View Full Version : Taking half the farm down
I'm thinking about taking the slower half of my machine down for two reasons. First is to make room and give me an excuse to build faster machines. Second is to learn how to build a mosix cluster with some friends. With his cluster of 7 p100 machines we were able to complete a wu in 7mins.
http://setiathome.ssl.berkeley.edu/fcgi-bin/fcgi?email=greg@greg.gann.org&cmd=user_stats_new
It's okay buddy, no need to justify your addiction to us :D Now I can't let you off the hook without lots of details on this cluster... How is it breaking up the units? Did you have to tamper with the fft routine in the code? If not that would be awesome since it doesn't violate the "patched client" warning from Berkeley.
craigiz1
06-29-02, 04:22 PM
Originally posted by LuKE
I'm thinking about taking the slower half of my machine down for two reasons. First is to make room and give me an excuse to build faster machines. Second is to learn how to build a mosix cluster with some friends. With his cluster of 7 p100 machines we were able to complete a wu in 7mins.
http://setiathome.ssl.berkeley.edu/fcgi-bin/fcgi?email=greg@greg.gann.org&cmd=user_stats_new
Can you explain this 'cluster' that you are talking about. 7 min. seems incredible with only 7 machines. I want to do that.
What does it do, split the w/u and then put them together?
rogerdugans
06-29-02, 04:41 PM
No excuses needed...but lets hear more about this cluster and crunching!
How about trying the benchmark wu?
Morph could probably tell you something about this concept too. Basically you need some software to split the work unit data up and distribute it to each machine and then append the returned data in the proper sequence. A friend of mine actually worked on a cluster at GA Tech, but they used 20 systems running linux and they had to break into the seti code, so ultimately Berkeley would not accept the results. That's why it has been stated that Benz could be doing something similar if his work unit times are actually that low, but Berkeley will not accept results from altered clients, usually refered to as patched clients.
craigiz1
06-29-02, 05:25 PM
Originally posted by TC
Morph could probably tell you something about this concept too. Basically you need some software to split the work unit data up and distribute it to each machine and then append the returned data in the proper sequence. A friend of mine actually worked on a cluster at GA Tech, but they used 20 systems running linux and they had to break into the seti code, so ultimately Berkeley would not accept the results. That's why it has been stated that Benz could be doing something similar if his work unit times are actually that low, but Berkeley will not accept results from altered clients, usually refered to as patched clients.
OK. Thanks TC. I'll just keep cruncing away.
Morpheus
06-29-02, 05:28 PM
TC is right...
I did A LOT of research into this (Beowulf cluster though... search Beowulf+Stanford if interested), and collected MANY parts before I came up one of the issues TC mentions... that is, the intergrity of the code must be maintained...
Secondly, in a Beowulf, the "network" device is the true cost... and Grendel (the software) is still evolving...
At one point, I had all the hardware for more than 14 P-166s w/ 64Mb of RAM... I could never find/afford the "fast" switch to make it all happen... (the get an idea, Standford used a very spendy switch donated by AT&T on the first Beowulf, and it was NOT fast enough)....
These are the issues that helped me "get a grip" and abandon the project... Luke sound like he may be onto something though... I NEED TO KNOW MORE!!! :eek: :rolleyes: :eek: :rolleyes: :eh?:
Elaborate for us LuKe..
What are we going to do tomorrow night Brain? The samething we do everynight, try and take over the world!
http://www.mosix.org/
This doesn't sound illegal to me. The OS divides the workload of the application. The client isn't hacked.
They're also using fast ethernet. Ethernet switches are cheap.
Originally posted by Greg M
http://www.mosix.org/
This doesn't sound illegal to me. The OS divides the workload of the application. The client isn't hacked.
They're also using fast ethernet. Ethernet switches are cheap. Yes indeed. I have 5 linux nodes here that I may try it on this evening if time permits.
craigiz1
06-29-02, 07:28 PM
Originally posted by TC
Yes indeed. I have 5 linux nodes here that I may try it on this evening if time permits.
TC-
Please keep us posted. If it works, can we find out if SETI will except this? It could mean a lot to our team.
