that they had the Apache config set to allow up to 127 servers, and they had not raised file_max from 1024 to something decent. Do the math. Each Apache server has 8 file handles open just sitting there doing nothing. If it then goes to open a file to serve it, and file_max has been exceeded, guess what happens? Yep, Apache collapses!
In general there are some areas where Linux lags NT. IIS, for example, outperforms Linux on static page displays because it has a page cache and does not have to always cross the user-kernel boundary to fetch the page (system calls DO have a cost, even though 2.2 sped up the open() call considerably with the dcache). And it may very well be that a well-tuned ultra-high-end NT machine will beat a well-tuned Linux machine at file serving, given that NT will support the full 4gb of memory while Linux only supports 2gb of memory. But this "test" was not such a test -- it compared a well-tuned NT machine against a totally untuned Linux machine.
And of course I'll point out that on tests on more modest hardware, like this, Linux blows away NT handily. To be fair, that Smart Reseller test was just as biased in its own way as this joke test we're talking about... Smart Reseller chose a machine that's too small for NT to comfortably stretch its legs, albeit that the machine they chose is rather typical of small office web servers.
Thanks. I just checked that out. It does appear that they asked a single question about Apache performance. I remember seeing that posting myself and blowing it off because there wasn't enough info to tell him anything and I didn't feel like going into the give-and-take to get enough info to do something. (I do enough of that supporting my own customers!). Now, in hindsight, knowing what he did not do to Linux, the answer is obvious: he was running out of file handles. Do the math. An inactive Apache server has 8 file handles open. 127 servers max * 8 = 1016. Default file_max is 1024 for Linux, of which 150 or so are usually open while the system is at a rest. Apache could not bind a socket to a file handle for incoming connections because there were no file handles. So Apache was basically deadlocked, waiting for file handles to come free so it could accept() the socket but it was already holding all the file handles!
If there had been questions about general tuning of such a large system, that would have solved the problem because someone would have remembered about file_max. But one cryptic query that didn't give enough information to get help does not an honest effort make.
Anyhow: I guess I have to post a partial retraction. They did post a *SINGLE* query to the net.
And: These people are *LIARS*. They say the posted messages asking for help on the Linux newsgroups. There are *NO* messages from the mindcraft domain anywhere on the Linux newsgroups. So I did DejaNews searches of "performance tuning", "performance tune", "kernel tuning", "kernel 2.2 tuning", between January 1 1999 and today, and examined the results to see if there were any messages that may have been by MindCraft researchers (i.e., that were referring to performance problems with a large-memory machine). There were *NONE*. Zero. Zilch. Which means that if they did ask any performance tuning questions, they did not use those words in the message.
Anyhow: VA Research already loaned a quad-processor Xeon machine to PC-Week and it blew away NT 4.0 in their SAMBA benchmarks. VA Research's quad-processor Xeon machine is the same machine that we sell, and the same machine that Penguin Computing sells (we all get them from Intel, and then dress them slightly differently once we get them, e.g. VA Research uses a Mylex RAID card while we and Penguin use ICP-Vortex RAID cards). So we already have the benchmark that shows that their SAMBA benchmark is full of ****. But that's not going to matter to pointy-haired bosses because they recognize only those reports and studies that say what they want to hear.
1) Dell is almost totally a "white box" operation, with the exception of the very high end of their line. In fact, the Compaq guys make fun of Dell because so much of Dell's stuff is just off-the-shell commodity hardware. The Dell guys just smile all the way to the bank.
2) I assure you that LHS and VA Research, at least, have people on staff who came from Compaq, Apple, etc. and are quite familiar with the design centers at those compoanies.
3) VA Research no longer builds their own low-end computers. Those have been outsourced to Flextronics (or did you miss their news release?). Most computer companies do this. Very few low-end computers are built by the company whose badge is on the front cover. (Heck, even the iWhack is contracted out to such an outsourcing firm!).
4) In general, almost all facets of designing and building a computer can be outsourced now. I can't tell you details (sorry, NDA), but basically it becomes a game of where your design dollars should go. In the $1K-$30K range Compaq cannot build better motherboards than ASUS or Intel, and in fact they don't. What they do accomplish is creating motherboards that are cheaper to add into manufactured computers. Ever wonder why so many components are integrated into Compaq motherboards? Well, that saves a few dollars because those components can be automagically placed onto the motherboard by robot pick'n'place machines, rather than having some poor slob have to spend thirty seconds slamming a video card into a computer and spinning the retaining screw home. But the thing about outsourcing is that the poor slob is now in Malaysia or Thailand and is being paid $3 per day. Given that, why invest design dollars creating a board that reduces manufacturing costs, but which is actually technically inferior to what can be bought off the shelf from ASUS or Intel? (I say technically inferior because integrated peripherals reduce the flexibility and repairability of a computer, both of which are qualities which I personally value).
