Slashdot Mirror
The Transmeta Pushme-Pullyou?
tired.cranky writes: "An article on LinuxDevices.com sez that Transmeta is about to ship a quasi-distro slash embedded development toolkit featuring Linus' new super-efficient cramfs and ramfs filesystems. Apparently, a reasonably normal Linux system can be shoehorned into 8MB of storage, with zlib decompression-on-demand and such. It sounds like it could push a fair few hobbyists and embedded developers in Transmeta's general direction, too... and reads nicely next to a Register piece on Transmeta's leaked server initiative. Does one end of Transmeta know where the other is pointed?"
It might not be a big deal now, but how about in 6 months or so when the real cost of electricity is passed on to the consumers? If you started paying 10x more for your electricity than you're paying now, will power savings be an issue?
-jon
Remember Amalek.
The low power, many microprocessor solution has been propounded many times, and it has routinely failed in the marketplace. Why?
People don't care about low power on servers. Even if the power costs and cooling costs are trippled, only a handful are in a position where optimizing for power, not speed is important.
Theoretical peak, N cheap processors is much more powerful than one expensive processor
Then why are sales of low-power servers skyrocketing, especially in San Francisco, which someone from Berkeley should know about. Hint: read the local business pages sometime. I do up here in Seattle, and it's reported here.
There are high-demand servers, mid-demand servers, and back-up servers. For high-demand, constant traffic servers your arguments make a lot of sense. For mid-demand or failsafe rollover servers, the heat cost and power cost are a higher fraction of the operating expense of the server, which is likely a cheaper box with dual drives but chock full o RAM.
Servers don't come in one size or one shape. Some people serve up lots of static pages, some people have worldwide operations. Others have frequent low-traffic periods and may have a primary hit box with some image servers that can afford to have a little lag time and would get a high return on code-replication as it serves the same file over and over and over and over and over.
Just because you don't like Dietzel doesn't make you right. Think before you post, do some research, a few base google queries, check the local trade mags, before leaping to assumptions.
--- Will in Seattle - What are you doing to fight the War?
And we're supposed to be impressed? Let me break the news here. QNX did it ages ago when they squeezed their kernel, GUI, web browser, mail client and tcp/ip stack into a SINGLE 1.44 MB floppy. Oh, the power of Closed Source!
Your pizza just the way you ought to have it.
The low power, many microprocessor solution has been propounded many times, and it has routinely failed in the marketplace. Why?
Similarly, non-SMP machines have repeatedly failed when compared to SMP machines, because a message passing machine is much more difficult to program. Yes, if you can rephrase your programs to run on a cluster, then you see the impressive possibility in cost/performance, but it is really difficult to program well. People pay the extra cost for SMP machines simply because they can actually program them for a wider variety of applications.
Most problems which can be translated to a message passing structure have already been migrated to clusters of cheap machines. It is problems like databases, which don't map well to message passing, which is why people buy "servers", and it is these problems which are why people buy E10ks.
Dietzel knows better, he is an intelligent scientist, but he is acting like some corporate PR flack. Such dishonesty I can accept from someone who is ignorant, but he knows better.
Nicholas C Weaver, Winged Rat Consulting
nweaver@cs.berkeley.edu
Test your net with Netalyzr
Let us assume a 1u high rackmount mini-server, which is burning 100W of power [1], and which will have a 2 year lifetime before you call it obsolete and scrap it because the space you are renting would be better served by new machines.
Now let us assume a ridiculous power cost of $.50/kwh (note, current rates are around $.10/kwh, and may rise as high as $.15 when all is said and done. Even if you are being charged for cooling as part of the power budget, 50 cents a kilowatt hour is a ridiculous charge).
This means the server costs $.05/hr to operate, or about $900 in lifetime power costs over the 2 year lifetime, assuming vastly overpriced power.
Now given a miracle transmeta server which burns only 50W (after all, you still have to power the ethernet, disks, memory, etc). EVEN IF it has identical performance, it can only cost $450 more for it to be cost effective. If it is lower performance along with lower power (which it will be), there is no point in purchasing it.
[1] Current dual processor, stacked, HIGH end (2x933 MHz PIII) 1u systems will burn ~200W. Couple that with a still excessive but slightly more reasonable power cost of $.25/kwh and the calucation is roughly the same: a lifetime power cost of $900. Something like the BriQ is only burning 40W, but being sold for the remarkably low form factor (can probably fit 4 in a 1U rackmount), not ops/W.
Nicholas C Weaver, Winged Rat Consulting
nweaver@cs.berkeley.edu
Test your net with Netalyzr
Transmeta is about to ship a quasi-distro slash embedded development toolkit
And how does one develop under such a system? Post your programs and hope they get modded up to compile-threshold level?
DataSquid.net, a little about me.