Protein folding, for starters. The point you're missing is that a beowulf class supercomputer is -MUCH- cheaper to put together than some "super-powerful" single machine (forgetting that most of the -really- big machines are themselves clusters of a sort).
Money played (and plays) a rather large role in the whole COTS supercomputing arena:)
Head Node: Cluster users connect to this node to kick off their jobs onto the rest of the cluster. Basically, it's the enduser accessible machine.
Management node: cluster monitoring tools, image server, etc.
On a -small- cluster (16 or 32 nodes, although 32 is pushing it, IMHO), the head and management nodes are typically the same machine. There really isn't a need for there to be more than that, as it's fairly quick and easy to monitor all of 32 machines. On larger clusters (64/128/etc), it starts becoming almost -imperative- to seperate purposes, as you don't want management tasks to eat into the bandwidth available for computation. It's also why the management node(s) is(are) occasionally on it's seperated network (100bT is fine for it, BTW) subsystems.
Having said all that, I'm *really* looking forward to Opteron. We're getting some in a week or so. 64 bit + SSE2 support is going hard to beat. This is -specifically- why I'm waiting until the next Opteron stepping before specing out our next cluster. The current rig is P4 Xeons (1.8GHz), and is "ok" performance wise, but that's ok, since it was more a "proof of concept" type of thing than a real install *heh*
Re:Alabanza sucks ass.
on
C&W Bails Out
·
· Score: 1
Who was your provider? There were several who seemed to have had those sorts of complaints, even though the -vast- majority of their clients seem to be quite happy with them. *shrugs* I really don't care as I haven't worked there for 6 months.
Re:This really doesn't make sense....
on
C&W Bails Out
·
· Score: 2, Interesting
This is -precisely- why I keep scratching my head when I see folks advertising "CoLo $100/mo no bandwidth limit!". There's absolutely -no- way they're making money. At -all-. An OC3 runs 12-15 thousand a month, minimum, now imagine a company that has 10 datacenters with 4 or 5 of those into -each-, and each one has multi-million dollar security systems, 3 dozen employees, etc. I can easily see C&W losing a million bucks a day. It's kinda funny. Folks have been saying for -years- that my previous employer's pricing was out of line with "reality" and was "too expensive". Come to find out, they were pricing things almost -precisely- right, and not caving to pressure from outside to cut prices, and they're -profitable- (and have been for, like, 2 years now)...the company is 7 or 8 years old). However, they -don't- do CoLo. There's no money in it. Managed hosting and hosting services, on the other hand, is a cash cow that -very- few companies are doing (or doing -well-).
*shrug* I'm not suprised. Exodus imploded, Verio's been in the process of imploding, Rackspace supplements their income through sales of hardware and cuts their costs by building their own rigs, Alabanza is profitable, Interland may as well have imploded, etc. Basically, Alabanza and Rackspace are the two winners out of the entire industry, and -noone- is really competing well with either of them.
*drools* I'd like to have a 4 way opty...oh yesh...yesh I would. So would the research group I'm attached to, too *laff*
You're right, though. Multiple paths is definately the way to be when doing HPC. I -do- want a chance to futz around with Red Storm though (never happen, but I can dream).
Look at the average Joe Schmoe, or even us uber-users, who really needs a 3+ GHz machine? Even some of the cornerstones of fast computing such as computational problem solving are being addressed by grid/cluster based solutions which typically don't use high end machines.
Bullshit. While I'm not going to comment -too- much on the "who really needs a 3+GHz machine" it needs to be said that gamers tend to push them. It -also- needs to be said that developers on large projects will also push them. Downtime while compiling -sucks-, man.
As to the rest of it, you obviously have no clue how large hardware orders are handled, do you? It typically takes a few months from the time a machine is spec'd out and purchased, to delivery, assembly, burn in and release. I can -easily- see this process take 12 months when talking about larger clusters. So, when it's ordered, it's likely with the most power procs available AT THAT TIME. ie. If I were to order a new cluster, right now, I would likely be going for dual Opty 244's. We -all- know that the next stepping is due out sometime in the next 6 weeks, but the order has to be placed by Friday due to budget issues. (that said, I'm -not- ordering such a machine at this time, it's just an example).
Actually, when you say you did you take a look at the top 500 list, you should put actions behind your words. The top cluster is at #5 on the most recent list (LLNL's NetworX machine - http://www.top500.org/list/2002/11/), and is less than 20% behind the #2 spot. Guaranteed that within a year, linux clusters will indeed fill the #2 spot on down.
