change in system bucks - you don't even need a pricey mobo. A dual-core 2.6 Opteron is about US$1079. whereas a single core is about $460. So for about $200. more I can build 2 2.6Ghz systems that give me a lot more bang for my buck than 1 dual-core system. Intel isn't quite in play yet, since a dual-core Pentium D doesn't give me the performance of a single 2.6Ghz Opteron, so there's no point in even considering it. Woodcrest/Conroe will change things, of course. Again, I'm talking about getting past the wall, so the lower end stuff doesn't buy me anything. DT __ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com
RE: Any serious production servers yet?
A dual-core 2.6 Opteron is about US$1079. whereas a single core is about $460. So for about $200. more I can build 2 2.6Ghz systems that give me a lot more bang for my buck than 1 dual-core system. Well, the bleeding edge is always at a premium. But you mention a wall. A wall doing what? Single threaded monte-carlo? Single postgres query? Almost any real-world load that will stress a modern server box comes from multiple requests. As for pf performance - who the hell cares? Are you routing between two 10GBit LANs? If you're using that much CPU, then if you care what OS you're using, your app is badly written, cos you should first avoid entering the kernel anyway as much as you can. Personally I value a good pthread/libc much higher than parallelism in the kernel for exactly this reason.
Re: Any serious production servers yet?
--- Matthew Dillon [EMAIL PROTECTED] wrote: Er. Well, if I were talking about today I would be talking about today. I'm talking about the near-future, 2-3 years from now. It would be the height of stupidity to have programming goals that only satisfy the needs of today. It might be the height of stupidity, but it takes 2-3 years to convince people that you have something worth using, even if you have something great, so are you prepared to wait 4-5 to have a mainstream O/S? Once you get the groundwork done, you should ramp up to be production quality so you can get some noteworthy people using the O/S in real-world servers. Then you can keep it stable and work in your roadmap. We're approaching a critical point where a lot of companies are going to start looking to move into MP who haven't been there before. Saying you'll have something really great in 2 years isn't going to get people involved with the project. Your project will move ahead at an exponential pace once you get more funding and more important people working on it. You can't do that with an OS that can't be used by anyone with the money to contribute. You're thinking like an engineer, and not a marketeer. Its sort of like a kid going to college part-time trying to pay as he goes, earning $10/hr, while he could borrow the money and pay it off by making $50. an hour by graduating earlier. Its a backwards approach to product development. In 2-3 years single-core cpus will be relegated to niche status. You won't be able *BUY* Intel or AMD single-core's at all for general purpose computers. It won't matter a bit whether the average consumer is able use the extra computing power, it will be there anyway because it doesn't cost Intel or AMD any more to build it verses building single-core cpu's, it doesn't eat any more power either (in fact, it eats less, for more aggregate computing power). So regardless of what you believe the future is quite clearly going to become permanently multi-core. In anycase, it's a mistake to assume that the extra computing power is wasted just because you can't think of anything that can use it right now. That mentality is what caused Bill Gates to make the statement that no computer would ever need more then 640KB of memory. Nobody is saying that it can't be used, I'm just saying its not worth the $$$ today because the marginal cost of the hardware can't be utilized by the O/S and the applications that most use. So why buy MP today? Surely it makes sense to begin developing O/S applications (which is what I need to do), however I need an OS that is production ready, even if its not as good as its going to be, because I can't reasonably test the performance of an application on an OS that can't handle production loads. There are plenty of applications both existing and on the horizon that would be easily be able to use the additional computing power. Even on a fast machine today SSH can still only encrypt at a 25-40MB/sec rate. Filesystems such as ZFS are far more computationally expensive then what we use today, but what you get for that price is an unbelievable level of stability and redundancy. Photo-processing? It takes my fastest box 4 hours to run through the fixups for one trip's worth of photos. Since that workload is primarily userland, it only takes 2 hours on my dual-core box. Encryption, Graphics, Photo-processing, Database operations. Well I'm talking about servers here, because thats where the big $$$ are. With photoshop you're talking about saving some seconds here and there. Maybe productivity but most people do multiple things at once so its not really clear that being able to do something in 2 seconds instead of 5 matters. We're not talking floppy disks vs scsi raid here. The big picture issue is capacity. I don't really care if my server has 2ms or 4ms latency, I want it to be able to handle the load without barfing. At gigabit speeds filtering devices (like firewalls, soft routers, bandwidth management boxes) are pushing the envelope now. They can't relinquish capacity in order to have a smoother feel to the user interface. The capacity has to at LEAST be the same, or very close. DT __ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com
Re: Any serious production servers yet?