You know what ****es me off, that completing 1 wu puts a person above 34% of the total SETIers. (click the link in LuKE's post and you'll see it) I know some of those 0 wu'ers are new members who just haven't completed a unit yet but most are people who've signed up and then done nothing.
rogerdugans
06-29-02, 08:33 PM
7 p100s completing a wu in 7 minutes is WILD: I have a work pc here for the weekend (had to replace a bad VLB ide card- lol!) and the thing is a p90. I tossed Seti on it just to see: after 26 hours its about 17% complete!!!!!!!!!
Luke- or TC: please do a benchmark wu for a comparison with LandShark- if results are valid......
I can get hold of at least a dozen sub 200mhz pcs within a week!
Talk about a Farm!:D
(Side note: with the low power needs of early Pentium-class chips, do you think 1 modern psu could power a couple of 'em?)
Add me to the list of those interested in this cluster setup! If it works and is legal then what an awesome project to work on . :D
Cy
I'm going to take a look in a bit. All of my systems here are red hat 7.3. The site mentioned a version all ready to go on slack 8 which is what we have on the servers at work. I can't risk taking the servers down, but if red hat doesn't work I'll snag the slack cd from the office and load it here.
Don't try this at home kids - I hosed my red hat install. Gotta start over from scratch.
heezer7
06-30-02, 12:17 AM
wow this sounds really neat. i have 4 or so k6-2 and mobos lying around. alone they take for ever to process a WU bout together it sounds like it might be an idea. keep us updated.
heezer7
06-30-02, 12:33 AM
ok this is sounding better as i read more. so it sounds like i could link my dual Athlon, and/or just three k6-2 550. is this easy to set up? they don't really say much on their site. i am new to linux. i have used it a few times but never really got that experienced at it. tell us how it goes. i am thinking this is very cool.
This looks possible even for the slightly newbie linux people. I did screw up a red hat box last night, but I think I know why. The installation program wanted a path to the kernel source files, and I couldn't find them so I tried to skip that section and move on. When I got done I couldn't shutdown or reboot my system.
Why do some work units take very little time to complete?
Why did the progress bar suddenly jump to 100% only part way through the analysis?
Occasionally, a work unit will contain strong radio interference; these strong interfering signals typically come from satellites and radar from our own civilization. If the interference is very strong, the SETI@home program can not analyze that part of the spectrum, and after trying for a few minutes and detecting thousands of strong signals of earth origin, the program stops early in the processing and gets a new work unit. You will still get credit for the work done.
Were thinking that's what happened but were still checking this out. I'm selling him like 10 more pentium class machines and a larger hub so maybe we can get this puppy running quicker. Thanks for the interest guys!
rogerdugans
06-30-02, 06:39 PM
Hey, my interest is self-serving Luke: I can get SLOW pcs easy:D
Thank you for posting about what you were trying: if the project works well, I think a few of us are going to be digging out the old junkers again!
I need some help with this. I'm not good enough with linux to get this running. I managed to recompile the kernel with the mosix patch applied, but things didn't go too well after that. I'd like to try it out, but I'm going to need some coaching.
Digital Pimp
06-30-02, 07:06 PM
TC needs an obi wan
but really, good luck tryin to figure this out, we could jump up the rankings in quite a hurry.
Digital
But would breaking a wu into chunks provide any total benefit? Sure it'll lower your times per unit, but if you take a wu that'll crunch in 100 hours on a P1 100mhz, and split it into 5 equal chunks then each of those chunks should take 20 hours right? If so then in the end the volume of crunching you get done in a day, week or month would be the same right?
yeah were pretty sure that was just a freak incident. My oh my what 7min wu times could do to a users ranking. Since the clusters uses smp it's still like having 7 p100s. I'm not giving up tho.
LandShark
06-30-02, 11:49 PM
"subscribe" me into the "List of interested" people list pls!!! ;)
if it really works out as thought, i could imagine LOTS of farmer will born!!! :D
hmmm....."only" 7 P100 huh? what IF i link my own farm together...??? :rolleyes:
Just shoot
07-01-02, 11:36 AM
Please keep me posted as well, I have 10 p166 at my disposal in approx a month.