Actually, I'll have to disagree with this, to a point.
I expect to see no -more- than half the top 10 be linux clusters. The reason for that is that most of the folks that I've seen using them are content with a max of 32 nodes (-generally- 64 procs). The really big clusters are being built at LLNL, LANL, etc, and they're still primarily using Power-based systems outside of the occasional Linux Beowulf...
That said, LANL is supposed to be getting a new Beowulf in the next 6-12 months, and supposedly, it kicks much ass.
What spead Athlon are you using? And don't forget that the HPC/Cluster folks are absolutely -drooling- over the Opteron (which -CRAY- is putting into their next Big system). The I2 has been getting panned:)
I assume you're speaking of the Top500.org list, yes? Since that -is- the list most people refer to:)
Now, to break it down: 1. Earth Simulator: classive vector super computer. 2 & 3. ASCI Q: Alpha Cluster. 4. ASCI White: Power3 rig, not sure if it's a cluster or not... 5. MCR Linux Cluster: name says it all. 6 & 7: Both are Alpha clusters. 8. HPTi: Linux Beowulf. 9 & 10: IBM p690's.
So, out of the top 10, -6- are clusters, not all but 1. I don't know what list you're using, but it certainly isn't the posted one (if you're using the not yet released list, I'm sure the ISC folks would be interested to find out how you got it, though).
I -do- agree that using the high speed interconnects tends to price theses systems out of the COTS category, though *heh* Myrinet cards are something like US$1200 -each-.
Oh I'm with you on that. Now to convince my wife that I really -do- need that 75 acre farm that's for sale just up the road... *twitch*
The best beers I've ever had are "craft" brewed, either by myself (and I make what a number of professional beer adjudicators consider a world class Altbier....and there -are- brewing/beer competitions around the country) or from a local brewpub (like Cap City in DC and Baltimore). The mass produced shit is good to get a buzz going, and that's about it. When I'm in the mood for a Beer, I spend the money on the good stuff (if it's -less- than 5 bucks a bottle it's not something I'm going to spend the time enjoying...except for Murphy's and Guiness:)
Re:GPL the best bet
on
OSI vs SCO
·
· Score: 2, Interesting
Considering that SCO/Caldera is still distributing the SRPMS for OpenServer 3.1.1. Now, considering that SCO is claiming damages in pre-IBM "involvement" as well as post-IBM "involvement" in kernel development, they're pretty much pooched.
Basically, since SCO/Caldera now owns that code, and they are still distributing it, they are doing so with the full knowledge and understanding that they are bound by the GPL. Pretty much, they're fucked. They may very well be -right- (although I think it more likely that Linux kernel code made its way into Unixware), but they're hosed by the GPL (as they would if it were BSD licensed):)
Pretty much. We decided that the only way to effectively roll SA out in their environment would have required setting up a number of different machines for it. For those clients that -really- wanted it, they look the other way. Basically, with as small as the admin team is, there simply isn't the man power (and to sound egotistical, skill, they've had to hire a couple of folks to replace me since I left). Incidentally: 330 servers is a small fraction of the number of machines said previous employer had 6 months ago. They've more, now.
Actually, open source is employing about a hundred and fifty people at at least one company in Baltimore, MD (and I hear they're looking for a couple of -good- SysAdmins). Add to that all the sysadmins at the various research universities that are running Beowulf and Mosix/openMosix clusters, and you start to see a different picture.
Honestly, I personally haven't had a problem finding a job...
IIRC, the postmaster@ at AOL goes to the same group that handles spam for them. It's been a while since I talked with any of those individuals though (I'm out of the anti-spam game. too much stress...seriously). Basically, postmaster is received, but probably not -replied to- by anyone because of the sheer volume of crap that comes in to those addresses. Having -been- an abuse desk, it gets read, and added to the complaint count for whoever you're complaining about.
'sides, the RFC only requires that it be deliverable. Not read.
Um, no. Having worked for a reasonably large hosting company (largest in the world based on number of active IP addresses...by domains it's still a pipsqueak, but not insignificant) I can say, with confidence, that SpamAssassin will -not- work in that environment. Not easily, at least. And yes, deploying it was considered, but the CPU hit (as mentioned) would kill all of the servers (since they were all Web/App/Mail in a box type set ups).
And to add to this, folks doing bioinformatic and genetic research will -=really=- put the Opteron to use, and for a -hell- of a lot less than Itaniu (and I2).
