Dataram Launches SAN Accelerator Appliance
September 28, 2009 Timothy Prickett Morgan
To midrange shops, the name Dataram is pretty much synonymous with clone server main memory that is a heck of a lot less expensive than the main memory that IBM, Hewlett-Packard, Sun Microsystems, Dell, and others peddle (and grossly profit from) for their boxes. But this week, Dataram is taking the expertise it has in memory and that it gained with solid state disks through and acquisition to launch a caching appliance for storage area networks that can radically boost the performance of servers and cut down on the number of disks that servers need.
As shops using the AS/400 and its progeny are well aware of, the number of disk spindles, not the disk capacity on disk drives, is a critical aspect of a machine configured to do online transaction processing. As disks have gotten fatter with capacity, they have generally also spun at faster rates. But the increase in rotational and latency speeds has not kept pace with the much easier Moore’s Law-driving increases in disk capacity, and that means as disk drives prices per unit capacity have gone down dramatically, they have not really gone down at all in the past decade by spindle count. So companies have been short-stroking disks, using an ever-smaller slice of data on the outside of the platters to boost performance, essentially ignoring all the capacity they supposedly bought.
For its part, IBM has been making the case for new solid state disks it announced earlier this year for Power Systems machines. IBM is taking a 2.5-inch SSD drive it gets from STEC, formatting it down to 69.7 GB, and adding algorithms into the i and AIX operating system to move data somewhat automatically from disk arrays into the much faster flash-based SSDs. This means a server can do the same work with a lot fewer disk drives because the flash drives IBM is selling can deliver 135,000 I/O operations per second, compared to the several hundred IOPS (not to be confused with AS/400 architecture I/O Processors) a 15K RPM SAS drive can yield. However, the IBM SSD costs $10,000 a pop for Power 520 and Power 550 machines and $13,325 for Power 575 and Power 595 machines.
But these SSDs for Power Systems do zero for companies that have servers accessing external, shared, SAN-style disks through Fibre Channel links.
Solid state disks are nothing new to Dataram, which created SSDs and shipped them back in 1977. (That’s not a typo.) And a year ago, Dataram bought a solid state disk maker called Cenatek, which was founded in 2000 and which had created a product called RAMDisk that uses main memory in servers to create a RAM-based disk drive for applications and operating systems to access. Jason Caulkins, chief technologist at Dataram and the founder of Cenatek, says that Dataram has taken its many years of experience in solid state disk storage to create a different kind of storage acceleration device, unique from the network storage and database caching appliances that are popping up all over the place, and offering significant advantages over just plunking flash drives into servers and moving datasets around by hand.
Today, Dataram is taking the wraps off a new product called XcelaSAN, which it characterizes as being “an intelligent SAN optimization appliance.” Rather than being a file-based accelerator, as most current flash implementations are when they are sitting inside of a server or in a disk array to speed up I/O access, the XcelaSAN is a block-based device that is completely agnostic about the server, operating system, file system, or storage arrays being used. By the way, the appliance doesn’t care if the device fielding requests is actually the Virtual I/O Server that IBM uses to virtualize the server side of I/O for the i, AIX, and Linux platforms.
The appliance sits between a Fibre Channel switch that links to one or more servers and the storage area networks that are shared by servers to give access to servers for common data sets. SANs are sometimes partitioned, of course, by operating system. But the XcelaSAN doesn’t care about that. When a server asks for data, it finds the blocks and gives them the data out of memory and flash instead of reaching back into the SAN for the data. As data changes, it is updated back on the SAN at the block level. The important thing is that servers get access to data through the XcelaSAN appliance a lot faster than the normal SAN can. Like hundreds of times faster, according to Caulkins, in terms of IOPS.
Dataram is only putting out the XcelaSAN appliance out as a technology preview right now, so it is being a bit cagey about what is in the box, but it has 128 GB of DDR2 ECC registered memory (using 4 GB DIMMs) and 360 GB of some kind of flash memory. The appliance has eight 4 Gb/sec Fibre Channel ports, and Caulkins says this is plenty fast enough since most customers have either 4 Gb/sec or 2 Gb/sec Fibre Channel links to their SANs in the midrange market it is targeting, and no one has 8 Gb/sec Fibre Channel. As customers deploy faster Fibre Channel, or even move to converged 10 Gigabit Ethernet (which supports the Fibre Channel protocol over Ethernet to link servers to storage as well as normal network traffic between servers and the outside world), Dataram can add support for these technologies to the appliance. All of this is crammed into a 2U rack appliance, and they can be daisy chained to provide redundancy and extra bandwidth between servers and SANs.
The XcelaSAN appliance has been in beta for several months, and will be demonstrated throughout October at various trade shows. The product will begin shipping in November. There is only one configuration, and it costs $65,000. That’s about what IBM is charging for four of its SSDs for high-end Power Systems boxes.
To make the case for the XcelaSAN box, Dataram cooked up this comparison. Take a typical midrange server (the architecture doesn’t really matter) and host a 1 TB database on it. To configure such a database for performance, not capacity, might take as many as 200 SAS disk spinning at 15K RPM. If that server was running, say, the TPC-C online benchmark test, the server and storage configuration (in this case no doubt an X64-Windows box), Dataram calculates that the box would have $124,000 in hardware costs, $18,000 in operating costs (over three years), and deliver 232,000 transactions per minute (TPM) on an OLTP test. Now, taking the same server and configuring it with 25 SAS drives and an XcelaSAN appliance would drop the capital expenses down to $104,000, the operating expenses for three years down to $6,000, and the resulting machine would do about 500,000 TPM on OLTP work. That’s a 22.3 percent savings in CapEx and OpEx, as they say in accounting speak, but a 63 percent improvement in price/performance.
Obviously, this XcelaSAN appliance is not aimed at entry Power Systems shops, but midrange and high-end customers would do well to investigate it further. And Dataram would do well to create a smaller appliance that can either be tucked onto a PCI-Express bus or somehow integrated into Power 520 and Power 550 boxes with a price tag and performance tuned to these smaller customers.