What would really be slick would be to tie everyones slow machines together and crunch under a account kind of like amdmb The pond.
Antimatter
07-01-02, 11:49 AM
Originally posted by LuKE
Why do some work units take very little time to complete?
Why did the progress bar suddenly jump to 100% only part way through the analysis?
I remember reading somewhere if seti finds enough spikes, pulses etc, it will just finish, even if it's not at 100%.
You have to realize that 7 minutes is not the norm even for a cluster, but a cluster could potentially do work units in maybe 45 minutes to an hour. The 7 minute one was clearly because of a bad work unit.
Just shoot
07-01-02, 02:37 PM
Well guys I'm going to jump in. I'm a newbie to linux but this looks pretty straight forward, YEAH RIGHT, LOL.
Anyway I've got 3 p166 and 1 1gig duron (morgan) I'm short a harddrive at the moment but will have it this weekend.
I've spent most of the afternoon reading the Mosix site and the K12os.org How to, lots of info here if anyones interested.
At the bottom of the how to guide it says that they use seti for cluster testing, so its got to be legal no hacked client.
Keep me posted, maybe we can figure this out. I'd like to link two dual XP machines via gig ether and see how that does. I downloaded the mosix patch et al, but the readme had different instructions from what I've seen on some of the how to's like this one for instance:
http://www.tldp.org/HOWTO/Mosix-HOWTO/index.html
You may also want to have a look at this:
http://psoftware.org/clumpos/#START
Just shoot
07-01-02, 05:44 PM
I looked at those as well TC, but they didn't make a whole lot of sense being alittle different.This (http://k12os.org/modules.php?op=modload&name=News&file=article&sid=14&mode=thread&order=0&thold=0) is what I was going to try it seems pretty straight forward, heck if a high school networking class can do it we should be able to. Anyway I'll keep intouch with my progress. Later
meb3269A
07-01-02, 05:50 PM
luKE
If you need any info building that cluster give me a shout. I just constructed a 32 node setup at school and wrote a paper on it. We used a Gigabit ethernet configuration as opposed to a myrinet/fiber setup which was too expensive. I am now working on a cluster at home. Still in the planning and node collecting phase. I am also working on the software to break up the work units. Do you have any info about Berkelys rules pertaining to how the client is used?
Good luck
meb3269A
Originally posted by meb3269A
luKE
If you need any info building that cluster give me a shout. I just constructed a 32 node setup at school and wrote a paper on it. We used a Gigabit ethernet configuration as opposed to a myrinet/fiber setup which was too expensive. I am now working on a cluster at home. Still in the planning and node collecting phase. I am also working on the software to break up the work units. Do you have any info about Berkelys rules pertaining to how the client is used?
Good luck
meb3269A Keep us informed on your project. I am working on it as well. I was thinking of hooking up 3 dual XP machines on their own switch with gig ether. I downloaded the clump OS and it works extremely well - just need to get a manually configured box up to get the cluster going. That's where I'm getting into trouble. As far as the Seti client is concerned - it cannot be altered in any way whatsoever to be legit. So if you have to decompile it and do something then there's no point in bothering with this.
Just shoot
07-01-02, 06:00 PM
meb3269A, by chance could you post a link to the paper you wrote or uplink a copy? I'm sure that others would also appreciate it, its got to cut a ton of questions and hours of digging for answers. Thanks Just shoot
meb3269A
07-01-02, 11:27 PM
Your right about the many hours of searching and digging for answers! Our project group of five members spent an entire semester working on that project and we barely scratched the surface. The paper is concentrated mainly on the high speed network connections, lightly touching on the operating system. We chose to use a Windoze platform against my better judgement at the recommendation of our professor. It turned out better than I thought though. We used Win 2K advanced server for the head node and Win 2K pro for the workstation nodes. I will post a link to the paper after I receive permission from the project team members to do so. Just a formality and a little professional courtesy. I can give you a good link to get started try, www.microsoft.com/windows2000/hpc/. Bill and his cohorts actually have a Windows based cluster starter kit for an extremely modest price.
meb3269A
LandShark
07-02-02, 01:08 AM
but the only thing that concern me & all of us is could we do it w/ SETI??? will it still be legal, or not???
do u guys know anyone that did that w/ SETI and berkely said ok??
i wanna check this out before spending too much time on it first......:rolleyes:
I understand learning about this as a growth / development thing but I wouldn't suggest using this for SETI long term. If a group of 10 machines were used individualy and one of those ten was giving faulty results then 10% of the results would be bad, but if those 10 machines were linked then that one machine would be corrupting part of every unit, resulting in 100% bad units.
