Slashdot Mirror
Open Blade Servers?
Greg Smith points to this ZDNet story on new Intel chips aimed at blade servers, writing "Proprietary blade servers are coming on strong from IBM, Dell and HP. Where are the open blade servers? How did Google roll out 10,000 servers at such a low cost?"
I got an ad for AMD server solutions for that story..
WWJD.... for a Klondike bar?
Imagine a beowulf cluster of those...
Really, though, the fact that this blade server consumes so much less electricity would be very meaningful to me. The server room at our school was not intended to be a server room. The wiring also is lacking, and every once in a while the breakers go pop!
BTW, is it possible to use this in a laptop? Just imagine the power (or less consumption thereof) if you packed two processors in parallel on a laptop...
I'm the Devil the Windows users warned you about.
If you're too lazy to read the article and don't know what a blade server is...
Server blades got their name because of their design. A blade server typically resembles a circuit board more than anything else. They're made to be stacked vertically. These types of servers are growing in popularity for more mundane tasks such as delivering Web pages or housing protective firewalls because they use less floor space and electricity than racks of traditional servers. Server blades also share a power supply, cables and memory, which further cuts down on costs and space. Although the down server market has dampened sales, analysts believe blades will eventually form a substantial part of the market.
Maybe I'm retarded, but I didn't immediately picture exactly what a blade server was when I saw the name...so there it is.
A blade server is a hardware product, it really has nothing todo with software, outside of the Operating System Clustering/Scaling functionality.
Google does not use blade servers, last I knew it was just a large amount of x86 boxes running Linux.
Open Source hardware? Does that even make sense? Either have drivers (or release the specs) that allow your hardware to be used on an Open Source operating system, or dont.
Want an "Open Source Blade Server"? Yeah, thats called an HP with Clustered Linux on each blade...
The usage of Pentium IIIs for these monsters of serious computing only goes to show how much of a badly designed marketing ploy the Pentium IV is.
That could be dangerous. Over 50 people* die per month in server blade accidents. Sadly, these needless deaths could've been prevented by simple server blade cover kits.
Only YOU can prevent fatal server room mishaps.
Keep those blades covered, kids.
This is a public service message brought to you by Sally Shark, official mascot for server room safety.
*note: statistics may be fictional.
Well, I am sure free advertising has a lot do with the Google roll out.
The <insert powered by DELL,COMPAQ/YOMAMA) tags will start appearing all over google.
I also reckon they were free or at pretty much close to cost. Companies know what there doing. Cost on that kinda margin is probably at 200 bucks a pop straight outta the factory when you consider markup is about 500 percent on computer parts. Remember buying in bulk is power.
Example, you can get a pc on pricewatch with a 20 gig drive, 256 megs of ram, and a giga or more processor for, 250, so think about it.
Puto
The Revolution Will Not Be Televised
So a balde server is sort of like a pie but in reverse. Instead of making it smaller you make it bigger by adding some more stuff, but you still share the same pan.
"What we have here is a failure to communicate"
The Warden, Cool Hand Luke
Is it bad to when you see stuff like this to think how you can use it to further boost your Setiathome scores
..........FULL STOP.
Blade servers are not supposed to be stacked vertically, and you can fit *way* more than 42 blade servers in a single rack. The author is thinking of 1U boxes, which have only been around for say... 10 years!
i ndex-bl.html
look at : http://www.compaq.com/products/servers/platforms/
280+ servers in a rack.
The article is quite old now - March 19 - and HP appears to favour the blade servers from the former compaq. That being said the advantage that blade servers give is that they save a great deal of space, and make cabling much easier. In essence you can stuff a lot of proccessors in a rack, also put in a small disk farm, network switch using copper or fiber, and away you go.
Semper ubi sub ubi
The fastest AMD part listed in Pricewatch is the 2400+ (1.93 GHz) for $191. The equivalent Pentium 4 (2.4 GHz) is $188 (i.e. less expensive than AMD). Furthemore there are substantially faster Pentium parts available (up to 2.8 GHz now).
