This is just interesting for scientists, who want to transport enormous amounts of data. TCP has a performance problem on networks with a high bandwidth-delay product (so-called "long fat pipes"), where a typical TCP connection will only fill less then 10% of the available bandwdith. At iGrid2002 , scientists were schocked to see their great demos fail miserably by just adding 100 ms of delay in their network.
One of the reasons is that TCP has a congestion control mechanism (it will back off if the sender doesn't get all acknowledgement back), but with such a high bandwidth-delay product, the congestion control mechanism is triggered, even when the routers are not dropping any traffic. It is currently unclear what the exact reason for this is (I've heard possible reasons from the sender floading the receiver to the random early packet dropping that most modern routers do). Scientists are looking into this right now. Problem is the sheer complexity of TCP Reno (The TCP implementation as you are using right now).
Of course, UDP can be used: that is much more aggresive then TCP, and can fill 98% of the available bandwidth, but has no congestion control, and can easily bring down a whole university network to it's knees if applied improperly. So a whole range of new TCP-alternatives is emerging, FastTCP one of them.
Currently, the Global Grid Forum (GGF), a standardization organisation is evaluating all these TCP-alternatives. You can check their most current results at:
http://www.evl.uic.edu/eric/atp/
(Note that this is very much a work in progress!)
Re:mac problem
on
OS X Hacks
·
· Score: 0, Offtopic
M.A.C. = Media Access Control, a sublayer of the Datalink layer (layer 2) in the OSI protocol stack. It's main purpose is to take care that on a shared network, like ethernet, not two hosts start talking at the same time.
Apperently, the author is babbeling that anyone on an ethernet (M.A.C. Users) use old hardward (perhaps he would like everyone to use a 10Gb/s optical line?)
Other than that, I have no clue where he's talking about. Obviously it's off-topic since it is not Macintosh related:-)
I don't get it -- the numbers the author gave are perfectly in order to the thing which is often referred to as "Moore's law": 16 years of doubling hard disk size gives indeed 2^16 = 65,000, which is indeed the factor of 60,000 he gave. So: well his numbers are correct. So what's new here?
In fact, let's just restate what many probably have said thousants of times before: * Moore did not say anything about money or hard disks. All he observed was an exponential growth in the number of transistors per integrated circuit. http://www.intel.com/research/silicon/mooreslaw.ht m * In day to day talk, the term "Moore's law" has been (ab)used for three things: 1) The speed of CPU seems to double every 18 months (assuming it is a CPU of the same price, it is unclear if this is just the number of Gigaflops, or the clockspeed) 2) The speed of hard disk storage doubles every 12 months 3) The speed of network capacity doubles every 9 months
Regrettably, the writer just made a pretty common observation. Too bad. It would have been a much more interesting story if drew an interesting conclussions, like: despite the exponential growth, the CPU proceessing speed really is getting behind compared to network capacity, so in a matter of a decade we can expect a lot more distributed computing, just because it will be cheaper to buy more bandwidth than to buy more CPU's
I do not buy anything with rebates. I'm living in the Netherlands, and typically, I have to contact companies abroad (mostly USA), and that is just too much hazzle (if possibly at all).
Actually, debates are a clear indication for me that I will pay too much for a product, so it has in fact some use for me: they act as a big warning sticker: "Do not buy. Too expensive".
Slashdot Mirror
One of the reasons is that TCP has a congestion control mechanism (it will back off if the sender doesn't get all acknowledgement back), but with such a high bandwidth-delay product, the congestion control mechanism is triggered, even when the routers are not dropping any traffic. It is currently unclear what the exact reason for this is (I've heard possible reasons from the sender floading the receiver to the random early packet dropping that most modern routers do). Scientists are looking into this right now. Problem is the sheer complexity of TCP Reno (The TCP implementation as you are using right now).
Of course, UDP can be used: that is much more aggresive then TCP, and can fill 98% of the available bandwidth, but has no congestion control, and can easily bring down a whole university network to it's knees if applied improperly. So a whole range of new TCP-alternatives is emerging, FastTCP one of them.
Currently, the Global Grid Forum (GGF), a standardization organisation is evaluating all these TCP-alternatives. You can check their most current results at: http://www.evl.uic.edu/eric/atp/ (Note that this is very much a work in progress!)
M.A.C. = Media Access Control, a sublayer of the Datalink layer (layer 2) in the OSI protocol stack. It's main purpose is to take care that on a shared network, like ethernet, not two hosts start talking at the same time.
:-)
Apperently, the author is babbeling that anyone on an ethernet (M.A.C. Users) use old hardward (perhaps he would like everyone to use a 10Gb/s optical line?)
Other than that, I have no clue where he's talking about. Obviously it's off-topic since it is not Macintosh related
I don't get it -- the numbers the author gave are perfectly in order to the thing which is often referred to as "Moore's law":
t m
16 years of doubling hard disk size gives indeed 2^16 = 65,000, which is indeed the factor of 60,000 he gave. So: well his numbers are correct. So what's new here?
In fact, let's just restate what many probably have said thousants of times before:
* Moore did not say anything about money or hard disks. All he observed was an exponential growth in the number of transistors per integrated circuit. http://www.intel.com/research/silicon/mooreslaw.h
* In day to day talk, the term "Moore's law" has been (ab)used for three things:
1) The speed of CPU seems to double every 18 months (assuming it is a CPU of the same price, it is unclear if this is just the number of Gigaflops, or the clockspeed)
2) The speed of hard disk storage doubles every 12 months
3) The speed of network capacity doubles every 9 months
Regrettably, the writer just made a pretty common observation. Too bad. It would have been a much more interesting story if drew an interesting conclussions, like: despite the exponential growth, the CPU proceessing speed really is getting behind compared to network capacity, so in a matter of a decade we can expect a lot more distributed computing, just because it will be cheaper to buy more bandwidth than to buy more CPU's
I do not buy anything with rebates. I'm living in the Netherlands, and typically, I have to contact companies abroad (mostly USA), and that is just too much hazzle (if possibly at all). Actually, debates are a clear indication for me that I will pay too much for a product, so it has in fact some use for me: they act as a big warning sticker: "Do not buy. Too expensive".