--- Kevin L. Kane [EMAIL PROTECTED] wrote: So, 2-3 years tops, and there won't be any more single-core offerings from AMD or Intel. Probably not even for laptops. This is really already happening, ALL of Apple's new latops are dual core only and the only single core Intel based mac is the cheapest Mini. Apple is not really a great example, as they control the entire animal, and there's more margin in the high end. Here's a question for Matt, will dual-core designed chips (as opposed to chips with 2 independent cores on once chip) be used on an UP OS as a single core? Say if I wanted to use a dual-core chip on Freebsd 4.x in UP mode since SMP sucks wind? Or do the cores designed as dual-core with the shared caches require that both be used? DT __ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com
Re: Any serious production servers yet?
Danial Thom wrote: Surely it makes sense to begin developing O/S applications (which is what I need to do), however I need an OS that is production ready, even if its not as good as its going to be, because I can't reasonably test the performance of an application on an OS that can't handle production loads. *sigh* Is this going to be another of those half-yearly Danial vs. the rest threads? How about this: You restrain yourself from stealing people's time with your annoying discussion for discussion's sake and I promise to get back to you in personal email as soon as I think that DragonFly has reached the point where it could be interesting to you? Sascha P.S. All others: Just dig the archives for past threads of DT. -- http://yoyodyne.ath.cx
Re: Any serious production servers yet?
On Wed, May 31, 2006 11:50 am, Danial Thom wrote: --- Matthew Dillon [EMAIL PROTECTED] wrote: Er. Well, if I were talking about today I would be talking about today. I'm talking about the near-future, 2-3 years from now. It would be the height of stupidity to have programming goals that only satisfy the needs of today. It might be the height of stupidity, but it takes 2-3 years to convince people that you have something worth using, even if you have something great, so are you prepared to wait 4-5 to have a mainstream O/S? Once you get the groundwork done, you should ramp up to be production quality so you can get some noteworthy people using the O/S in real-world servers. Then you can keep it stable and work in your roadmap. We're approaching a critical point where a lot of companies are going to start looking to move into MP who haven't been there before. Saying you'll have something really great in 2 years isn't going to get people involved with the project. Your project will move ahead at an exponential pace once you get more funding and more important people working on it. You can't do that with an OS that can't be used by anyone with the money to contribute. You're thinking like an engineer, and not a marketeer. Its sort of ^^^ Thank god for that. It's about time!! There are two many OSes out there that are managed by the marketing dept. Marketers should not get involved with an OS at the get go, or even in the first few years. We need more engineers working on OSes, not more marketing droids and script kiddies testing out their compiler for the first time. :D
Re: Any serious production servers yet?
:Here's a question for Matt, will dual-core :designed chips (as opposed to chips with 2 :independent cores on once chip) be used on an UP :OS as a single core? Say if I wanted to use a :dual-core chip on Freebsd 4.x in UP mode since :SMP sucks wind? Or do the cores designed as :dual-core with the shared caches require that :both be used? : :DT Sure. You can always choose to use just one of the cpu's. Or, alternatively, if we were to implement the userland scheduler cpu masking feature that was discussed recently, one could run MP but partition the use of the cpus. -Matt Matthew Dillon [EMAIL PROTECTED]
Re: Any serious production servers yet?