Yes, and RISC processors will outperform comparable speeds of Intel and AMD chips.....But anyways, Intel's are pipelined as well, goofball. AMD just seems to do it slightly better. However why Intel outperms AMD is because they are able to jack the clock speed up better than AMD seems to be able to. I'd bet they're purposefully decreasing the effectiveness of their pipeline to boost the clock speed. Anyways, the point of the article is that Intel is making Lower Power chips. So their main counterpart would be Transmeta, or say Motorola for PPCs, or SUN, et cetera, anyone of those are players in this particular arena.
Karma: Not Particularly Funny.
I'm currently involved in a server consolidation project where the customer has dictated that they want to see some blade. Our primary platforms are some kickin' Intel servers (8-way 1.6GHz, 8GB RAM, max 16-way 64GB) running VMWare ESX, but the customer is insisting on seeing some blade. I am personally unimpressed by them. You need to make sure that your apps can and are built to either cluster or failover cleanly when you get blade involved. Or just not run any mission critical stuff on it.
I prefer the VMWare ESX on our nearly-non-stop Intel hardware, the x440.
Intelligent Life on Earth
Ok, this statement confuses me. IBM's BladeCenter line is based on Intel's new standard for blade servers, which means that the blade rack should be able to accept blades from ANY manufacturer that follows the new Intel server blade standard.
HP/Compaq is also (supposedly) planning to use this standard for their new blade servers, so you'll be able to use HP blades in an IBM rack, and vice versa.
The only server blade company that seems to be sticking with a "proprietary" design is RLX technologies, which uses a more compact blade system that was originally designed to use Transmeta Crusoe processors. They also have Intel blades as well, which use a simular RLX proprietary form factor.
We won't see open blade servers for quite a while, if ever. Normal servers are only "open" because they use a common set of interconnects (standard power, ps/2 keyboard, 100BaseT), etc. On a blade server, you have to unify all of those interconnects in a hot-swappable fashion. The result? A customized connector and backplane architecture.
In addition, there's no incentive for companies to open a standard for blade servers - they'll make more money by selling the chassis and blades, as well as the management software that is generally required for these types of servers.
As far as Google goes, they rolled out their infrastructure for such a low cost because they did the following things:
1) didn't use blade servers(more on that in a sec)
2) bought in large quantities
3) bought generic/semi-generic servers (by which I mean "not IBM")
Not using blade servers was a sharp idea because the real advantages of blade servers come in certain particular situations. These include where power/heat/space is really expensive or where you need a lot of hosts without a lot of performance (like QA, staging and development environments). Remember, that while they use less space, power, etc., they also use laptop/low-power cpus and hard drives, so the performance can be lower, especially for i/o intensive operations. If you're not hugely space-constrained, using 1U servers will save you money in the long run.
Thanks,
Matt
me@mzi.to
Where Linux will really shine is the new PICMG-2.16 standard. It's an enhancement/alternative to CompactPCI where a chassis uses Ethernet signalling on the backplane instead of CompactPCI signals. That means a single chassis can have an intel, Sun, and/or motorola blade in the same chassis and they communicate via TCP/IP instead of hardware-specific signalling. It also means that a Linux-based blade can work in *any* manufacturers chassis. This removes a big barrier of entry for the Linux in the telecom market.
Other cool things about PICMG 2-16 Blades:
- Blades (like ethernet hosts) are more easily hot-swappable
- Depending on the chassis switch, bus speeds could approach 24GB/s in the near future
- Device drivers need only speak TCP/IP (one driver works on multiple blade operating systems)
For more info see: The Next Big Thing (pdf) and there might be something here since these guys designed part of the spec.Just because the blades are "proprietary" doesn't mean they're bad. They're denser, thus easier to physically manage and run with lower power requirements than other types of servers. Just because they weren't created by a committee of "free-thinking" open source advocates doesn't mean they're useless to companies who need more processing power at lower cost.