5) All that nonwithstanding, you are correct about being able to put together a system equal to that of VA Research, Dell, or Linux Hardware Solutions out of off-the-shelf parts. In fact, any Joe Consultant in Cheyenne Wyoming can do the same. That is what is so amazing about today as vs. 20 years ago, when that was definitely not the case. Given that, vendors like VAR, LHS, etc. are not going to win by trying to out-engineer ASUS and Intel. Engineering resources have to be carefully allocated to those areas where off-the-shelf hardware currently doesn't exist (I don't think I can say more there, NDA etc). In the mainstream $1K-$30K server market, though, where any Joe Schmuck can buy the exact same parts off the shelf, it's the services of systems integration and support that we sell. We sort through the dozens of video cards and network cards and etc. out there so you don't have to, and then set things up and often times hunt up or write drivers for things that need it (like for the Symbios 53c896-based stuff or the Mylex stuff). If you don't need those services, build your own for crying out loud! That's why we put the parts lists up on our web sites, after all. But don't diss those of us providing a valuable service just because you don't need that particular service!
Cache coherency etc. is not part of the problem. The problem is the granularity of the semaphoring within the Linux kernel. That is, how much of the kernel can be running in parallel. Right now, once you get up past 4 processors, the kernel becomes the limit.
There are also some process control and migration issues, as well as memory management, all of which can be lumped under the banner of "scheduling issues". Matt Dillon of the FreeBSD project did an interesting analysis of the Linux VM, for example, noting how it was a simple and elegant design (and making suggestions for doing similar simplifications to the FreeBSD VM subsystem) but somewhat simple-minded with some obvious performance problems when dealing with large memory sets (like when we're talking 8-way processors!).
Anyhow: the Linux kernel guys know what needs doing. It's just that they're now at the hard part -- going from protecting entire subsystems with semaphors (so that they run on only a single processor) to allowing those subsystems to run in parallel on multiple processors (by protecting the data structures within those subsystems with semaphors). The known scheduling and VM issues need addressing too, of course. All of which is on the table for 2.3.
I've been bouncing around this business since 1982, basically, when as a kid I got my first little Commodore computer and got a subscription to Byte. How things have changed since then. Back then, you bought an Apple II if you were a mere mortal needing a computer, and building one was not in the cards unless you wanted a boat anchor filled with S-100 cards for an ungodly sum of money (anybody remember Godbout? Cromemco? All those other S-100 hobby vendors that are long gone?).
Today, thanks to the mass market in commodity components created by the IBM PC platform (see The Commoditization of Computers), any Joe Consultant in Hoboken, Michigan can put together his own computers that are every bit as high in quality and low in price as those from Dell. It's an amazing democratization of the computer industry, totally unlike anything that has ever happened in any other industry. Suddenly any schmuck off the street can build a computer just as good as what he can buy, often for less!
Given all that, folks like VA Research, Penguin, or Linux Hardware Solutions would have to be nuts to design their own motherboards. People don't buy our hardware because it is somehow better than what Joe Schmuck can put together in his back room with an issue of Computer Shopper in hand. People buy our hardware because we are *LINUX* people. We know Linux. We can choose the best hardware for Linux out of that vast array of commodity hardware just sitting on the shelf for the picking. We can configure Linux to best work on this hardware (and for the guy who says Red Hat 5.2 won't work with the 2.2 kernel, every single one of our SMP machines ships with the 2.2 kernel, and probably 90% of those are Red Hat 5.2). We can set up the automounter so that people don't have to mount and dismount floppies and CD-ROMs. We can install "X-CDRoast" when people buy a CD-R from us.
When every Joe Schmuck can put together a box that's every bit as good as what he could buy from Dell, Gateway, or LHS, we're no longer in a business where engineering is the difference. Rather, the difference is going to be service and quality of components. Fundamentally speaking, folks like VA Research, LHS, and Penguin sell the service of pulling together Linux-compatible components and installing Linux on the resulting computers. How well we do this is what detirmines our success or failure -- not how many components we manufacture ourselves.
Note: I'm talking about the "mainstream" market, between $1,000 and $30,000 in price... past $30,000, we're talking about engineering making a difference again, and below $1,000, you need massive quantities that Joe Schmuck can't do in his back room. Still, you get the point, right?
Let's put it this way. At the low end, Compaq has maybe a $50 profit on each machine that goes out the door. On the high end, they may mark it up by as much as 50% to cover the engineering costs (high end stuff is usually highly customized).
Now you know why Linux Hardware Solutions and VA Research aren't aiming at the low end, and why a large-volume distributor of Wintel computers is the first one to take Linux to the low end.
Personally, I'd love for a Linux-centric vendor to attack the low end. But the marketing is going to be a killer. I think the only way to make it work is to have a Linux equivalent of the iWhack, i.e., plug it in, turn it on, it comes up and asks you to set up your ISP, and the whole thing costs under $500. The only way to sell something like that is in the mass market, which takes $$$ to buy your way onto the shelves at Best Buy and Circuit City. The other possibility I see is for sort of a WebTV without the TV part, i.e., a diskless machine with a built-in monitor that operates off of flash disk that runs a browser and that's pretty much it. Think "iWhack" with a flash card instead of a disk drive. But again, we're talking about massive quantities needed to make a profit here, and a marketing strategy that requires millions of dollars to make work. I have actually played with a similar diskless "thin client", but right now it's too expensive for that kind of use. Right now it's just a diskless Linux workstation for use in, e.g., loading bays and other such high heat/high vibration areas where you don't want a hard drive. The only way to get the price down is by selling massive quantities, and the only way to do that is by investing massive amounts of money in advertising.