Just shoot
07-02-02, 05:54 AM
TC here is a e-mail from the gentleman that wrote the k12os.org how to, interesting to say the least.
QUOTE
Hello Everett,
Due to Mosix going closed source I have moved to OpenMosix. The good
news that I have created an install package for OpenMosix. Go to the
"slobian package" at http://www.campworld.net/index.php/Linux.link
This is a prelease package and only works with i686 CPUs (ppro or
greater). Setup your K12LTSP server, install my package, and reboot.
Your diskless nodes can either boot from floppy or PXE. I would like to
get some more feedback on the package.
Seti doesn't use threads or subprocesses to divide up the workunit.
Hence on a SMP machine you should run two copies for a dual cpu machine.
Since this is the case I let the cluster work on multiple wus at once.
Create a shell script to run multiple copies of seti.
Important things to keep in mind:
1. All nodes should be around the same speed. Example, lets say you
have a 3 node cluster. Server is a 2Ghz cpu and the 2 nodes are 1Ghz.
You would run 3 copies of seti. But you find a node isn't doing any
work. Since the server is 2x faster than the nodes 2 copies of seti
will run on the server before Omosix migrates any processes. So you
would have to run 4 copies to keep the cluster busy.
2. A 100Mbit switch is required. When seti migrates 15Meg of data goes
with it.
3. Open Mosix and Seti will push your hardware to the edge. If there
is a weak link this combo will find it. This will test to see if a CPU
can be overclocked.
My current cluster is a dual PII 400 server and 4 400Mhz celeron nodes.
My nodes have no monitors or keyboards. Everything I run is done from
the server. With Omosix you don't need to log into a node to runn a
program. To seti or other program the cluster looks like on large
computer with multiple CPUs.
But its all about the numbers. The cluster runs a 16-18 hour per wu.
The cluster crunches 6 wus at a time. The CPU speeds add up to be
2.4Ghz. So the cluster turns in a wu every 2.6 to 3 hours. My 1.6Ghz
P4 with RDRAM turns in a wu every 5-6 hours.
Also I am looking for a source to buy celeron cpus between 400Mhz to
600Mhz. I need 5 to 20 cpus. Prices are too high on ebay.
Good luck
Richard Camp
QUOTE
I think it still worth a try, if anything to take advantage of alll those slower chips laying around.
craigiz1
07-02-02, 10:26 AM
Originally posted by Just shoot
QUOTE
Hello Everett,
But its all about the numbers. The cluster runs a 16-18 hour per wu.
The cluster crunches 6 wus at a time. The CPU speeds add up to be
2.4Ghz. So the cluster turns in a wu every 2.6 to 3 hours. My 1.6Ghz
P4 with RDRAM turns in a wu every 5-6 hours.
Good luck
Richard Camp
QUOTE
I'm not sure if I see any gain to this. If he can get a w/u from his cluster in 2.6-3 hours at 2.4Ghz and then he get's a w/u in 5-6 hours with his 1.6Ghz P4, it seems you would get about the same timing as the cluster with a P4@2.83Ghz with less expense and hassle. Maybe I'm not getting the point.
Please tell me if I'm missing something, because this idea is very interesting.
Thanks and good luck.
I would think that a big part of this would be just building the mosix (Omosix) cluster itself. The big question I have is if that cluster can churn out w/us on average of one every 2.5hrs etc. how does it register the times with your account? Meaning, is your account still credited with 18hr w/us or is it credited with w/us at 2.5hrs each? That would be something to consider!
I have two dual PII400 Xeons that I would not mind playing with as they are running RH already. I also have two duron 1.3s at 1.404 in the same rack I could use if it was worth while.