Seriously, the commercial market offers added value in their products that still lacks in many open source projects.
From there, they figured out a functional failover system and set up four geographically distributed data centers.
Oh, and they coded up a search engine thing at the same time.
No, because honestly the question didn't make any sense.
AMD chips with officially supported 166 mhz FSB will be ariving shortly. The revamped thoroughbred "b" core can run at a 166 mhz FSB and the 2100 mhz XP 2600 has a 16x multiplier. That is a 2.6 ghz Athlon or a 3400+ in AMD speak.
Now if these were actually real chips and not paper launches, it woudl all mean something.
... Governments are instituted among Men, deriving their just Powers from the Consent of the Governed...
I'm imaging a new market on the horizon for low-powered PCs. Laptops that can run for 6 hours+ without heavy batteries, and the sub-micro ATX form factor systems. The latter are an interesting case, useful for roll-your-own multimedia appliances and servers that you can leave on but won't chomp on your power budget. I have a PC I keep on constantly in a corner and use for my firewall, mailserver etc. I used an underclocked celeron to keep the heat down and to keep the power usage to a minimum. But it could do so much more if it wasn't so lightweight. The LP Pentium III would easily outpace what I currently have.
So, it would be cool to see these chips and motherboard commoditzed for just this use. For a bit of extra money up front you can get double your money back in power savings (vs a high-power CPU). There aren't many sub 11W IA processors that can get the job done.
Black holes are where the Matrix raised SIGFPE
True Open Standard Blade Servers are just around the corner. Up until now the current offerings by RLX, HP and IBM have been proprietary blade server designs. The next generation blade servers will be based on an open hardware standards where different vendors blades can be swapped with each other the same way that Compact-PCI is a standard blade design where all cpu boards are interchangeable with each other.
Low power CPU's are needed for the current crop of blade server designs since they forgot to deal with any heat management. The current blade designs rely entirely on airflow across the cpu package for cooling in a 2U or 3U high blade with 0.7" between each blade. Oops!!... how many blades can you stuff into a rack with each processor pulling 30 - 60 watts each and keep the temp down to 1K cpus per 42U rack) while still using Xeon and other x86 processors that produce over 60W of heat each.
Quidquid latine dictum sit altum viditur
The really interesting thing is that as it is used it appears to be faster than the same clock speed pentium. What? you say. How can this be, since transmeta has a rep for being slow.
Well it truns out that for scientific applications, ones where you tend to sit in tight loops a lot the thing is faster. It's meta chips compile the intel instruructions into its internal processor code. Once the overhead of compiling is over its faster internally than a pentium 3
The reason it got a bad rep for being slow is that for GUI type applications where the code is running all over the place and never doing the same thing for very long, it loses out.
given the incredible stability (120 days no reboot), the increacing speed of the transmeta chips (1.2 Ghz), and the extreme low power, high density and no need for special cooling these things may revolutionize scientific and industrial computuing. But they may not dent the desktop market for raw power in GUI applications.
Some drink at the fountain of knowledge. Others just gargle.
nope. because you can get 48 P3 blade processors in a 3U rack sharing power supplies and swithces. And most importantly it works without thermal problems. Lower speed cpus also mean lower speed everthing else so costs go down dramtically. Cooling costs especially. You can make up for slow speed with more cheap blades. for many web type applications cpu speed is not too big a deal. for many industrial and scientific apps the they are easy to effieciently parallelize. And for many server applications the disk acess is what limits you anyhow.
Some drink at the fountain of knowledge. Others just gargle.
How did Google roll out 10,000 servers at such a low cost?" Am I supposed to know what timothy is talking about here? I sure don't. Google hasn't informed me of any "low cost" and the article and timothy's write-up don't say anything (else) about it. Perhaps if he's going to make a big deal of something he should explain what is revolutionary or amazing here, or at least do more than imply a special amazing but unmentioned price.
