Dr. Peter J. Varman, professor of electrical and computer engineering and computer science, has received a three-year grant from the National Science Foundation for research related to data storage.
The project, titled, “Architecture and Software for Scalable Persistent Memory,” seeks to blur the distinction between volatile memory and non-volatile storage, using a single memory abstraction to combine the advantages of both interfaces.
Main memory, made up of DRAM, has the advantage of being fast and can be directly accessed by the processor without going through the Operating System’s layers of software. The downside to DRAM is that if there is a power failure, the contents of memory are lost.
Traditional storage systems, such as magnetic hard disks or Solid State Disks, are slow and store memory in block format, which requires software translation for the data to work when brought over to memory. But, these formats can survive power failure.
Dr. Varman proposes to use Storage Class Memory (SCM) in his research. SCM devices retain data should the power fail, and are also directly accessible by the processor.
“SCM combines the speed and access characteristics of DRAM with the persistence of traditional storage. SCM-based systems can get the performance benefits of in-memory databases and have the additional advantage of being non-volatile and persistent,” Varman explained.
“In current systems, programmers have to wear two hats – manage data that resides on magnetic hard disks or Solid State disks, as well as manage the same data when it is brought to memory to be processed. By having a single device that is both persistent and can be treated like DRAM, programmers don’t have to switch between these views,” Varman said.
“There are a lot of advantages to using SCM devices. But, they come with their own set of headaches. Any time we deal with persistent data we need to make sure that data is updated consistently”.
“Think of transferring $100 between two bank accounts. If I subtract $100 from one account and the machine crashes before I add $100 to the other account, we have created an inconsistent database. So, database management systems and operating systems do a lot of work behind the scenes to make sure that even if the machine crashes the system recovers to a consistent state – either both accounts are updated or neither is,” he explained.
With support from Intel Corp, the National Science Foundation and the Ken Kennedy Institute at Rice, Varman’s group is exploring scalable hardware and software techniques that guarantee atomicity and durability of storage transactions along with application-determined isolation and consistency mechanisms. The research has direct impact on in-memory computing identified by Gartner as a major industry trend.