Slashdot Mirror
Demand For Custom Datacenter Servers Rising
With his first posted submission, SpaceCracker writes "According to this Bloomberg article, hardware giants like HP and Dell are losing out to Intel and others who've adapted more quickly to the trend of shifting from traditional, off-the-shelf servers to custom-tailored machines. 'Buyers say custom servers provide a cheaper, more efficient way of meeting the boom in demand for personal data shared via the Web. A lot of that demand can be met by less expensive machines shorn of the components, upgrades and backup services that server makers traditionally offer to large corporations.'"
Not custom, just not bloated. Bloat, both in hardware, operating systems, and "standard desktops" has become so bad that people are willing to pay a premium to remove some of it. The high end Intel server I got from System76 had nothing on it I did not need, and was cheaper than others, even while being "custom." The Linux drivers were a plus as well.
If you want 12 core Opterons you are automatically required to get a 2U machine from dell. It does not matter than the 1U machines could use those parts, they do this to improve their margins.
Lots of stuff like that, feature X only is enabled on Platform Y. Intentional crippling of hardware leads to this big buyers side stepping these vendors.
I don't know about Dell or anything precisely along the lines of what you describe, but there *is* more to server design than 'does it fit in the socket'. If the components chosen for a 1U skimped on cooling to a certain TDP and the Magny Cours exceeds that, then they may not have enough room around the socket to accommodate a heatsink that can dissipate the heat given the flow rate.
In terms of 'custom' unclear to what extent they are talking about board components (which have been increasingly sparse in the tier one vendors) or mechanical issues. In either case, the truth is when you are building entire datacenters and promise a vendor a run of tens of thousands of servers, you can get what you want. Most of the market isn't doing things that cleanly, and that's the market that is front and center in the Dell/HP/IBM offerings that you see on their respective websites. In the long haul, engineering everything from CRAC to network switch, to rack, to board, to chips is going to give you the best/most efficient approach, *but* it involves a large expense in a single shot, *can* get you stuck in a difficult situation if the state of hardware changes in 2 years.
XML is like violence. If it doesn't solve the problem, use more.
I tried these guys (just noticed they switched from com to nl, interesting, wtf?
They are not as refined as the big ones, but they provide bang for the buck.
You can't handle the truth.
With Compaq SmartStart.
What a hell! Then, they required it for initial boot! And every other manufacturer saw this as worthy of emulation.
They were awfully captive to a Windows-oriented market, and driven by MS.
"Flyin' in just a sweet place,
Never been known to fail..."
HP and Dell are also killing themselves on SSD prices as datacenters move to these for there increased reliability and performance. HP and Dell are both anywhere from 10 to 20x the prices of other parties.
Also TFA is about datacenter servers, which always come with either get re-imaged by large customers to whatever there operational image is. Almost none of the people listed here would deploy on OS that "ships" with the box, so all the bloatware complaints are idiotic.
IBM, HP, Dell, etc have quit innovating LONG ago. Now, they are busy either shifting operations offshore to make up for horrible management, or they simply sell the unit (again to make up for horrible management, marketing, sales, etc).
I prefer the "u" in honour as it seems to be missing these days.
We're buying large quantities of SuperMicro blade servers. Pretty minimal, but just right for our high density stuff.
Only "advanced stuff" is good PCI-X riser for a SAS card and good IPMI, plus 4 blades in a 2U chassis...
If they think 98% of the market is going to go with the onboard, it's cheaper for them to have a single part with an unused component 2% of the time than to maintain two motherboards with different chips populated with independent replacement stock (or to make it a field pluggable module). If they relax warranty promises or have a handful of customers driving tens of thousands of servers for one run, they can (and do, as pointed out in the article) make exceptions.
XML is like violence. If it doesn't solve the problem, use more.
Next generation of Windows server running on small, cheap, low power, ARM, system-on-chip servers. Cheap, commodity stuff. Parallel to MS starting to use this to supply their hosted Azure services, I can see private cloud in enterprise running on much the same sort of thing. Stateless machines, much like Azure. It won't suit everybody, but I bet this is the way it's going to go. Why burn power on Intel machines when you can save heat, power and space with SoC stuff?
This is a guess only, but I think it'd make sense for MS...
Where do these vendors source their hardware from? Do they go straight to Taiwan with specs in hand and have a production run of boards done? What happens when they need spares? Do they just buy a whole slew of spares and then when those run out, move onto the next design?
What happens with industry wide problems, like with what Intel experienced with the SandyBridge CPUs? I ran into some weird issues where a couple of my Dell blades were not seeing all of the RAM. They had a firmware fix available for it. What happens in similar situations with these custom built solutions where companies like Google are designing their own boards, and presumably do not have as much leverage with the BIOS developers.
The mac pro in sever use fails at a lot of that,.
Back in my day, if you wanted a special server at a datacenter you built your own. Still have a dual 550 xeon sitting around here somewhere in a 2U case that I built many moons ago. It's not an unheard of concept.
Standard off-the-shelf datacenter hardware sucks.
Facebook had a rack from their Prineville datacenter on display at OSCON. It was beautiful. The rack is actually three racks wide, and holds 90 servers total. Each server is about 1.5U high, and it's basically a motherboard, PSU and hard drives in a sliding aluminum tray. There are no front or back panels and no top covers, because none of that makes sense in a secure climate-controlled datacenter with filtered air. The DIMM slots are oriented front to back so they don't block airflow. No video cards, because you shouldn't need video on a server. There's a little release button that locks into place when you slide the tray into the rack. Simple, easy, cheap, effective.
Except, of course, that it's not actually cheap yet - Facebook's racks cost significantly more than standard racks, since they can't be produced in large quantities yet, but they make up for it with reduced operational costs. I hope they standardize and commercialize the design, because the first thing I thought of when I saw it was, "when can I buy one?"