Not only that, but the last line of the ZDNEt write-up says The new 800MHz chip, which uses ServerWorks' LE3 chipset, will list for $289 each in 1,000 unit quantities. OK, low power is nice, dual processors are OK, but hardly anything special, particularly when they only run at 800 mhz. After all, the reason for a dual processor is to gain more processing power and speed, but a dual processor 800 meg chip will not perform as well as a simple single processor 1600 mhz chip and is more complex to program for. A single processor 1600mhz AMD chip is less expensive and will outperform this chip. I see no reason to get excited if the cpu chip price is $289!
I'm an American. I love this country and the freedoms that we used to have.
Applications?
1. Citrix farm. They're NOT disk intensive. You can do load balancing on them. If one goes down the ICA client only has to hit reconnect. .. back up no biggy, nobody sees or knows the difference.
2. Web service farm. One server goes down (MS), kernel panics for some reason...remote reboot
3. Novell (or NT) clusters. Exchange or Groupwise. Box dies / need to upgrade..
4. Home control system..Building control system. have 2-6 blades controlling different things..
there's a lot of benefit from cheap blade dual proc boxes..
= Grow a brain...
First of all, the power consumption of an AMD XP running 1600Mhz (or even a 1600+) uses significantly more power than two of these 800Mhz CPU's. Talking about one CPU is one thing, but 40? Or 200? Now you're talking about enough heat generated to cause spontaneous combustion. Low power is a good thing.
Also, it is not just the Mhz that determines the usefulness of a given configuration. Case in point, for many large multi-user database applications the number of concurrent processes (so many per CPU based on the app itself) that the system can do is much more important than the clock speed of the CPU's. Hence the need for dual, quad and oct servers and clustering with shared storage.
Moekandu
"It is a sad time when a family can be torn apart by something as simple as a pack of wild dogs."
Mediocrity knows nothing higher than itself; but talent instantly recognizes genius. -- Sir Arthur Conan Doyle
Not quite. Sun's LX50 mentions support for the three Ultra3 SCSI hard drives, as does the 1RU 1000, but the 1RU 1000 also has two internal drivebays for ATA hard drives, which Sun doesn't mention. All in all, I think they are about as similar as any other 1U server out there.
The notion of "open" makes sense for hardware, although it is slightly different than from software. "Open" hardware that is documented, hardware that conforms to standards, hardware that has well-defined interfaces for software, hardware that is at least licensed under reasonable and non-discriminatory terms. RS232C, parallel ports, PC104, PCI, ISA, USB, IDE, etc., all can be considered reasonably open. Stuff that comes only from a single company, requires proprietary drivers, etc., is not open.
An "open" standard for blade servers would be nice. And, in fact, there are such standards: passive PCI backplanes, networking backplanes, and EuroBoards. Look around the web--there are plenty of systems to build open blade servers on--servers that are open in terms of both hardware and software.
just to nitpick...
After all, the reason for a dual processor is to gain more processing power and speed, but a dual processor 800 meg chip will not perform as well as a simple single processor 1600 mhz chip and is more complex to program for.
Well, it depends. Well written web applications under a moderate to heavy load tend to perform better under the multi-processor configuration. More complex to program for? Yessss....scalability often is.
meh.
Here's your pictures. I can't believe you missed this, /. covered it way back in April.
Click here if you just like to click on shit.
How did Google roll out 10,000 servers at such a low cost?
Certainly not by using blade servers. Contrary to popular belief, blade servers cost more tran their non-blade equivalents. Just like notebooks vs. laptops. Their selling points are (in some vendors' opinions) integrated management and supposed flexibility.
I wonder whether we will see a small blade housing being sold for desktop use. A box that sits under your desk and holds between one and four blades (possibly made to look like a single machine with MOSIX). In other words could the 'blade' form factor become a rival for ATX?
