| |
| |
| |
|
Page: 1 2 3 4 5 6 7 8 9
Comments:
<0> And still sound good
<1> My schedulser is nearly roughed out
<1> I'm now getting hostname and load avwerage from remote hosts and I'm building the data parser now.
<1> I'm an optomist.
<1> it will work with up to 1024 nodes
<0> cool
<0> whats it do
<1> well, it will swcan a cluster of available machines and send a job to the one with tle lowest 1 min load average
<1> for example it is called "launch"
<1> and "launch foo <args> " woill launch program "foo" on the most available machine
<0> thats awesome
<1> you see, my cluster runs on a single system image
<0> :d slick
<0> couldn't you then make it automatic, so you had the power of 1024 pentium Is at your hands
<1> and all nodes share the same /usr and /home
<2> I think you should offer shell time on your cluster
<2> :D
<0> yeahhh :P lol
<0> make it a dedicated *something cool* environment
<1> lol, it is not a very fast cluster
<0> oh
<1> but it is very parallel
<0> well fasten it
<1> got $$$
<1> ?
<2> I'll be on my way to a little cluster of my own soon
<1> g'night
<1> good
<1> we can grid
<2> yay :D
<1> all my nodes have internet access
<2> It'll be nice, I'm getting some older Xeons from my uncle, as his workplace is scrapping their older servers
<1> cool
<2> I think I'll be able to get all of mine on the net too
<1> well, I was building an opteron cluster, but the heat the first node made changed my mind
<2> hahaha
<1> so I wntr back to my mini itx boards
<2> nice :)
<1> mostly because wile fast is good, I am more into the parallelism
<1> and I can run a roomfull of mini itx bards for what a few dual opteron bioxes umake in ehat
<2> Yeah, this'll be a nice learning experience for me in parallel development
<3> it also seems like a good way to prototype the software, and then possibly scale it up to larger systems afterwards
<1> well, parallel is where it is at
<1> intel hit the wall. Moore's law is dead
<1> so the only go-fast is more cpus
<3> moore's law is not dead, it's simply changing
<3> since moore's original law only stated that cost would go down and number of transistors would go up by whatever scale every 18 months
<1> well if they change the definition of it, howe can it still be Moore's law ?
<4> note that it is not known whether inherently serial problems exist.
<1> well, indeterminant problems exist
<4> so multiple processors need not give you any speedup at all.
<3> if the transistors are running in parallel instead of going faster, it seems to me that the law still holds
<1> and parallel machines don't do so wel lon them
<4> "indeterminant"?
<1> correct
<1> for some things parallelism does no good
<4> tmyklebu@beta:~$ oed indeterminant
<4> tmyklebu@beta:~$
<4> this also has not been proven.
<1> indeterminant means no known outcome
<4> no, that's 'indeterminate'.
<1> the big poroblem is the coarse grained nature of a beowulf
<1> you can't know the state of any cpu
<1> if it becomes impossible to know when to communicate your process may fail
<1> -or hang for eternity...
<1> in any case, there are a lot of things where parallelism helps
<1> plus is is just really neat. :p
<5> parallel is not the solution
<5> think how many 1980 cpus we'd have to put in parallel to equal your pc
<5> so in 2020, we wont be putting lots of 2006 cpus in parallel
<3> 1980? standard desktop processor would be a Z80 or an 8086 in a high end machine
<1> well, IBM used 32,000 pda cpus in blue gene
<5> the quest for higher serial speed will not end
<1> Z-80 was the big dog back then
<1> 8086 was slow
<5> people always say 'this is the end, this is as fast as they're going to make'
<5> i say bull****, we'll see faster, and it wont be due to parallel
<2> the z-80 is a beautiful processor
<1> well, the physics of silicon is pretty much a solid wall.
<1> yes it is.
<5> note i didn't mention silicon ;)
<1> I think it is still in use in industrial and space applications
<2> And in calculators :D
<1> they are doping silicon with germanium in the labs now and they are getitng below the 0.6V threshold
<1> plus it allows for higher temperatures
<5> high temp is bad :P
<6> why bother you have diamons
<1> but they don't know how to fabricate it yet
<6> diamonds
<4> computers will get faster and less reliable.
<1> germanium semiconductors can take a lot more heat than silicon
<1> but it is more expensive
<5> evilgeek: computing of the future will have to take the google approach
<1> that is back to parallelisin
<4> what, just hire lots of people and hope that **** starts working again?
<5> parallelism for reliability
<7> evilgeek: worked for Amazon.
<1> that is called high availability clustering
<5> more so than speed
<4> that doesn't necessarily work. you need to read about complexity theory.
<5> raid5 for cpus
<1> oh yah there are all kinds of failoversystems
<1> none of them work, really.
<4> not all errors can be corrected in this way.
<1> it is back to the indeterminant nature of the state of an outside cpu.
<4> we've been over this. "indeterminant" is not a word.
<1> I wish I could show you here
<1> I can ask for the hostname of all my cpus at the same time
<1> and they never come back in the same order
<1> it is gaussian
<7> evilgeek: dictionary.com disagrees. http://dictionary.reference.com/browse/indeterminate
<8> _Aegis_: It is a fact, not something that is being argued.
<1> the link is stalled
<1> their server is crap
<4> Akashra: that's a different word.
<1> indeterminant
<1> if I spelled it correctly...
<7> ah true :)
<1> in any case, you can't know the outcome, because you can't know the state of another cpu precisely
<4> why are you using random words whose meanings you do not understand?
<4> "gaussian" refers to a very specific kind of probability distribution.
<1> yes it does
<1> and an identical parallel operation across many nodes, started at ther same moment, will return in a different order every time
<1> and the statitical reperesentation of it is a gaussian distribution
<1> a bell shaped curve
<4> that is impossible, since there are a finite number of possible orderings of a finite number of nodes.
<1> across identical nodes ?
<4> yes.
<1> I can demonstrate it to you
<4> no, dude, you can't. you're claiming a mathematical impossibility.
<1> I'll have to paste to a pste bin
<1> not impossible.
<4> if you have n nodes, run your process n!+1 times and you are guaranteed to have a duplication.
<4> it is impossible, dude.
<4> you are rejecting the pigeonhole principle.
<1> it is a fundamental characteristic of a distribuited memory supercomputer
<4> this is a fundamental property of a finite set.
<1> lol, well build one and find out. :p
<4> would you care to explain how you can run the same process arbitrarily many times without it ever outputting the same permutation of {1, 2, ..., n} twice?
<1> I swend it in parallel
<1> the states of the cpus are never identical
<4> okay, so let's say you have two CPUs, CPU 1 and CPU 2.
<8> _Aegis_: No dude, you're wrong.
<1> so the response time for each machine will be differenmt each time you do it
<4> there are two possible ways to order these, not infinitely many.
<8> Order is finite if the set is finite. Period. Think about it.
<1> yah and that finite set has a gaussian distribution in it's probability of ordering
<8> What you're saying is; "There is an infinite number of ways in the order to put 3 marbles into 2 buckets"
<4> either CPU 1 says "hi" first, or CPU 2 does. if one says "hi" first, the other one says "hi" second.
<1> the randomness is not in the number of nodes
Return to
#c
or
Go to some related
logs:
#nhl
routing-and-remote-access seperate ports
#hardware
#firebird
#beginner
#beginner
bioremidation
#firebird
amoroso39
#cph
|
|