June 26, 2008 | Comments: (0)

Good-bye DRAM, hello flash?

TAGS: Storage

For some datacenter operators out there, insufficient server processing power isn't driving them to adopt more and more servers. Rather, it's the lack of precious server memory, necessary to deliver results at the lightning speed users have come to expect -- nay, demand -- from search engines, social networking sites, e-commerce sites, and similar Internet-based applications. A pair of companies, Virident and Spansion, have announced a remedy to the problem: replacing (or, more accurately, supplementing) the traditional DRAM found on servers with a flavor of flash memory called EcoRAM, capable of boosting a single server's memory beyond today's 32GB limit to a capacity of 512GB -- without increasing the machine's power envelope.

EcoRAM comes from Spansion (previously a joint venture of AMD and Fujitsu), while Virident brings to the table a silicon, software, and hardware platform called GreenGateway, designed to effectively transform flash RAM into DRAM-class memory. Combining the technologies opens up the opportunity for organizations running data-centric Web apps -- ones that perform frequent reads and infrequent writes -- to do with one server what they previously could do with four, the companies say. Given the struggles organizations are facing with high energy costs, limited power budgets, and finite datacenter space, that's significant.

[ Other vendors are touting memory as a means of boosting server efficiency. For more, please read "Memories of green." ]

"There is a class of applications in datacenters that are constrained by the addressable main memory a server can handle," says Rufus Connell, vice president, North American ICT (Information and Communication Technology) Research Practice for consulting company Frost & Sullivan. "As memory-bounded apps such as search use grow, conventional server architectures are failing to keep up with the increasing number of queries per second from consumers, which means more servers are needed and more power is needed to run and cool them. It is a real problem for datacenter operators globally, and a solution to help save energy and reduce total cost of ownership is needed sooner rather than later."

The kind of applications Connell is alluding to require a high level of memory sitting on the motherboard, close to the CPU, such that it can access and deliver data quickly. "Today's servers were not built with the data-centric needs of the Internet in mind. As a result, compute-centric servers in Internet datacenters can be made far more efficient with faster access to larger main memories," says Raj Parekh, co-founder and CEO of Virident (and former CTO at Sun). "We created GreenGateway to enable flash memory to replace DRAM and deliver performance and energy efficiency. ... The platform will enable Internet companies to access far larger main memories, achieving growth while living within the pressing power, space, and cost constraints of the datacenter."

The companies are currently targeting search as the application for their technologies, but they expect EcoRAM and GreenGateway can be effectively applied to applications such as data mining, databases, and biometrics. "I wouldn't call it a niche technology. We're probably talking about 40 to 50 percent of servers out there," says Hans Wildenberg, executive president of Spansion's media storage division.

The companies argue that their flash memory trumps DRAM in a number of ways. For example, DRAM is limited by energy constraints at the DIMM level and a limited number of DIMM sockets. "The typical DIMM is 2GB to 4GB. We can build up to 32[GB]," says Wildenberg, and that's in part thanks to the fact that EcoRAM consumes one-eighth the power of DRAM.

Notably, the companies say they've devised a way to rearchitect the memory subsystem so that it is still compatible with existing server designs. This is accomplished by leveraging the standard memory DIMM form factor while requiring little platform modification.

The result is higher density: "If you have a 32GB server today [from DRAM] ... we can move that server straight to 128[GB] or 256[GB] or 512GB," he says.

A year from now, in fact, he expects to see servers with a full terabyte of flash memory.

Moreover, they say that EcoRAM is superior to NOR flash, which has slower write performance and lower density than required by the data-centric applications they're targeting.

Performance is a critical component in all this, of course, given that the companies are targeting applications that require extremely fast reads. The company says, for example, that its EcoRAM is around 800 times faster at reads than NAND flash. At the moment, however, it performs at 80 percent of the speed of DRAM-only, according to John Nation, director of corporate marketing at Spansion. "[We] expect to be at virtually the same level with final silicon and software from the two companies," he adds.

Now rather than messing with new flavors of in-server memory to boost performance, what about using attached storage such as SSDs (solid-state drives) as a supplement, you might wonder. Wildenberg says they can't deliver the performance users need from these types of applications. "Solid state is not able to handle these types of applications. Users would not be able to have the experience of instance access," he says.

I decided to run this technology past InfoWorld senior analyst Mario Apicella for his perspective, as he lives and breathes storage. His take: "Storage vendors like EMC, who don't have control over servers, will swear on deployment within the storage domain and will ridicule others such as Sun, HP, and IBM, who are considering instead deployment inside the server."

He continues: "I am convinced that these two different approaches serve different requirements and that there is room for both in a datacenter. However, the server I/O bottlenecks we bumped into while doing the SATA drives review are real, and deploying flash as an extension of DRAM will bypass that narrow I/O path."