-- Ed Avis ed@membled.com
It also means that a Linux-based blade can work in *any* manufacturers chassis. This removes a big barrier of entry for the Linux in the telecom market.
Power Appliance is already available, running Oracle 9i on Linux on up to 48 dual-processor (PIII, 800-1200 Mhz) blades in a single chassis, and the software to manage them.
The p4's fpu might be weak without using the sse2 extensions,
How is that supposed to be a problem? CFLAGS="-mfpmath=sse2"
Yes, remember Transmeta's Crusoe processor?
http://www.rlx.com - they're usually at LinuxWorld showing off their blade server units. They sell them with either Pentium III chips or Crusoe chips.
Pretty cool stuff..
1. Start with a 24" rack, 72" tall. Rip the doors off the front and back.
2. Get sheet-metal 24" trays to fit into the rack. Mount them every 2U, on both the front and back of the rack. Leave a few U open in the middle of the rack for your switch and KVM.
3. Contract a company to build you custom power supplies that are 1U tall, use 90w of power, and only have 1 ATX connector and 1 molex hookup for a hard drive.
4. Put two Tyan dual-PIII mini-ATX motherboards w/ onboard LAN and video side-by-side on each tray. Slap two 1ghz PIII's in there with good passive heatsinks. Add a small amount of RAM (128-256mb) and strap a 10-20gb hard drive to the free space on the tray using a velcro strap.
5. Cluster 'em up! Heat is a HUGE problem, even with using the relatively-cold PIII's instead of P4's or Athlon MP's.
After seeing the Ashburn facility in person a year or so ago, I figured out that it would have cost about $700 per node to build the cluster. Considering it was an approximately 960-node setup, it was most likely around $700,000 for the 1920-processor cluster. That's REALLY freakin' cheap!
.... um, i lost you after "0110100001101001".
One major limiting factor is a 500W per square foot limit in most hosting facilities (that's why a lot of their systems are still PIII based). But if a low-power blade cost only 20% more per "MIP", it might still be cheaper to pay the server facility for the extra cooling and power.
They said the HP/IBM/Dell salesmen just cry because there's no way those vendors can compete with the cheapest daily far-East motherboard import prices. The only salesmen who must like Google are the ones who sell them the diesel locomotives (err. backup power generators) when they exceed the power limit of some hosting facility.
If you're using your own real estate, it's pretty easy to power the things, but if you're buying commercial hosting space, blade servers and 1U rack servers quickly start running into problems with electricity. The problem is that Intel/AMD CPUs are fairly power-intensive, and increasing the density by a factor of 5-20 over traditional PC designs also increases the amount of power that a rack of servers needs to levels beyond what the typical hosting center is designed for. If you're getting a rack with 2 20-amp circuits, you've got 4KW to play with - doesn't go very far if you've got to feed 200 or 336 Xeon chips, and for that matter, isn't really ideal for 42 1U rack-mounted boxes, if you want to have redundant power supplies and you're burning 75W per CPU plus some more power for the disk drives.
And of course, all those watts of heat require cooling. If you're planning to do it, have a serious talk with your real estate suppliers.
Bill Stewart
New Fast-Compression-only CPR http://preview.tinyurl.com/dy575ks
Getting the heat out won't be a problem if you can't get the electricity in to power them. If you're using your own real estate, it's one thing, but if you're actually using 60KW of electricity in one rack, that's about how much power a typical colo center provides for 10-40 racks of servers, depending on how you're counting redundant power feeds. If you're trying to fit that many processors in one rack, and using heavy-power Xeons instead of low-power Transmetas, you need to start looking at room airflow and not just in-box airflow. The obvious solution is to imitate a Cray-2, and use Fluorinert or some other liquid fluorocarbon coolant piped in from a big honking Air Conditioner outside your building, possibly combined with some kind of gas turbine to turn some of that waste heat back into electricity.
Bill Stewart
New Fast-Compression-only CPR http://preview.tinyurl.com/dy575ks