Cy
craigiz1
07-02-02, 12:11 PM
Originally posted by craigiz1
I'm not sure if I see any gain to this. If he can get a w/u from his cluster in 2.6-3 hours at 2.4Ghz and then he get's a w/u in 5-6 hours with his 1.6Ghz P4, it seems you would get about the same timing as the cluster with a P4@2.83Ghz with less expense and hassle. Maybe I'm not getting the point.
Please tell me if I'm missing something, because this idea is very interesting.
Thanks and good luck.
Bump.
Just shoot
07-02-02, 12:50 PM
Originally posted by craigiz1
I'm not sure if I see any gain to this. If he can get a w/u from his cluster in 2.6-3 hours at 2.4Ghz and then he get's a w/u in 5-6 hours with his 1.6Ghz P4, it seems you would get about the same timing as the cluster with a P4@2.83Ghz with less expense and hassle. Maybe I'm not getting the point.
Please tell me if I'm missing something, because this idea is very interesting.
Thanks and good luck.
I can only think of one benefit other than just being neat to try, and that would be to take advantage of all the old hardware laying around.
I'm not really seeing the advantage in this. This guy is running a total of 6 CPUs in the cluster, running 6 clients and averaging roughly 6 results per 18 hours of crunch time. This is relatively a result every 3 hours but this is moot...it is still taking 18 hours of completion time to complete one WU. You can't compare that to a single P4 system running only one client...it's just not a logical comparison.
Without the cluster and running a client on each individual machine per processor, he is going to average the same thing. Each rig will still produce a WU in about 18 hours meaning a total of 6 WUs will be returned in an 18hr timeframe.
Is it just the "coolness" factor that makes this desirable or am I missing something terribly?
craigiz1
07-02-02, 01:01 PM
Originally posted by Jon
I'm not really seeing the advantage in this. This guy is running a total of 6 CPUs in the cluster, running 6 clients and averaging roughly 6 results per 18 hours of crunch time. This is relatively a result every 3 hours but this is moot...it is still taking 18 hours of completion time to complete one WU. You can't compare that to a single P4 system running only one client...it's just not a logical comparison.
Without the cluster and running a client on each individual machine per processor, he is going to average the same thing. Each rig will still produce a WU in about 18 hours meaning a total of 6 WUs will be returned in an 18hr timeframe.
Is it just the "coolness" factor that makes this desirable or am I missing something terribly?
This is my point too.
rogerdugans
07-02-02, 01:20 PM
Thats why I am interested: I expect results to show no real gain, but if, and I mean IF there is any decent gain, then the hassles of getting set up might be worth it.
I do not see how it could make any improvement, but I miss a lot sometimes:D
I can only see where a loss in performance could happen, unless you're using gigabit networking. Even that may not help if using very powerful machines for this.
In no way can a 100Mbps network connection substitute for memory bus bandwidth. On slower machines this won't make much difference and I would expect that this is only attractive for several slow machines anyway.
I have to agree with Jon's logic.
The cluster would run slower due to the overhead involved and the relatively slow communications links.
Morpheus
07-02-02, 02:58 PM
As I posted earlier, the beowulf is SEVERLY limted by the communication links (front plane/back plane)... Standford used a switch designed for fast communications traffic (which i would assume far exceeds gigabit-per-second networking) and was limited on clusters of DX-100 systems... How does Mosix get around this bottleneck?
Additionally, the overhead of the software needs to be considered... in a beowulf, there is a "control" node that compiles/decompiles and maintains order in the cluster... if Mosix does this resident on each system, there will be a performance loss... without a doubt...
Now after this doom & gloom.. will someone prove me wrong so that I can get further interested... :rolleyes: :D :cool: :rolleyes:
Unfortunately I have nothing good to add to this gloom and doom. Last night a friend of mine helped me get it running with two nodes - both dual XP systems. I started two seti clients on the first node, and nothing happened on the second. When I started a third seti client the second node showed activity on one cpu. When I started a fourth client on the first node the second cpu on the second node showed activity. Since Seti is not multi threaded to begin with, what Mosix did was move the third and fourth clients from the first node over the network to the second node. It was essentially no different than if I had manually installed the client on the second machine and started two instances of the client. Bottom line - complete waste of time for seti. If Berkeley ever did release a true multi-threaded client it would be a different story. With gig ether and a switch running in cut through mode it might produce some nice work unit times, but I suspect that they don't want to do this on purpose because it would create more bandwidth problems for their servers.