danial_thom wrote @ Mon, 29 May 2006 16:59:06 -0700 (PDT): Is anyone using DragonflyBSD in any serious production servers yet? Any feelings about how it measures up in its current state performance-wise? One of the ftp.fortunaty.net mirrors ran DragonFly for 2 years or so. Latest uptime was 360 days before the server was phased out. Testing showed it to be really good performace wise. Matt's and especially Hsu's improvements of the network side showed. My current setup is based on Xen and unfortunatly there is currently no DragonFly-xen. What hurt back then was the missing support for nullfs in a realease and the ABI breakage(s). Package management was als a point. I had pretty much worked out everything i could think of with ports. Of course there was an investment made into the setup prior to it being good. Pkgsrc seemed a step back for me and i had had to invest again into my setup, so i choose not to. Currently I'm working on packages made with a customized version of pacman from archlinux (http://archlinux.org/). It seems much easier to handle than pkgsrc. See http://wiki.dragonflybsd.org/cgi/moin.cgi/Pacman_Packages for a bit more about this approach. -- Andy
Re: Any serious production servers yet?
On Tue, May 30, 2006 at 02:54:56PM +0300, Yiorgos Adamopoulos wrote: http://www.dbnet.ece.ntua.gr/~adamo/howto/DragonFlyBSD/ftp-proxy.txt Try redirecting to an address outside of 127/8 instead. Joerg
Re: Any serious production servers yet?
On Tue, May 30, 2006 11:20 am, Danial Thom wrote: I guess I should have qualified my question. If you're pushing less than 100Kb/s then there's really no reason to spend 3X the dollars on a multi-core system. So the only real value of an MP system is how it performs under heavy load, if you're talking about a server and not a desktop box. Is there a certain specific setup and threshold you are looking for? Like, performance of DragonFly with an MP kernel on a multi-core processor on a system acting as a firewall with traffic at level X? (Guessing from your writing above)
Re: Any serious production servers yet?
--- Justin C. Sherrill [EMAIL PROTECTED] wrote: On Tue, May 30, 2006 11:20 am, Danial Thom wrote: I guess I should have qualified my question. If you're pushing less than 100Kb/s then there's really no reason to spend 3X the dollars on a multi-core system. So the only real value of an MP system is how it performs under heavy load, if you're talking about a server and not a desktop box. Is there a certain specific setup and threshold you are looking for? Like, performance of DragonFly with an MP kernel on a multi-core processor on a system acting as a firewall with traffic at level X? (Guessing from your writing above) At this point I'm looking for someone using it on *any* server running at levels that might justify using MP. The entire point of MP, on a server anyway, is to get better performance than you can on a UP server. So I guess the criteria is whether anyone is running a server that pushes the envelope of what a UP server can do, even if its just once and a while. Stability at 10% utilization and stability at 50-60% utilization are very different animals. Of course if you can tell me that its not really ready for a serious production environment then that's enough information right there, which was the general sentiment 6 months ago. DT __ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com
Re: Any serious production servers yet?
Well, keep in mind, that in 2-3 years time there won't *BE* any single-cpu computers any more, at least not for consumer offerings. both Intel's and AMD's entire manufacturing line is going to be dual-core at a minimum, and higher-end products will be at least quad-core. Even as we speak Intel has begun repricing its dual-core cpus to be on par with their single-core cpus. AMD has already dropped their dual-core prices considerably, too. The writing is on the wall. -Matt
Re: Any serious production servers yet?
Danial Thom wrote: cut As of NOW, the price differential between a single core 2.6ghz Opteron and a dual-core one is about 120%. I can't think of many applications that are going to push a 2.6Ghz opteron that justify spending more than twice as much. Of course that's all going to change in a few months when AMD loses their advantage. I don't see the point of using a dual-core 2.0 when I can get equal performance from a single 2.6 for less money. So the price drops so far are meaningless, given the OS's ability to utilize the cpus. DT I think you hit the nail with that one, the only reason I going to buy an AMD x2 is because xp feels more responsive, for all the other things, since the 1ghz barrier was broken anything (for my usage) was fast enough, of course this is desktop usage, on my servers I recently have used VMWare to centralize a lot of servers (9) to a quad cpu machine because most of them where eating out of there nose anyway. Saves me a fortune in electricity ;-) I am eager to see how ever how xen would develop out with using the new virtualization technique in both the newer AMD and Intel CPU's. I for one like to ditch VMWare. -- mph
Re: Any serious production servers yet?