And no, Engineers -aren't- the only ones that could use them on their desktops. You also forgot graphics designers and testbed machines for applications programmers, and testbed machines for a larger 64bit environment.:P
And you're getting this pricing for the Opteron from where? Everything I've been hearing has the low end stuff -=undercutting=- the current Xeon pricing, while still running more than standard Athlons/P4's.
There's a reason when you look in a "pro" kitchen -everything- is gas powered. If it's not, find a different restaurant to eat at. Gas is -always- better than electric.
Period.
And yes, a bit of experience cooking on gas will teach you how to gauge flame height/intensity and temperature. One thing I -do- want to note, though: try not to let the flame lick up the sides of your pots and pans. Seriously.
How important was that cluster to computational research? How important is that cluster to the mission of the school? I'm not asking facetious questions here, I'm honestly curious.
The machines that I work with support millions of dollars worth of research projects. Having unskilled, or even semi-skilled students futzing around as root on theses systems could be potentially disastrous (even -with- me backing everything up nightly).
We were employed by the university This is -exactly- what I -don't- mind. Students who's job and responsibility it is to admin a machine are fine by me. Well, with one caveat: there's some sort of oversight from a more senior admin (be he/she another student, or a pro).
I'm not against students gaining experience, and being given responsibility of important/semi-important machines. What I'm against are students who, just because they know something about computers, are given root on the university's central servers to do with as they will without oversight.
To Paraphrase: if you're paid job as a student employee of a college or university is to admin a couple of machines, cool. There's -going- to be some oversight (at least there should be) as well as consequences for actions taken. You've also been through some sort of an interview process to (hopefully!) only put someone who's capable of doing the job in place.
And what is the machine being used for? If the -job- of that computer is to teach the students systems administration, then more power to 'em. If it's a system out in a student computing lab, or a system used primarily for the programming classes or whatever, then no, students should -not- have access. There's absolutely no reason.
Slashdot Mirror
Protein folding, for starters.
:)
The point you're missing is that a beowulf class supercomputer is -MUCH- cheaper to put together than some "super-powerful" single machine (forgetting that most of the -really- big machines are themselves clusters of a sort).
Money played (and plays) a rather large role in the whole COTS supercomputing arena
Head Node:
Cluster users connect to this node to kick off their jobs onto the rest of the cluster. Basically, it's the enduser accessible machine.
Management node: cluster monitoring tools, image server, etc.
On a -small- cluster (16 or 32 nodes, although 32 is pushing it, IMHO), the head and management nodes are typically the same machine. There really isn't a need for there to be more than that, as it's fairly quick and easy to monitor all of 32 machines. On larger clusters (64/128/etc), it starts becoming almost -imperative- to seperate purposes, as you don't want management tasks to eat into the bandwidth available for computation. It's also why the management node(s) is(are) occasionally on it's seperated network (100bT is fine for it, BTW) subsystems.
*shrugs*
Having said all that, I'm *really* looking forward to Opteron. We're getting some in a week or so. 64 bit + SSE2 support is going hard to beat.
This is -specifically- why I'm waiting until the next Opteron stepping before specing out our next cluster.
The current rig is P4 Xeons (1.8GHz), and is "ok" performance wise, but that's ok, since it was more a "proof of concept" type of thing than a real install *heh*
Who was your provider? There were several who seemed to have had those sorts of complaints, even though the -vast- majority of their clients seem to be quite happy with them.
*shrugs*
I really don't care as I haven't worked there for 6 months.
This is -precisely- why I keep scratching my head when I see folks advertising "CoLo $100/mo no bandwidth limit!". There's absolutely -no- way they're making money. At -all-. An OC3 runs 12-15 thousand a month, minimum, now imagine a company that has 10 datacenters with 4 or 5 of those into -each-, and each one has multi-million dollar security systems, 3 dozen employees, etc.
I can easily see C&W losing a million bucks a day.
It's kinda funny. Folks have been saying for -years- that my previous employer's pricing was out of line with "reality" and was "too expensive". Come to find out, they were pricing things almost -precisely- right, and not caving to pressure from outside to cut prices, and they're -profitable- (and have been for, like, 2 years now)...the company is 7 or 8 years old).
However, they -don't- do CoLo. There's no money in it. Managed hosting and hosting services, on the other hand, is a cash cow that -very- few companies are doing (or doing -well-).