Anyhow, enough of that. (And note that I am DEFINITELY not speaking for Linux Hardware Solutions in this message... believe me, my boss just smiles and ignores me when I start raving about going after the consumer market !)
I haven't checked out his parts list, but behind hard drive (most expensive), usually comes processor and motherboard. If he's using a cheap Cyrix processor, the motherboard is probably the third most expensive item. Then memory and processor would come in at a tie for fourth most expensive item.
Linux SMP is not currently limited by by memory bandwidth on the 8-way stuff. They use multiple memory busses to gain extra bandwidth.
The real limit is simply that the Linux kernel's locking isn't granular enough. From listening to the discussions on the kernel list, the basic limitations seem to be in the filesystem and SCSI device drivers (you don't try doing a machine like this with IDE drives!). Theodore T'so popped up a while back saying he was going to work on making the filesystem work better w/SMP, but that was the last we heard of him (presumably he has discovered that it's harder than he thought!).
Anyhow: The VA Research machine is apparently, from reading their press releases, a machine that was developed by a Japanese manufacturer in conjunction with Intel (was it Hitachi?). I suspect that the Penguin machine is the exact same machine, just as the rest of their "big" servers are the exact same machines that VA Research sells (heck, the exact same machines that Linux Hardware Solutions sells as our dual processor Xeon workstation and quad processor Xeon servers, except that for our top-of the-line quad server we use the AMI platform rather than the Intel one, and we deck them out a bit differently as far as network card and RAID card). I haven't the foggiest clue who has been putting LSD into Sam's drinking water lately with these hallucinogenic press releases we've been seeing, but I must admit that I get a bit of a laugh out of them. Maybe he's doing like UserFriendly etc. and stringing the April Fools jokes out?
Just more MS FUD -- "Promise of Future Bliss" (I think that's #4 in the FUD 101 paper). Microsoft feels threatened by thin servers from Cobalt, HCC/Corel, and Linux Hardware Solutions (okay, so LHS isn't on their radar scope, grin, but we do have a thin server), so they're doing this vaporware "thin server" initiative in order to get customers to go with their promised future bliss rather than with already-shipping products like the Qube, Raq, or Netier.
The sad part is that it'll probably work. After all, it worked back in the days when IBM did it all the time against rivals Amdahl and Fujitsu, and I don't see that the buying public has gotten any smarter in the meantime.
First: VA Research doesn't sell Alpha because the demand for Alpha (or lack thereof) doesn't justify them spending resources on it. VA Research is no longer the little guy on the block with some guy puttering about in the back room. They sell thousands of identical systems, all of which are now being outsourced to a 3rd party manufacturing firm. (Nothing new there -- for example, the Commodore 64 was not built by Commodore, it was built by a Taiwanese outsourcing firm).
Secondly, Red Hat was just reflecting their sales data, not trying to insult Alpha on Compaq. Basically, the reason for running Alpha has decreased now that you can get 500Mhz PIII chips. A 500Mhz PIII chip comes within 15% of Alpha performance, while costing 25% less.
All that nonwithstanding, LHS still sells Alpha Linux systems for those few who do demand it. If we ever get into a situation where we must streamline our product line, though, the Alpha systems will be the first to go -- we just don't sell of them for them to matter much.
I agree with your assessment of Intel's commitment to Linux, BTW. Intel is out for #1 -- Intel. But that's true of all the large companies now coming into the Linux business. You think Dell cares about Linux? Heck no! All they care about is whether there's enough demand to put a few engineering dollars into creating a Linux line. Linux could be a toaster as far as they're concerned -- just another widget to sell.
Well, it said that my Matrox G200 was an OAK SVGA and refused to configure it at any resolution higher than 640x480 in 256 colors. It refused to recognize my Symbios 53c8xx-based SCSI card AT ALL, and nothing I tried would get it to recognize it no matter how much I wiped off the hard drive and reinstalled with the Symbios driver on a disk and etc. It didn't detect my old NE-2000 clone network card at all (non-PnP, but still...).
Red Hat 5.2 detected my video card fine, detected my SCSI card fine (and let me boot off of it!), in fact, the only way I could get Win98 to install on my machine (since it wouldn't detect my SCSI card, and my CD-ROM is hooked to the SCSI card) was to boot into Linux and copy everything to the FAT32 partition that I made with the Windows boot floppy.
Point: Auto-detection depends on what hardware you have in your system. Red Hat's installer is very good at auto-detecting PCI devices, not so good at detecting ISA devices, and doesn't do parallel device autodetection at all. For PCI devices, Red Hat actually auto-detects a larger percentage than Win98 does, in my experience. But if you still have a bunch of old PnP ISA devices, well, Win98 will detect them, Red Hat won't. Time to upgrade, eh?