On 2006-05-30, Danial Thom [EMAIL PROTECTED] wrote: And what kind of volume are you pushing through your firewalls peak, in terms of bandwidth and pps? The main reason that we chose DragonFlyBSD was that it could *install* on a Siemens RX200S2 when all the other BSDs failed (and we prefer pf for packet filtering). 1.2.0 at the begining and _Preview now. Our bottleneck in terms of pps was our cisco card (which we now changed and we do not see any problems). The last thrity minutes it served 23,112 connections, 3,369,221 packets and 2,578,423,409 bytes at a load arround 0.02. The only thing that goes slow (compared to inferior and IDE DragonFlyBSD systems that we also run elswhere) is make buildworld and I am guessing that it has to do with the disks. However it is more of a nuissance than a problem and have not investigated further. I guess I should have qualified my question. If you're pushing less than 100Kb/s then there's really no reason to spend 3X the dollars on a multi-core system. So the only real value of an MP system is how it performs under heavy load, if you're talking about a server and not a desktop box. Well for now we use uniprocessor systems, so my mistake - it did not cross my mind that you wanted oppinions on MP systems only.
Re: Any serious production servers yet?
So, 2-3 years tops, and there won't be any more single-core offerings from AMD or Intel. Probably not even for laptops. This is really already happening, ALL of Apple's new latops are dual core only and the only single core Intel based mac is the cheapest Mini. On 5/30/06, Matthew Dillon [EMAIL PROTECTED] wrote: :Matthew Dillon wrote: : Well, keep in mind, that in 2-3 years time there won't *BE* any : single-cpu computers any more... : :I'm still confused about the difference (if any) between a motherboard :with two separate CPU's and a new mobo with a dual-core CPU. : :From your point of view as a kernel programmer, is there any difference? Basically, single-processor motherboards of yesterday still look the same as today, but that physical cpu chip you plug into the board can actually contain two cpu's instead of one if you are using a multi-core cpu (and a multi-core capable MB, of course). So you get the same computing power out of it as a dual-socket motherboard would have given you. It gets even better for SMP boards. The two-socket opteron MB of yesterday could accomodate two single-core cpu's giving you 2xCPU worth of computing power. A two-socket opteron MB of today can accomodate two multi-core cpu's giving you 4 cpu's and about 3x the performance. A 4-socket box gives you 6x the performance. It IS true that the initial dual-core parts run at a slightly slower clock rate. But this isn't because they are dual-core, it is simply because both AMD and Intel know that they had gone over the heat dissipation limits in their attempting to max-out the clock frequency of their single-core CPUs and going to dual-core gave them just the excuse they needed to back down to more reasonable dissipative levels. The result is that cooling requirements for dual-core cpu's, not to mention case and power supply requirements, are far lower. Lower requirements == costs less money to maintain, and in a server room == costs less money in electricity and costs less money in cooling. People don't care about single-core performance any more these days, because most computing jobs aren't single threaded. Even a mail server isn't single threaded... it forks off a process for each connection. As these cpu's approach price parity, dual-core makes more and more sense *EVEN* if you don't actually need the extra computing power, simply because the clock frequency reduction results in far, far less power use. What matters now is computing performance per watt of electricity used. That isn't to say that people are dropping AMD and Intel and going to ARM... there are still minimum performance requirements simply due to the cost of all the other hardware that makes up a computer. But it does mean that people would rather have two 2.0 GHz cpus which together use LESS power then a single 2.6 GHz cpu. So from my point of view, we win both ways. Not only are there plenty of applications that can use the newly available computing power, even if it is just in 'burst' usage, but you get that new power at lower cost. Even without price parity on the cpus (and as I have said, price parity is rapidly approaching anyhow), it is STILL worth it simply due to savings in electricity costs. I spend over $2500/year JUST to power my machine room. That's down from the $3500/year I was spending two years ago when I had 1/4 of the computing power shoved into that room as I have today. This means that I really don't give a damn whether a dual-core cpu costs a bit more money or not. The extra few hundred I spend on it today is easily made up in the power savings I get over the course of a single year. The reason? More aggregate computing power eating less electricity. If you think about it, this equation... 'more aggregate computing power eating less electricity' is PRECISELY what multi-core gives us. As Martin mentioned, using virtualization to concentrate computing power results in a huge savings. Multi-core allows you to concentrate the computing power, eating less electricity in the process, even more. -Matt Matthew Dillon [EMAIL PROTECTED] -- Kevin L. Kane kevin.kane at gmail.com
Re: Any serious production servers yet?