*shrug*
I'm not suprised. Exodus imploded, Verio's been in the process of imploding, Rackspace supplements their income through sales of hardware and cuts their costs by building their own rigs, Alabanza is profitable, Interland may as well have imploded, etc. Basically, Alabanza and Rackspace are the two winners out of the entire industry, and -noone- is really competing well with either of them.
*drools*
I'd like to have a 4 way opty...oh yesh...yesh I would. So would the research group I'm attached to, too *laff*
You're right, though. Multiple paths is definately the way to be when doing HPC. I -do- want a chance to futz around with Red Storm though (never happen, but I can dream).
Look at the average Joe Schmoe, or even us uber-users, who really needs a 3+ GHz machine? Even some of the cornerstones of fast computing such as computational problem solving are being addressed by grid/cluster based solutions which typically don't use high end machines.
Bullshit.
While I'm not going to comment -too- much on the "who really needs a 3+GHz machine" it needs to be said that gamers tend to push them. It -also- needs to be said that developers on large projects will also push them. Downtime while compiling -sucks-, man.
As to the rest of it, you obviously have no clue how large hardware orders are handled, do you?
It typically takes a few months from the time a machine is spec'd out and purchased, to delivery, assembly, burn in and release. I can -easily- see this process take 12 months when talking about larger clusters.
So, when it's ordered, it's likely with the most power procs available AT THAT TIME. ie. If I were to order a new cluster, right now, I would likely be going for dual Opty 244's. We -all- know that the next stepping is due out sometime in the next 6 weeks, but the order has to be placed by Friday due to budget issues. (that said, I'm -not- ordering such a machine at this time, it's just an example).
*shakes head*
Actually, when you say you did you take a look at the top 500 list, you should put actions behind your words. The top cluster is at #5 on the most recent list (LLNL's NetworX machine - http://www.top500.org/list/2002/11/), and is less than 20% behind the #2 spot. Guaranteed that within a year, linux clusters will indeed fill the #2 spot on down.
Actually, I'll have to disagree with this, to a point.
I expect to see no -more- than half the top 10 be linux clusters. The reason for that is that most of the folks that I've seen using them are content with a max of 32 nodes (-generally- 64 procs). The really big clusters are being built at LLNL, LANL, etc, and they're still primarily using Power-based systems outside of the occasional Linux Beowulf...
That said, LANL is supposed to be getting a new Beowulf in the next 6-12 months, and supposedly, it kicks much ass.
What spead Athlon are you using? And don't forget that the HPC/Cluster folks are absolutely -drooling- over the Opteron (which -CRAY- is putting into their next Big system). The I2 has been getting panned :)
:)
*shrug* I only admin the bloody things
I assume you're speaking of the Top500.org list, yes? Since that -is- the list most people refer to :)
Now, to break it down:
1. Earth Simulator: classive vector super computer.
2 & 3. ASCI Q: Alpha Cluster.
4. ASCI White: Power3 rig, not sure if it's a cluster or not...
5. MCR Linux Cluster: name says it all.
6 & 7: Both are Alpha clusters.
8. HPTi: Linux Beowulf.
9 & 10: IBM p690's.
So, out of the top 10, -6- are clusters, not all but 1. I don't know what list you're using, but it certainly isn't the posted one (if you're using the not yet released list, I'm sure the ISC folks would be interested to find out how you got it, though).
I -do- agree that using the high speed interconnects tends to price theses systems out of the COTS category, though *heh* Myrinet cards are something like US$1200 -each-.
Seriously :) :)
I -like- the smell of wort bubbling with the various hops going on
*ponders brewing this weekend*
*remembers he has a concert to go to in DC on Saturday*
*heee!*
Oh I'm with you on that.
:)
Now to convince my wife that I really -do- need that 75 acre farm that's for sale just up the road...
*twitch*
The best beers I've ever had are "craft" brewed, either by myself (and I make what a number of professional beer adjudicators consider a world class Altbier....and there -are- brewing/beer competitions around the country) or from a local brewpub (like Cap City in DC and Baltimore). The mass produced shit is good to get a buzz going, and that's about it. When I'm in the mood for a Beer, I spend the money on the good stuff (if it's -less- than 5 bucks a bottle it's not something I'm going to spend the time enjoying...except for Murphy's and Guiness
Considering that SCO/Caldera is still distributing the SRPMS for OpenServer 3.1.1. Now, considering that SCO is claiming damages in pre-IBM "involvement" as well as post-IBM "involvement" in kernel development, they're pretty much pooched.