As always, the proof will be in the pudding. For the time being, the companies have had undisclosed companies testing prototypes of the memory technology. They're also currently in talks with server vendors and OEMs. Expect to see EcoRAM products shipping once summer ends.

Posted by Ted Samson on June 26, 2008 03:00 AM

October 04, 2007 | Comments: (0)

Storage grows greener

TAGS: Green technology, Hardware, Hitachi Data Systems, MAID, Storage

Servers tend to take center stage in green datacenter discussions, and for good reason. Those data-crunchin', app-runnin' machines are currently the most notorious power hogs in the datacenter joint. But storage hardware is rapidly catching up in terms of power consumption, according to The Green Data Project. By 2008, power consumption by disk is expected to exceed that of all other equipment.

[InfoClipz: Green tech | Video: Hitachi CTO talks green storage ]

There are a couple of reasons for that. First, data simply continues to pile up. Second, as that data piles up, datacenter operators have an unfortunate habit of throwing more storage at the problem, rather than making better use of what they've got. That approach ends up costing you in several significant ways. You're laying down cash for new or leased storage arrays. You're paying monthly bills to power and cool those arrays. And you're sacrificing precious datacenter space for potentially superfluous hardware.

Surprise, surprise, storage vendors out there are keenly aware of the problem and are devising storage technology and products that could play a critical role in your green datacenter blueprints. Among those vendors is Hitachi Data Systems (HDS). The company's CTO, Hu Yoshida, recently sat down with me to talk about the subject. (You can watch the video here.)

Data diet
One resource-saving storage technology being pushed by HDS and other vendors, including HP, is thin provisioning. The premise behind the technology is pretty straightforward: It lets IT admins view all of their storage hardware as a great big pool and divvy up slices as needed, rather than allocating separate arrays for different business units that might not be taking full advantage of the pricey hardware you've set aside for them.

The result, if all goes well: You can purchase just enough storage machinery to meet your organization's collective needs, which means you're paying less money for arrays that are sitting there running at 20 percent utilization while you pay for 100 percent of their drives to spin. That doesn't make a lot of sense economically or environmentally. (There's also the benefit of easier management, but my sites are set on green.)

For thin provisioning to be effective, careful planning is, of course, critical, a point that's certainly not lost on InfoWorld storage guru Mario Apicella. If various departments estimate they'll need 8TB of storage but, lo, it turns out they collectively needed 10TB (oops!), you're left in an unenviable situation of playing storage catch-up.

But then, that's the challenge of running a green datacenter in general: You have to be sure you forecast and budget intelligently such that you always have just enough servers, storage arrays, and other hardware up and running at one time. That's fodder for another blog post, perhaps. On to additional green storage approaches.

Put the brakes on drives
In my interview with Yoshida, he talked about the new power-management capabilities, called Power Savings Storage Service (or secretively, PSSS), which the company recently introduced in its mid-range Adaptable Modular Storage (AMS) and Workgroup Modular Storage (WMS) systems. Once again, the premise is pretty straightforward: You have an array with various drives, but this feature gives you the ability to put drives to sleep when they're not being used. That means, of course, that you're not paying to power something pointlessly.

This technology might sound familiar to you if you've read about MAID (massive arrays of idle disk) technology. MAID storage systems also have drives which can be woken up individually and put to sleep. But there are key differences, which Yoshida has taken pains to clarify in his own blog. He goes so far as to say that PSSS is the opposite of MAID: PSSS works by putting active drives to sleep whereas MAID works by waking up sleeping drives.

The approach differs because the applications are different. The AMD and WMS arrays are RAID systems storing production data. "These HDS RAID arrays are production arrays that have no restrictions on writing or reading and can have high-performance [Fibre Channel] RAID groups and/or lower cost SATA RAID groups," Yoshida writes. "Instead of keeping all the disks idle until they are accessed, this feature enables a user or scripted application to power down a RAID group when it is not being accessed and power it back up when it needs to be accessed."

MAID, on the other hand, is geared for SATA drives storing write once, read occasionally data; that is, data stored on large-capacity drives that you access infrequently and can afford to wait for.

Yoshida concedes that MAID does technically consume less energy, which isn't surprising, considering that California utility PG&E has offered incentives for companies to enlist MAID. Byte and Switch reports that "users have already shaved around 75 percent off their energy bills through MAID, thanks to the fact that disks are typically idle. HDS, on the other hand, is touting potential savings of around 20 percent as a result of its reliance on active disks."

Of course, there's no reason you can't employ MAID and HDS's arrays at the same time. In our interview, Yoshida noted that they can certainly be complementary. Yes, MAID might use fewer kilowatts, but when you're building a green datacenter, you can't simply focus on the product that uses the least energy -- another point that Yoshida emphasized. Rather, the idea is to assemble the equipment that meets your present and future needs while wasting the fewest resources possible. Again, that takes plenty of planning, but the payoff can nonetheless be significant.

Posted by Ted Samson on October 4, 2007 03:00 AM