Sherman's march and Lee's surrender ended the American Civil War, for example. But the rights of the blacks in the Southern United States could only be guaranteed by Federal troops. After ten years of occupation, the North had lost its stomach for that. The moment federal troops were withdrawn, the black population were back to being virtual slaves again.
Force of arms can win a war, but cannot change the hearts and minds of the losers of that war.
The main reason most of the bombings during WWII were against civilian targets was because of the policy of night air raids. Until the German air defenses were destroyed or at least significantly impaired, losses were too heavy during day raids for extended day bombing campaigns.
The technology of that day did not allow hitting individual targets during night raids. For that matter, day raids were no piece of cake either -- even with the famed Norden bombsight in broad daylight, American B-17's missed more often than they hit.
Go look at a map, fool. There's no way for a Caspian Sea pipeline to go through Yugoslavia to get to Turkey. The Caspian Sea is on the wrong side of Turkey for that to happen.
See http://www.eia.doe.gov/emeu/cabs/caspian.html for more info on the geography of the Caspian sea oilfields area.
Oh -- Noam Chomsky is going senile. Brilliant linguist, nice person, but he keeps thinking up bizarre things like oil pipelines that somehow skip over Turkey to get to Yugoslavia, then turn around and go back into Turkey. (huh? doesn't make sense? Now you get it!). Or else he's just doing like Ed Muth, and making it all up as he goes along. Either way, his web site is NOT providing facts, but, rather, is providing bizarre conspiracy theories similar to the right wing sites with their black helicopters baloney. Except that Noam is a left-wing conspiracy freak. Same difference.
When my FUD101 paper was referenced on Slashdot, we didn't even notice the load until somebody pointed out that FUD101 was on Slashdot.
Of course, the fact that everybody was at Comdex might have something to do with that too (grin).
Our web server box was (and is) a Pentium 166 with 64mb of RAM and a 1gb IDE hard drive. At the time it was hanging off of a single T1 line. I think I had just updated it to Red Hat 5.1 and the 2.0.35 kernel. Some things that probably helped it survive:
1) I don't do a lot of bulky graphics. Lots of bulky text, but not bulky graphics.
2) Everything was being statically served. No dynamically generated pages anywhere on the system.
3) When I updated to Red Hat 5.1, the Apache configuration got slaughtered but I didn't notice. Thus it was using the default Red Hat configuration. So it wasn't trying to do a reverse DNS lookup on each incoming connection, it was just logging the raw IP numbers. This greatly reduces the load on a system, at the expense of making it hard to see where people are coming from. (But I found a great utility that takes log files and "DNS-ifies" them, i.e., does the reverse lookup on all the #'s and creates a file as if Apache had its reverse lookups turned on, but without the load of having multiple Apaches all trying to do it at one time).
4) That box does nothing but serve web pages. No EMAIL, no FTP, no proxy server, nothing but web pages. When I logged in to find 100 EMAILs in my box from Slashdot readers, none of those EMAILs contributed to the load on the web server.
I think the important parts were that it was serving static pages, and the reverse DNS lookups were turned off. In general, Apache will saturate a T1 on quite modest hardware if you are just serving static pages. (And at the time, Slashdot had "just" a T1, meaning they couldn't hit me any harder than they were hitting Slashdot!). So if you have a 64mb Pentium 166 hanging around with a spare 1 gig IDE hard drive, don't worry about beefing up your hardware to survive a Slashdotting -- beef up your pipe (and get rid of those bulky CGI scripts!).
At this point, waiting is the right choice. 2.2 is settling out faster than 2.0 did (remember, 2.0 had EIGHTEEN kernels in three months!), but there will probably be a few more before it's really stable.
2.0 took 29 revisions to be really solid. Then it took another 5 revisions to be fixed again after 2.0.30 screwed everything up (groan).
Take a look at Arkadia Software or EST Inc.'s products for Linux (Arkadia Backup, BRU). Both support enterprise-wide backup and restore onto a central Linux server.
Look, I managed over 200 mixed Linux and SCO Unix boxes scattered across a two state area -- single-handedly, in less than 2 hours a day. Every night those systems backed up their data to their district office, even the ones not hooked up to a WAN (thanks to the power of UUCP and dialup modems!). Only an idiot can't do network-wide management of Linux boxes with tools like "rdist", "rsync", "ssh", etc. available.
My ISP charges $500/month for T1 service, and BellSouth charges $500/month for T1 access to their frame relay cloud so that I can get to my ISP. That's $1,000 per month. And that's not even direct T1, that's going through a frame relay cloud with a maximum guaranteed rate of 960,000 bits per second. Direct T1 access would add another $500/month to that charge.
If the government cut my taxes by $12,000 to $18,000, they'd be paying ME!
Moral of the story: Tax cuts aren't going to do it.
Newt might be a combative jerk, but he's not stupid.
BTW, I wouldn't read too much into his "T1" statement. I think he was trying to say that everybody needed access to T1-type speeds at home for a reasonable price in order to make the Internet work the way it should. Newt's not a technology person, so I'm sure he wasn't talking about using any particular technology to get those speeds.