Basically, since SCO/Caldera now owns that code, and they are still distributing it, they are doing so with the full knowledge and understanding that they are bound by the GPL. Pretty much, they're fucked. They may very well be -right- (although I think it more likely that Linux kernel code made its way into Unixware), but they're hosed by the GPL (as they would if it were BSD licensed):)
Pretty much.
We decided that the only way to effectively roll SA out in their environment would have required setting up a number of different machines for it. For those clients that -really- wanted it, they look the other way.
Basically, with as small as the admin team is, there simply isn't the man power (and to sound egotistical, skill, they've had to hire a couple of folks to replace me since I left). Incidentally: 330 servers is a small fraction of the number of machines said previous employer had 6 months ago.
They've more, now.
*snicker* You'd like to think that.
I've written more code as a SysAdmin then most of the professional developers I've run into, that are at the same point in their careers.
Troll.
Actually, open source is employing about a hundred and fifty people at at least one company in Baltimore, MD (and I hear they're looking for a couple of -good- SysAdmins). Add to that all the sysadmins at the various research universities that are running Beowulf and Mosix/openMosix clusters, and you start to see a different picture.
Honestly, I personally haven't had a problem finding a job...
IIRC, the postmaster@ at AOL goes to the same group that handles spam for them. It's been a while since I talked with any of those individuals though (I'm out of the anti-spam game. too much stress...seriously).
Basically, postmaster is received, but probably not -replied to- by anyone because of the sheer volume of crap that comes in to those addresses. Having -been- an abuse desk, it gets read, and added to the complaint count for whoever you're complaining about.
'sides, the RFC only requires that it be deliverable. Not read.
Um, no.
Having worked for a reasonably large hosting company (largest in the world based on number of active IP addresses...by domains it's still a pipsqueak, but not insignificant) I can say, with confidence, that SpamAssassin will -not- work in that environment.
Not easily, at least. And yes, deploying it was considered, but the CPU hit (as mentioned) would kill all of the servers (since they were all Web/App/Mail in a box type set ups).
What area are you in? None of the jobs i've been looking at with 5 yrs pay less than 45. I'm well over 50 (but under 60).
That's without a degree.
So like, maybe proven track records do help (there's a -reason- I have a 3 page resume, folks. fuck that 1 page shit).
Short answer:
yes
And to add to this, folks doing bioinformatic and genetic research will -=really=- put the Opteron to use, and for a -hell- of a lot less than Itaniu (and I2).
:P
And no, Engineers -aren't- the only ones that could use them on their desktops. You also forgot graphics designers and testbed machines for applications programmers, and testbed machines for a larger 64bit environment.
And you're getting this pricing for the Opteron from where? Everything I've been hearing has the low end stuff -=undercutting=- the current Xeon pricing, while still running more than standard Athlons/P4's.
Geez.
*nodnod*
AMEN to this.
There's a reason when you look in a "pro" kitchen -everything- is gas powered. If it's not, find a different restaurant to eat at. Gas is -always- better than electric.
Period.
And yes, a bit of experience cooking on gas will teach you how to gauge flame height/intensity and temperature. One thing I -do- want to note, though: try not to let the flame lick up the sides of your pots and pans. Seriously.
*nodnod*
How important was that cluster to computational research? How important is that cluster to the mission of the school? I'm not asking facetious questions here, I'm honestly curious.
The machines that I work with support millions of dollars worth of research projects. Having unskilled, or even semi-skilled students futzing around as root on theses systems could be potentially disastrous (even -with- me backing everything up nightly).
We were employed by the university
This is -exactly- what I -don't- mind. Students who's job and responsibility it is to admin a machine are fine by me. Well, with one caveat: there's some sort of oversight from a more senior admin (be he/she another student, or a pro).
I'm not against students gaining experience, and being given responsibility of important/semi-important machines. What I'm against are students who, just because they know something about computers, are given root on the university's central servers to do with as they will without oversight.
To Paraphrase: if you're paid job as a student employee of a college or university is to admin a couple of machines, cool. There's -going- to be some oversight (at least there should be) as well as consequences for actions taken. You've also been through some sort of an interview process to (hopefully!) only put someone who's capable of doing the job in place.
And what is the machine being used for?
If the -job- of that computer is to teach the students systems administration, then more power to 'em.
If it's a system out in a student computing lab, or a system used primarily for the programming classes or whatever, then no, students should -not- have access. There's absolutely no reason.