Device and storage vendors are gradually embracing flash SSDs (solid-state disks) as a complement to spinning disk drives, but one executive at this week's Flash Memory Summit sees great potential in turning flash "upside down," as a lower tier of cache behind DRAM.
NAND flash allows faster reading and writing of data than HDDs (hard disk drives) because it has no moving parts, and it also consumes less energy than the spinning disks. However, it costs more per gigabyte than HDDs and isn't always needed, so in enterprises, flash SSDs are being implemented typically alongside disks as a higher, faster tier of storage for frequently accessed information. Research company IDC expects enterprise SSD sales to grow by an average of 165 percent annually until 2013.
Another promising use of flash is to extend the effective capacity of DRAM, said Andy Walls, the technical lead for SSDs in IBM's Systems and Technology division. This technique turns flash "upside down" because it treats the technology as the bottom tier of capacity beneath memory instead of the top tier of storage above disks, Walls said in a keynote address Tuesday..
There are key advantages to this type of implementation, especially cost savings. DRAM costs about 12 times as much per gigabyte as flash, and that gap is growing, according to analyst Jim Handy of Objective Analysis.
Extending DRAM with flash could be a boon to enterprises that run memory-intensive applications, including most applications that use Java, IBM's Walls said. Java applications can quickly consume the available memory in a server, he said. Any task that requires a lot of DRAM, such as logging database changes, could benefit, he said.
One place where flash might be useful is with in-memory databases, such as those that use IBM's SolidDB relational in-memory database software, which stores an entire database in DRAM for high performance, Walls said.
The biggest challenge to adopting flash in this role is the lack of software that knows how to treat flash as a secondary form of memory, Walls said. A confessed hardware guy, Walls said he's not leaving it up to application and operating system developers to change their code.
"My experience is that software doesn't change. For a long time, I suggested we change the name of software to 'impossibleware,'" Walls joked. The capability to use flash as memory in most cases will have to come from middleware, hypervisors or firmware, he said. That work is likely to be done within the next two years, Walls said in an interview at the conference.
While Walls emphasized enterprise data centers for this use of flash, Handy of Objective Analysis said he expects to see it appear in client devices as well. The main barrier there has been that PC operating systems are designed to treat memory and storage as separate things, and where SSDs have been used in PCs, they have been grouped in with the hard drive as storage.
However, with the cost benefits of flash versus DRAM, SSDs are likely to begin taking on a memory role, Handy said. While many observers have been playing up flash as a faster alternative to spinning disks, eventually it will be a greater threat to DRAM, Handy said.