Slashdot Mirror
that they had the Apache config set to allow up to 127 servers, and they had not raised file_max from 1024 to something decent. Do the math. Each Apache server has 8 file handles open just sitting there doing nothing. If it then goes to open a file to serve it, and file_max has been exceeded, guess what happens? Yep, Apache collapses!
-- Eric
Here is an SMB test on a large machine.
In general there are some areas where Linux lags NT. IIS, for example, outperforms Linux on static page displays because it has a page cache and does not have to always cross the user-kernel boundary to fetch the page (system calls DO have a cost, even though 2.2 sped up the open() call considerably with the dcache). And it may very well be that a well-tuned ultra-high-end NT machine will beat a well-tuned Linux machine at file serving, given that NT will support the full 4gb of memory while Linux only supports 2gb of memory. But this "test" was not such a test -- it compared a well-tuned NT machine against a totally untuned Linux machine.
And of course I'll point out that on tests on more modest hardware, like this, Linux blows away NT handily. To be fair, that Smart Reseller test was just as biased in its own way as this joke test we're talking about... Smart Reseller chose a machine that's too small for NT to comfortably stretch its legs, albeit that the machine they chose is rather typical of small office web servers.
-Eric
Thanks. I just checked that out. It does appear that they asked a single question about Apache performance. I remember seeing that posting myself and blowing it off because there wasn't enough info to tell him anything and I didn't feel like going into the give-and-take to get enough info to do something. (I do enough of that supporting my own customers!). Now, in hindsight, knowing what he did not do to Linux, the answer is obvious: he was running out of file handles. Do the math. An inactive Apache server has 8 file handles open. 127 servers max * 8 = 1016. Default file_max is 1024 for Linux, of which 150 or so are usually open while the system is at a rest. Apache could not bind a socket to a file handle for incoming connections because there were no file handles. So Apache was basically deadlocked, waiting for file handles to come free so it could accept() the socket but it was already holding all the file handles!
If there had been questions about general tuning of such a large system, that would have solved the problem because someone would have remembered about file_max. But one cryptic query that didn't give enough information to get help does not an honest effort make.
Anyhow: I guess I have to post a partial retraction. They did post a *SINGLE* query to the net.
-- Eric
I agree, it is hard to believe.
And: These people are *LIARS*. They say the posted messages asking for help on the Linux newsgroups. There are *NO* messages from the mindcraft domain anywhere on the Linux newsgroups. So I did DejaNews searches of "performance tuning", "performance tune", "kernel tuning", "kernel 2.2 tuning", between January 1 1999 and today, and examined the results to see if there were any messages that may have been by MindCraft researchers (i.e., that were referring to performance problems with a large-memory machine). There were *NONE*. Zero. Zilch. Which means that if they did ask any performance tuning questions, they did not use those words in the message.
Anyhow: VA Research already loaned a quad-processor Xeon machine to PC-Week and it blew away NT 4.0 in their SAMBA benchmarks. VA Research's quad-processor Xeon machine is the same machine that we sell, and the same machine that Penguin Computing sells (we all get them from Intel, and then dress them slightly differently once we get them, e.g. VA Research uses a Mylex RAID card while we and Penguin use ICP-Vortex RAID cards). So we already have the benchmark that shows that their SAMBA benchmark is full of ****. But that's not going to matter to pointy-haired bosses because they recognize only those reports and studies that say what they want to hear.
Am I steamed? You bet! I *HATE* liars!
-- Eric
However: There are more web app servers on the way for Linux. The CORBA plug-in for Apache should be a clue, eh?
-- Eric
GISA (General Internet Servelet Architecture) will also be a boon here, assuming it ever gets off the ground.
-- Eric
1) Dell is almost totally a "white box" operation, with the exception of the very high end of their line. In fact, the Compaq guys make fun of Dell because so much of Dell's stuff is just off-the-shell commodity hardware. The Dell guys just smile all the way to the bank.
2) I assure you that LHS and VA Research, at least, have people on staff who came from Compaq, Apple, etc. and are quite familiar with the design centers at those compoanies.
3) VA Research no longer builds their own low-end computers. Those have been outsourced to Flextronics (or did you miss their news release?). Most computer companies do this. Very few low-end computers are built by the company whose badge is on the front cover. (Heck, even the iWhack is contracted out to such an outsourcing firm!).
4) In general, almost all facets of designing and building a computer can be outsourced now. I can't tell you details (sorry, NDA), but basically it becomes a game of where your design dollars should go. In the $1K-$30K range Compaq cannot build better motherboards than ASUS or Intel, and in fact they don't. What they do accomplish is creating motherboards that are cheaper to add into manufactured computers. Ever wonder why so many components are integrated into Compaq motherboards? Well, that saves a few dollars because those components can be automagically placed onto the motherboard by robot pick'n'place machines, rather than having some poor slob have to spend thirty seconds slamming a video card into a computer and spinning the retaining screw home.