Kind of figured that. If you use this with single-threaded software all you're going to accomplish is a load balancing cluster.
Never hurts to try though and kudos for doing that :]
meb3269A
07-02-02, 07:07 PM
A properly configured and managed cluster will blow the doors off of a bunch of individual machines. To test our freshly built parallel setup we used some data-mining software that was written to utilize a parallel system. The setup crunched through a 50 gigabyte database in under 20 minutes on our trial run. If we had the time to fine tune that beast it would have been much faster. The key to fast sorting or calculations is the same principle as SETI uses by distributing the workload. Multiple work units being crunched at the same time is defeating the purpose and turns the cluster into just another LAN. The greatest challenge for us to crunch SETI in this manner is the client software. To follow the rules and guidlines set forth by Berkely, we can not alter the clients code. This is very understandable from Berkelys standpoint but hinders our effort. I plan to run a copy of the client on each machine and write some piggy back software to distribute individual chunks of each workunit to every machine and then recombine the WU for the final result.
wow, you guys are just amazing. This one thread has thrown my interest back into the team. From TC spending his time and money on this to Jon taking a step back and bringing everything back down to earth.
I really doubt Berkeley hasn't noticed the interest in cluster computing. What i do wish is that they would release something that could harness that power. Just think if dave from amdmb.com had been using a client of that sort. But, were always stuck saying what if what if at the end of the day. I guess that's what makes us like overclocking so much. Keep crunching guys and keep posting cause were still learning!!!!
MrMagoo
07-02-02, 11:24 PM
Hello everyone,
I'm the author of the K12LTSP+Mosix howto. My reasons for using a cluster are as follows: Cheap cost per node. The only software install to get a node running is making a floppy that boots the machine from the network.
Running a cluster is not easy. My goal is come up with a setup that is easy for most people to build and low cost. My setup isn't the best solution for some things.
The best part is that friends can bring over their machines and add them to the cluster. I just make a floppy disk for their machines boot.
I have been running seti on my cluster without any problems. I have not done anything to the client. As far as seti is concerned I am running multiple copies.
Unfortunately due to the high price of electricity in Calif, I can't run my cluster full time.
Richard
craigiz1
07-02-02, 11:28 PM
Originally posted by MrMagoo
Hello everyone,
I'm the author of the K12LTSP+Mosix howto. My reasons for using a cluster are as follows: Cheap cost per node. The only software install to get a node running is making a floppy that boots the machine from the network.
Running a cluster is not easy. My goal is come up with a setup that is easy for most people to build and low cost. My setup isn't the best solution for some things.
The best part is that friends can bring over their machines and add them to the cluster. I just make a floppy disk for their machines boot.
I have been running seti on my cluster without any problems. I have not done anything to the client. As far as seti is concerned I am running multiple copies.
Unfortunately due to the high price of electricity in Calif, I can't run my cluster full time.
Richard
WELCOMR TO THE FORUMS!
Hope you can stay and help with this idea. :)
Originally posted by MrMagoo
Hello everyone,
I'm the author of the K12LTSP+Mosix howto. My reasons for using a cluster are as follows: Cheap cost per node. The only software install to get a node running is making a floppy that boots the machine from the network.Richard - very nice to see you in the forums. I appreciate your email as well. What you're doing is exactly what is needed to bring clustering, and in a way, linux to the masses. Making it as fool proof and easy to setup as possible for those linux idiots like myself :rolleyes: I can get a basic linux box up and running, and even administer some programs once I know my way around, but as of yet I have not tried to compile kernels, and all these config files and scripts are duanting sometimes. If you can make this as easy as downloading the file and typing ./xxxx from a prompt then you'll have some very happy people. When I first ran the mosix patch utility it seemed like every question it asked (aside from the path to the kernel source) I was scratching my head thinking - what the heck is that! The clump os bootable cd is the easiest thing I've used so far. I brought several nodes up on my home lan that way, but I had no server to manage them. If you can make the server almost as easy to setup as that then you've got something truly awesome. :cool:
beau_safken
07-03-02, 12:13 AM
IF this 7 minute WU thing is the truth I can get together like 20 sub 200 PII's really quick so form a new thread if it works
Originally posted by meb3269A
A properly configured and managed cluster will blow the doors off of a bunch of individual machines.Not in all cases, as demonstrated by SETI in TC's experience above.
