Anything But a Flash in the Pan
There are only two kinds of storage devices - those that have failed, and those that are about to fail. That's the view most datacenters have about the traditionally mechanical devices pejoratively referred to as "spinning rust." All disk drives fail, cheap drives fail faster.
If the average time to fail is five years, you and your laptop can make do with the occasional backup. But when an average enterprise has 100, or 1,000, or increasingly 10,000 or 100,000 individual disk drives, failure is a daily, if not hourly occurrence. Mechanical devices fail.
And with failure comes the potential for losing data - using commodity disks to save your boss $500,000 does her no good if she's fined $50,000,000 for violating data retention regulations. Stock transactions, medical images or feature length movies - take your pick, some data has to be perfect. Not a decimal point or pixel out of place.
That's exactly why, years ago, Sun invented a storage platform called ZFS. ZFS makes a powerful assumption - that a reliable system must be built from unreliable parts. By using surplus computing cycles, ZFS constantly runs powerful integrity checks, giving data corruption no place to hide. With ZFS, customers can use the cheapest disks and simplest systems, and get exceptional data integrity, along with massive reductions in cost and complexity.
But there's a new option on the table, known to many by the memory cards they use in their phones, iPods or digital cameras - called Flash memory. Flash is very fast at reading and writing data, like DRAM (the memory chips in your computer). Its price sits squarely between DRAM and traditional disk drives. But unlike either alternative, Flash requires no power to remember data. And with the price of electricity escalating across the world, keeping 10,000 disks spinning at thousands of rpm can cost you in power what you pay for your storage. Power has become the dominant factor in high scale hardware decisions - and Flash is set to disrupt the industry.
Historically, there have been two impediments to using Flash in the enterprise.
The first was cost. Flash is more expensive per gigabyte than a comparable disk drive. But with every increase in the cost of electricity (and decay in the price of Flash memory), Flash's relative cost per available gigabyte is quickly improving - remember, disk drives have to be powered on to be available.
And although a gigabyte of mechanical disk might cost you less than a gigabyte of Flash memory, the latter is at least an order of magnitude faster in reading and writing data - so the cost per gigabyte served is exceptionally low.
But simply introducing Flash as yet another tier of storage in a datacenter isn't the real opportunity - that adds new costs and a set of new management hassles. To truly change the industry, adding Flash would have to be completely transparent to users and operators, alike, with no switching or operational cost. And that's exactly what we're doing with ZFS. ZFS will transparently incorporate Flash into the storage hierarchy of a running system, using the microprocessor cache for the most performance sensitive tasks, DRAM for the next, then Flash, then disk (then ultimately tape). ZFS will allow Flash to join DRAM and commodity disks to form a
hybrid pool automatically used by ZFS to achieve the best price, performance and energy efficiency conceivable. Simply put, our storage and server systems will get enormously faster - without any upgrade to the microprocessor. Adding Flash will be like adding DRAM - once it's in, there's no new administration, just new capability.
That's one reason we're so excited by Flash - the cost per available gigabyte served (the total operational cost of storage) plummets with Flash in the mix, especially for data or performance intensive applications (like MySQL, Postgres, Oracle or SQL Server). From the right systems designs, Flash has the potential of providing orders of magnitude improvements in economics and performance - and with the advent of Sun's xVM hypervisor, we can bring this performance benefit to any host operating system (running atop xVM, Windows operators can inherit the benefit of ZFS+Flash, too).
The second problem is trickier - simply put, although Flash memory can be read an infinite number of times, writing to Flash more than a few hundred thousand times can wear it out. Now, most normal humans will never hit 500,000 writes in a digital camera. But you might in your enterprise. What to do?
Again, ZFS to the rescue.
ZFS treats Flash memory like any other storage medium - remember, all storage devices fail - and manages data integrity whether failure's induced by a bad hard drive motor, write-fatigue or a hammer drill.
Increasingly sophisticated "wear leveling" algorithms are also maximizing Flash lifespans - evening out write activity to avoid hot spot failures. But the bottom line is this - with ZFS at the helm, wear is a non-issue (on hard drives or Flash - they both have wear limitations, after all).
These are the premises behind Sun's systems approach to Open Storage. We're integrating ZFS, Flash memory and some exceptional hardware/silicon innovations to deliver high performance, low power, general purpose storage and server appliances - accelerating any software that runs on our SPARC or x86 systems (MySQL users, in specific, will see a big turbo charge). For a fraction of what proprietary NAS storage ordinarily costs. Expect our first Flash systems to ship toward the end of this calendar year.
And as you might expect, ZFS and all the underlying software will be free without commercial support - OpenSolaris, ZFS, MySQL and Postgres are already available (click the image at left to get a free LiveCD - or check out Apple's release of ZFS in Mac OSX). Software revenue will come from those enterprises that want Sun's technical support for mission critical deployment. On the hardware side, Sun's Try and Buy programs allow any partner or customer to order one of our systems for free 60 day trial - if you love it, buy it. If not, we'll cover the return postage.
If the above doesn't make it obvious... our view is Flash is anything but a flash in the pan - as the price of power continues to increase, and the price of Flash continues to plummet, the combination of Flash, ZFS and true systems innovation will have an even bigger impact on datacenter economics than virtualization.
It's that big a deal.
UPDATE: You may have seen we added yet another company to the list led by Intel, IBM and Dell supporting Solaris (and thus ZFS) as OEM partners - this morning, we officially announced the addition of Fujitsu-Siemens. Congratulations, all...
Posted on 03:00PM Jun 10, 2008 | Comments[41]
Share this post
|
|
|
|
|
|