If some of the devs could do porting ftp-proxy (formerly pftpx) and ftpsesame I would switch immediately. Consider this as an argument. :-P Both are available via pkgsrc. ftp-proxy is in the pkgsrc/security/pflkm package. I didn't check if it was the newer pftpx rewrite though. And ftpsesame is at pkgsrc/wip/ftpsesame. I haven't test either of these on DragonFly. Maybe the pkgsrc/security/pflkm can split out the ftp-proxy part to its own package. Jeremy C. Reed echo '9,J8HD,[EMAIL PROTECTED]:[EMAIL PROTECTED];[EMAIL PROTECTED]@5GBIELD54DL@8L?:5GDEJ8LDG1' |\ sed ss,s50EBsg | tr 0-M 'p.wBt SgiIlxmLhan:o,erDsduv/cyP'
Re: Any serious production servers yet?
On 5/31/06, Jeremy C. Reed [EMAIL PROTECTED] wrote: If some of the devs could do porting ftp-proxy (formerly pftpx) and ftpsesame I would switch immediately. Consider this as an argument. :-P Both are available via pkgsrc. ftp-proxy is in the pkgsrc/security/pflkm package. I didn't check if it was the newer pftpx rewrite though. And ftpsesame is at pkgsrc/wip/ftpsesame. Hold on... I've been using ftp-proxy on DragonFly systems since it first got pf. It's always been in the base package, and its binary has been in /usr/libexec/ftp-proxy. The only difference I can find from the NetBSD version is that it has no 'pf/ipfilter' distinction modes. I don't know how the OpenBSD one works. At the very least, I have never needed to resort to pkgsrc to do FTP proxying with pf in DragonFly. Also, the minor difficulty in not being able to target 127.x.y.z is not as bad as it sounds, as long as you can actually bind the server to something else. The filter should take care of the security aspect. And yes, the whole setup is remarkably stable under DragonFly in UP and SMP rigs. I've had one repeated panic with pf in an SMP environment but it was declawed by turning off normalization in the rule set. Once the rest of the network stack is properly MPSAFE it'll be difficult to recommend anything *but* DragonFly for a high-load firewall or server. -- Dmitri Nikulin
Re: Any serious production servers yet?
Matthew Dillon wrote: ...but that physical cpu chip you plug into the board can actually contain two cpu's instead of one if you are using a multi-core cpu (and a multi-core capable MB, of course)... To reconfirm: if I bought a mobo today (at Fry's, for example, just because I'm tempted ;o) with a dual-core processor -- I would need to configure my DragonFly kernel with 'options SMP' [and friends] and DragonFly would recognize and use both cores just like an old dual- processor mobo? Thanks!
Re: Any serious production servers yet?
:To reconfirm: if I bought a mobo today (at Fry's, for example, just :because I'm tempted ;o) with a dual-core processor -- I would need to :configure my DragonFly kernel with 'options SMP' [and friends] and :DragonFly would recognize and use both cores just like an old dual- :processor mobo? : :Thanks! SMP and APIC_IO usually. If APIC_IO doesn't work you can try without it. But, as usual, the kernel might have problems recognizing new motherboards. Our ACPI implementation is pretty old. Your mileage may vary. -Matt Matthew Dillon [EMAIL PROTECTED]