But the thing about outsourcing is that the poor slob is now in Malaysia or Thailand and is being paid $3 per day. Given that, why invest design dollars creating a board that reduces manufacturing costs, but which is actually technically inferior to what can be bought off the shelf from ASUS or Intel? (I say technically inferior because integrated peripherals reduce the flexibility and repairability of a computer, both of which are qualities which I personally value).
5) All that nonwithstanding, you are correct about being able to put together a system equal to that of VA Research, Dell, or Linux Hardware Solutions out of off-the-shelf parts. In fact, any Joe Consultant in Cheyenne Wyoming can do the same. That is what is so amazing about today as vs. 20 years ago, when that was definitely not the case.
Given that, vendors like VAR, LHS, etc. are not going to win by trying to out-engineer ASUS and Intel. Engineering resources have to be carefully allocated to those areas where off-the-shelf hardware currently doesn't exist (I don't think I can say more there, NDA etc). In the mainstream $1K-$30K server market, though, where any Joe Schmuck can buy the exact same parts off the shelf, it's the services of systems integration and support that we sell. We sort through the dozens of video cards and network cards and etc. out there so you don't have to, and then set things up and often times hunt up or write drivers for things that need it (like for the Symbios 53c896-based stuff or the Mylex stuff). If you don't need those services, build your own for crying out loud! That's why we put the parts lists up on our web sites, after all. But don't diss those of us providing a valuable service just because you don't need that particular service!
-- Eric
Basic limitation is how to share processors? DUH!
Cache coherency etc. is not part of the problem. The problem is the granularity of the semaphoring within the Linux kernel. That is, how much of the kernel can be running in parallel. Right now, once you get up past 4 processors, the kernel becomes the limit.
There are also some process control and migration issues, as well as memory management, all of which can be lumped under the banner of "scheduling issues". Matt Dillon of the FreeBSD project did an interesting analysis of the Linux VM, for example, noting how it was a simple and elegant design (and making suggestions for doing similar simplifications to the FreeBSD VM subsystem) but somewhat simple-minded with some obvious performance problems when dealing with large memory sets (like when we're talking 8-way processors!).
Anyhow: the Linux kernel guys know what needs doing. It's just that they're now at the hard part -- going from protecting entire subsystems with semaphors (so that they run on only a single processor) to allowing those subsystems to run in parallel on multiple processors (by protecting the data structures within those subsystems with semaphors). The known scheduling and VM issues need addressing too, of course. All of which is on the table for 2.3.
-- Eric
Today, thanks to the mass market in commodity components created by the IBM PC platform (see The Commoditization of Computers), any Joe Consultant in Hoboken, Michigan can put together his own computers that are every bit as high in quality and low in price as those from Dell. It's an amazing democratization of the computer industry, totally unlike anything that has ever happened in any other industry. Suddenly any schmuck off the street can build a computer just as good as what he can buy, often for less!
Given all that, folks like VA Research, Penguin, or Linux Hardware Solutions would have to be nuts to design their own motherboards. People don't buy our hardware because it is somehow better than what Joe Schmuck can put together in his back room with an issue of Computer Shopper in hand. People buy our hardware because we are *LINUX* people. We know Linux. We can choose the best hardware for Linux out of that vast array of commodity hardware just sitting on the shelf for the picking. We can configure Linux to best work on this hardware (and for the guy who says Red Hat 5.2 won't work with the 2.2 kernel, every single one of our SMP machines ships with the 2.2 kernel, and probably 90% of those are Red Hat 5.2). We can set up the automounter so that people don't have to mount and dismount floppies and CD-ROMs. We can install "X-CDRoast" when people buy a CD-R from us.
When every Joe Schmuck can put together a box that's every bit as good as what he could buy from Dell, Gateway, or LHS, we're no longer in a business where engineering is the difference. Rather, the difference is going to be service and quality of components. Fundamentally speaking, folks like VA Research, LHS, and Penguin sell the service of pulling together Linux-compatible components and installing Linux on the resulting computers. How well we do this is what detirmines our success or failure -- not how many components we manufacture ourselves.
Note: I'm talking about the "mainstream" market, between $1,000 and $30,000 in price... past $30,000, we're talking about engineering making a difference again, and below $1,000, you need massive quantities that Joe Schmuck can't do in his back room. Still, you get the point, right?
--Eric
Now you know why Linux Hardware Solutions and VA Research aren't aiming at the low end, and why a large-volume distributor of Wintel computers is the first one to take Linux to the low end.
Personally, I'd love for a Linux-centric vendor to attack the low end. But the marketing is going to be a killer. I think the only way to make it work is to have a Linux equivalent of the iWhack, i.e., plug it in, turn it on, it comes up and asks you to set up your ISP, and the whole thing costs under $500. The only way to sell something like that is in the mass market, which takes $$$ to buy your way onto the shelves at Best Buy and Circuit City. The other possibility I see is for sort of a WebTV without the TV part, i.e., a diskless machine with a built-in monitor that operates off of flash disk that runs a browser and that's pretty much it. Think "iWhack" with a flash card instead of a disk drive. But again, we're talking about massive quantities needed to make a profit here, and a marketing strategy that requires millions of dollars to make work. I have actually played with a similar diskless "thin client", but right now it's too expensive for that kind of use. Right now it's just a diskless Linux workstation for use in, e.g., loading bays and other such high heat/high vibration areas where you don't want a hard drive. The only way to get the price down is by selling massive quantities, and the only way to do that is by investing massive amounts of money in advertising.
