A Flash in the Pan or the Future of Data Storage?
September 2, 2008 Timothy Prickett Morgan
IBM‘s Hursley development lab in England and its Almaden Research Center in California are known for a lot of different innovations, particularly in the area of disk storage. Before IBM sold off its disk drive business to Hitachi a few years back, the labs in Minnesota and outside of San Jose in California were also hotbeds of disk development. IBM invented disk drives, and has a lot of knowledge about data storage in its institutional memory.
Last week, the Hursley and Almaden labs announced that they have been collaborating on a new storage device, code-named “Quicksilver,” that will marry flash-based solid state memory to traditional disk controller and array technology. IBM did not provide a lot of details about the technology used in the Quicksilver project in terms of the architecture or components, except to say that Quicksilver involved marrying flash-based drives with its storage virtualization technology. Big Blue did, however, give some comparative feeds and speeds of the Quicksilver array, which is presumably a hybrid flash-disk array. IBM said that the Quicksilver array was able to deliver over 1 million I/Os per second (commonly abbreviated IOPS) of sustained data transfers with an average response time of under 1 millisecond. IBM then compared this to its SAN Volume Controller, an out-of-band storage virtualization controller, using arrays of disk drives based on the SPC-1 benchmark test for storage, and the Quicksilver array had 250 percent better performance, 1/20th the response time (which is way better), took up one-fifth the floor space, and required only 55 percent of the power and cooling of the disk drive version. IBM tested the SAN Volume Controller with its DS4700 disk arrays (basically, a Power Systems server running storage software) in the SPC-1 test.
IBM tested a 4.1 TB array of flash drives with the Quicksilver project. The flash drives were manufactured by a company called Fusion IO. That company’s ioDrive product comes in 80 GB, 160 GB, and 320 GB capacities and can deliver 125,000 sustained reads using 1 KB random data files and 100,000 reads using 4 KB files. Access time is 50 microseconds for reads (no data is given for writes), and Fusion I/O says the ioDrive can do 700 MB/sec of sustained reads and 600 MB/sec of sustained writes. The device plugs into a PCI-Express x4 slot and supports Red Hat Enterprise Linux 4 and 5 and Novell SUSE Linux Enterprise Server 9 and 10. The entry 80 GB drive costs $2,400. IBM did not say what it plugged the ioDrives into, and how these were lashed to the SAN Volume Controller to do the test.
Now, while flash drives have some great performance characteristics in terms of sequential I/O data transfer rates and low power usage, flash drives, gigabyte for gigabyte, are still a lot more expensive than disk drives and their performance is not very good on random I/O–you know, the kind that is going on inside the online transaction processing systems that are the hallmark of the IBM installed base of mainframes, AS/400s, Unix boxes, and even Windows servers with a smattering of Linux boxes these days. Giant flash arrays might, however, be particularly helpful in cases where large data sets have to be moved very quickly into systems for processing. IBM says that the Quicksilver flash array will allow certain kinds of applications—reservation systems, financial trading systems, and data warehousing and data mining, for instance–to see big performance gains in certain parts of their workloads.
Flash drives are not, however, going to replace disk drives in commercial computers–and particularly in servers–any time soon. At least not completely, and IBM is certainly not trying to suggest that flash drives–or any other kind of solid state storage–will. While read IOPS are high with flash, write speeds are on the order of one-sixth to one-tenth the speed of reading data.
Incidentally, IBM has been shipping a flash-based solid state drive for its BladeCenter blade servers since last June. The device was never expected to be more than a niche product for a select bunch of customers, since most blade server users go diskless and attach their blades to remote SANs through Fibre Channel or iSCSI links. With something more like Quicksilver, the use of flash-based disk becomes much more likely in certain data center environments. Particularly if IBM can get the cost of flash drives down by helping push them in greater volumes. That latter bit is a bit hard to predict. Many would have guessed we would have had memory-based solid state storage being a component of systems already. But faster and fatter L2 and L3 cache memories, cache memories on disk drives and on RAID disk controllers, as well as ever-faster and cheaper disk capacity have made it possible to not have to architect solid state storage into most commercial servers.