I just got back from the datacenter, building out new racks of servers, spec'ing out more servers for future expansion, etc. If anyone should be able to understand the story, I should... but I don't. The reporter clearly doesn't understand the topic, so the story becomes completely demented. They're using completely unrelated issues, and even opposing reasoning, to paint a picture of an emerging trend.
Let's shed some light on this murky discussion:
âoePeople want to be able to build it their way,â
No! No they don't. What people want is for servers to be less expensive. Cutting out unnecessary and redundant components is simply one way to get there, without losing anything.
Dell/Hp, why do I need VGA output on my server, along with serial ports, along with a BMC (IPMI - serial-over-lan), along with an add-in OoBM card (DRAC/iLO) that does KVM? Hell, why not video-output to the two-line LCD on the front, let's do that too!
Now, none of these are bad things per-se, until you consider the added hassle. If your server doesn't have VGA out, you can count on video ALWAYS going over the serial port, come hell or high water. But when your Dell/HP COTS server decides to suddenly STOP doing serial port redirection, now that VGA output you didn't want, becomes necessary, and you have to have something wrapped around it (ie. IP-KVM).
Those OoBM cards sound great, until you find 2% have some screwed-up setting somewhere, somehow, and aren't responding... Again, you've got to have some fallback to something else.
Okay, this quickly turned into an off-topic rant, so I digress...
Google didn't build their own servers because they wanted a special RAID controller... they did it because they want CHEAP. Facebook didn't build their own servers because they want DVI instead of VGA video connectors, they did it because they want CHEAP.
There are some other added benefits to custom servers. See my rant above about OoBM, or see Google's 12V batteries. But do you think Google would object if Dell sold them the servers they wanted, except with a few extra GigE ports built-in? Hell no! It's not "I don't want X", it's "I don't need X and don't want to pay a premium to get a server with it."
If somebody would productize Google's servers, maybe adding just a FEW features so it would meet the needs of a larger customer base, and sell it nice a cheap, this "custom server" craze would be over, as the COTS version would become cheaper still. Dell/HP hanging-on to that highest-common-denomination server design, with expensive features rarely needed, and big fat margins, is driving companies to look elsewhere, and elsewhere , right now, just happens to be custom designed. Yet even with the extras on Dell/HP servers, they could still pull customers back, if they'd just drive prices down on their existing servers, until a custom product run is more expensive.
The only question is, if Dell and HP don't want to provide these cheaper servers so many companies want, who will? I don't believe it will continue to be in-house for long, but who will jump in and serve this un-served market segment? Cutting margins is how Dell became so big, who will be the next upstart to come along and out-do them at their own game?
Slashdot gets worse every day... Pipedot: News for nerds, without the corporate slant
I'd love to convert my x86 based farm into servers that don't have any expansion slots, but have a pair of 10GbE LOM (LAN On Motherboard). I don't need internal disk, just CPU and RAM. I currently boot from SAN, but if I can get rid of the PCI slots, then I can boot from FCoE instead. All my disk is concentrated in the disk arrays, so I don't have to deal with disk in the servers and I can get away from remove the power inefficient HBAs. Now if my LAN and storage admins could stop fighting over "Who gets to upgrade the FCOE switch", I might actually get headache over with.
i think there are roughly spoken 2 types of server buyers ...
1. companies without the necessary staff to "roll your own" hardware-crap:
they want 3 or 5 years on-site warranties, support and wizards for their overloaded and/or badly educated staff.
and it really does not matter for them if a server systems costs 7000 or 9000 us$/euro/whatever ... just no outage/(hardware) failures, please!!
what's the real cost of a fairly good system technician? 100 us$ per hour? do the math and calculate the hours per year for a 5 years life-cycle of server systems if you save lets say 2000 us$ on the purchase price and "shit happens"...
2. facebook, google and other big companies who do not need the extended warranties etc. from the server vendor
now try to estimate who operates more servers on this world ... not per company, but seen as a whole...
3. nobody ever got fired for choosing ibm!!
lets think about the overloaded/undereducated staff at an average small company ... then i think this old mantra is still valid :)
cheers
v.
For the likes of Dell/HP/IBM, these scenarios present a problem. These datacenters architect their solution so that the manageability and service is no big deal. A system fails and it's going to be 3 weeks before you can get a replacement in? Fine. Can't get a replacement anymore because that model is done, upgrade it to something 'close enough'. Much of Dell/HP/IBM cost compared to, say, Supermicro is in maintaining stockpiles of replacement parts, keeping them distributed across the globe, paying for expedited shipping, and employing technicians to dispatch to customer sites.
So they cut cost on some offering by exempting it from this. Suddenly, a customer who wants "tier one quality" jumps on the system due to price. Then they realize there is no longer any particular difference, and they can then go off about how that vendor is nothing like they used to be.
Of course, hardware can factor into this as well, and the major players do something about this. In the not-to-distant past, there was an intel first-party server. You could buy it form Dell, HP, or IBM. The only difference was what logo Intel's BIOS showed you and the logo on the cheap removable bezel. This is them responding to the thought. This is the same platform popularized for these cheapo deployments, and so the tier-ones embraced it. The problem is for anyone who cares about reliability, serviceability, or management, the board was utter shit. As a result, "tier one" value eroded.
Basically, the 'tier ones' have to be careful playing this game. If they do this, they should probably establish new brands to slap on, like automotive companies do. This is asinine, but human psychology seeks simplicity and this is the only way to get there it seems.
XML is like violence. If it doesn't solve the problem, use more.
You want better tires, transmission and no speed limiter? You have to buy the package with the huge wasteful engine too. You want the sports suspension, low-profile tires and higher-tuned engine? You need the sports package with the automatic transmission (I kid you not).