Anyhow, enough of that. (And note that I am DEFINITELY not speaking for Linux Hardware Solutions in this message... believe me, my boss just smiles and ignores me when I start raving about going after the consumer market !)
-Eric
-Eric
The real limit is simply that the Linux kernel's locking isn't granular enough. From listening to the discussions on the kernel list, the basic limitations seem to be in the filesystem and SCSI device drivers (you don't try doing a machine like this with IDE drives!). Theodore T'so popped up a while back saying he was going to work on making the filesystem work better w/SMP, but that was the last we heard of him (presumably he has discovered that it's harder than he thought!).
Anyhow: The VA Research machine is apparently, from reading their press releases, a machine that was developed by a Japanese manufacturer in conjunction with Intel (was it Hitachi?). I suspect that the Penguin machine is the exact same machine, just as the rest of their "big" servers are the exact same machines that VA Research sells (heck, the exact same machines that Linux Hardware Solutions sells as our dual processor Xeon workstation and quad processor Xeon servers, except that for our top-of the-line quad server we use the AMI platform rather than the Intel one, and we deck them out a bit differently as far as network card and RAID card). I haven't the foggiest clue who has been putting LSD into Sam's drinking water lately with these hallucinogenic press releases we've been seeing, but I must admit that I get a bit of a laugh out of them. Maybe he's doing like UserFriendly etc. and stringing the April Fools jokes out?
Eric.
The sad part is that it'll probably work. After all, it worked back in the days when IBM did it all the time against rivals Amdahl and Fujitsu, and I don't see that the buying public has gotten any smarter in the meantime.
-- Eric
First: VA Research doesn't sell Alpha because the demand for Alpha (or lack thereof) doesn't justify them spending resources on it. VA Research is no longer the little guy on the block with some guy puttering about in the back room. They sell thousands of identical systems, all of which are now being outsourced to a 3rd party manufacturing firm. (Nothing new there -- for example, the Commodore 64 was not built by Commodore, it was built by a Taiwanese outsourcing firm).
Secondly, Red Hat was just reflecting their sales data, not trying to insult Alpha on Compaq. Basically, the reason for running Alpha has decreased now that you can get 500Mhz PIII chips. A 500Mhz PIII chip comes within 15% of Alpha performance, while costing 25% less.
All that nonwithstanding, LHS still sells Alpha Linux systems for those few who do demand it. If we ever get into a situation where we must streamline our product line, though, the Alpha systems will be the first to go -- we just don't sell of them for them to matter much.
I agree with your assessment of Intel's commitment to Linux, BTW. Intel is out for #1 -- Intel. But that's true of all the large companies now coming into the Linux business. You think Dell cares about Linux? Heck no! All they care about is whether there's enough demand to put a few engineering dollars into creating a Linux line. Linux could be a toaster as far as they're concerned -- just another widget to sell.
-- Eric
Linux is easy to install if you buy a real computer instead of a DELL.
For example, Linux installs on one of our computers in less than 15 minutes flat.
The DELL situation will be remedied whenever DELL gives a damn. Until then, don't buy DELL. Simple, see?
Why on Earth buy a machine labeled "Designed for Windows 98" to run Linux?!
-- E
Let's see what my Win98 did:
Well, it said that my Matrox G200 was an OAK SVGA and refused to configure it at any resolution higher than 640x480 in 256 colors. It refused to recognize my Symbios 53c8xx-based SCSI card AT ALL, and nothing I tried would get it to recognize it no matter how much I wiped off the hard drive and reinstalled with the Symbios driver on a disk and etc. It didn't detect my old NE-2000 clone network card at all (non-PnP, but still...).
Red Hat 5.2 detected my video card fine, detected my SCSI card fine (and let me boot off of it!), in fact, the only way I could get Win98 to install on my machine (since it wouldn't detect my SCSI card, and my CD-ROM is hooked to the SCSI card) was to boot into Linux and copy everything to the FAT32 partition that I made with the Windows boot floppy.
Point: Auto-detection depends on what hardware you have in your system. Red Hat's installer is very good at auto-detecting PCI devices, not so good at detecting ISA devices, and doesn't do parallel device autodetection at all. For PCI devices, Red Hat actually auto-detects a larger percentage than Win98 does, in my experience. But if you still have a bunch of old PnP ISA devices, well, Win98 will detect them, Red Hat won't. Time to upgrade, eh?
-- Eric
Unfortunately, I have to agree.
Sherman's march and Lee's surrender ended the American Civil War, for example. But the rights of the blacks in the Southern United States could only be guaranteed by Federal troops. After ten years of occupation, the North had lost its stomach for that. The moment federal troops were withdrawn, the black population were back to being virtual slaves again.
Force of arms can win a war, but cannot change the hearts and minds of the losers of that war.