A parallel cluster is great for things such as DB queries or rendering where each system receives a request and then works with only the portion of data associated to that system. Each system takes a piece of the work and runs its processes in parallel to each other system working on its own piece of the puzzle. This is much faster than a single system attempting to parse the entire DB or render the entire model in serial. FWIW, this is the configuration often seen with large data warehouses.
There are other components to computing where single faster systems will wax a cluster. They each have their positives and negatives.
If the goal is to get 50 buckets of water from one point to another, it is often faster to use 50 slower runners to get them there than to have a sprinter carrying one bucket at a time and making 50 trips. The 50 runners are working in parallel, each doing their part concurrently to reach the final goal. The sprinter is working in a serial fashion. If the goal is to get a single bucket there the fastest, the sprinter wins.
With the serial nature of crunching WU's with SETI@Home approved clients, a faster machine can do more work, and consume much less power, than a bunch of slow machines.
At the office, it takes almost 4 dual PIII 600MHz servers to keep up with a single dual P4 Xeon DP 2.2GHz system (using 4 logical processors).
The great part about this for Seti, which Richard pointed out to me, is that you can build a large diskless seti farm. Up to this point I have been building complete systems and taking the time to install and configure the OS, etc. With his diskless cluster you can build a ton of simple systems with nothing but the board, cpu, ram, and a floppy. The system will boot from the floppy and then load the necessary files across the network.
Just shoot
07-03-02, 01:39 PM
First I would like to thank MrMagoo (Richard) for responding to my emails and for taking the time to post in the forums, its always nice to see another member.
Maybe this is a stupid question but, doesn't seti only use 15 mb of information to crunch the WU. So with a 100mb switch in full duplex thats 200mb isn't the information that would be required to run seti such a small amount of the total network that it becomes insignificant? Maybe I'm way off base on this.
Very good point TC, in diskless mode, the cost of a PC or cruncher If you will, becomes very small. Pricewatch has Combos available for $72, thats a 1 gig duron, ECS K7AMA board with video, sound and 10/100 nic card ( video and sound not needed but would be nice for reselling ) total cost with power supply and floppy drive would be under $100. keep in mind this is bare bones stuff and nothing fancy. Memory would depend on what a individual would want to spend $34 dollars for ddr or $28 for sdram these are 256 meg sticks. This would be a cruncher anyway you look at it for under $150. If five of these were bought that would be 5 GHZ computer for under $750. I would have to say it would be pretty hard to build two systems for $750 to equal the same amount of output.
Everyone should be reminded to keep a open mind myself included Sometimes just learning how to do something is enough of a reason to try. Has anyone had a conversation with someone about overclocking and they ask why wouldn't you just buy what you want to begin with and not overclock, OMG.
We should keep it in mind that sometimes we do things for the heck of it just to say we did it.
Anyway I think I rambled on long enough, TC if you don't mind I would like to pick you brain on how you set up your cluster, I know it would save me sometime, and it would be greatly appreciated.
P.S. MrMagoo, I couldn't help in noticing that you don't crunch for a team at the moment, I'm sure I speak for everyone when I say we would be happy to have you as part of our team.
I'm not the right person to talk to about setting up the cluster. I tried in vain to get it going according to the mosix how to guide I posted a link to. I have an acquaintance whom is very good with linux, and he just zipped right through the instructions on the server node, and then it was easy to boot the clump os cd on the other node.
vBulletin® v3.8.7, Copyright ©2000-2012, vBulletin Solutions, Inc.