-E
The main reason most of the bombings during WWII were against civilian targets was because of the policy of night air raids. Until the German air defenses were destroyed or at least significantly impaired, losses were too heavy during day raids for extended day bombing campaigns.
The technology of that day did not allow hitting individual targets during night raids. For that matter, day raids were no piece of cake either -- even with the famed Norden bombsight in broad daylight, American B-17's missed more often than they hit.
Yugoslavia is to the WEST of Turkey.
The Caspian Sea is to the EAST of Turkey.
Go look at a map, fool. There's no way for a Caspian Sea pipeline to go through Yugoslavia to get to Turkey. The Caspian Sea is on the wrong side of Turkey for that to happen.
See http://www.eia.doe.gov/emeu/cabs/caspian.html for more info on the geography of the Caspian sea oilfields area.
Oh -- Noam Chomsky is going senile. Brilliant linguist, nice person, but he keeps thinking up bizarre things like oil pipelines that somehow skip over Turkey to get to Yugoslavia, then turn around and go back into Turkey. (huh? doesn't make sense? Now you get it!). Or else he's just doing like Ed Muth, and making it all up as he goes along. Either way, his web site is NOT providing facts, but, rather, is providing bizarre conspiracy theories similar to the right wing sites with their black helicopters baloney. Except that Noam is a left-wing conspiracy freak. Same difference.
-- Eric
When my FUD101 paper was referenced on Slashdot, we didn't even notice the load until somebody pointed out that FUD101 was on Slashdot.
Of course, the fact that everybody was at Comdex might have something to do with that too (grin).
Our web server box was (and is) a Pentium 166 with 64mb of RAM and a 1gb IDE hard drive. At the time it was hanging off of a single T1 line. I think I had just updated it to Red Hat 5.1 and the 2.0.35 kernel. Some things that probably helped it survive:
1) I don't do a lot of bulky graphics. Lots of bulky text, but not bulky graphics.
2) Everything was being statically served. No dynamically generated pages anywhere on the system.
3) When I updated to Red Hat 5.1, the Apache configuration got slaughtered but I didn't notice. Thus it was using the default Red Hat configuration. So it wasn't trying to do a reverse DNS lookup on each incoming connection, it was just logging the raw IP numbers. This greatly reduces the load on a system, at the expense of making it hard to see where people are coming from. (But I found a great utility that takes log files and "DNS-ifies" them, i.e., does the reverse lookup on all the #'s and creates a file as if Apache had its reverse lookups turned on, but without the load of having multiple Apaches all trying to do it at one time).
4) That box does nothing but serve web pages. No EMAIL, no FTP, no proxy server, nothing but web pages. When I logged in to find 100 EMAILs in my box from Slashdot readers, none of those EMAILs contributed to the load on the web server.
I think the important parts were that it was serving static pages, and the reverse DNS lookups were turned off. In general, Apache will saturate a T1 on quite modest hardware if you are just serving static pages. (And at the time, Slashdot had "just" a T1, meaning they couldn't hit me any harder than they were hitting Slashdot!). So if you have a 64mb Pentium 166 hanging around with a spare 1 gig IDE hard drive, don't worry about beefing up your hardware to survive a Slashdotting -- beef up your pipe (and get rid of those bulky CGI scripts!).
-- Eric
At this point, waiting is the right choice. 2.2 is settling out faster than 2.0 did (remember, 2.0 had EIGHTEEN kernels in three months!), but there will probably be a few more before it's really stable.
2.0 took 29 revisions to be really solid. Then it took another 5 revisions to be fixed again after 2.0.30 screwed everything up (groan).
Take a look at Arkadia Software or EST Inc.'s products for Linux (Arkadia Backup, BRU). Both support enterprise-wide backup and restore onto a central Linux server.
Look, I managed over 200 mixed Linux and SCO Unix boxes scattered across a two state area -- single-handedly, in less than 2 hours a day. Every night those systems backed up their data to their district office, even the ones not hooked up to a WAN (thanks to the power of UUCP and dialup modems!). Only an idiot can't do network-wide management of Linux boxes with tools like "rdist", "rsync", "ssh", etc. available.
-- Eric
A man who divorces his wife because she's sick in the hospital with cancer is HONORABLE?
Newt's smart, but he's also a real jerk.
My ISP charges $500/month for T1 service, and BellSouth charges $500/month for T1 access to their frame relay cloud so that I can get to my ISP. That's $1,000 per month. And that's not even direct T1, that's going through a frame relay cloud with a maximum guaranteed rate of 960,000 bits per second. Direct T1 access would add another $500/month to that charge.
If the government cut my taxes by $12,000 to $18,000, they'd be paying ME!
Moral of the story: Tax cuts aren't going to do it.
-- Eric
Newt might be a combative jerk, but he's not stupid.
BTW, I wouldn't read too much into his "T1" statement. I think he was trying to say that everybody needed access to T1-type speeds at home for a reasonable price in order to make the Internet work the way it should. Newt's not a technology person, so I'm sure he wasn't talking about using any particular technology to get those speeds.
-- Eric