There might be a trick to vastly improve the performance of most data centers and, consequently, the web itself — simply throw out the hard drives in data centers and replace them with dynamic random-access memory (DRAM).
John Ousterhout, research professor of computer science at Stanford and head of a new project called RAMCloud, proposes exactly that: Create a data center storage solution “where information is kept entirely in DRAM and large-scale systems are created by aggregating the main memories of thousands of commodity servers.”
It sounds simple, but also preposterous. Everyone would like to be able to keep all the information in fast, random-access memory, but it’s too expensive. Hence, for long-term storage, we use hard drives.
And indeed, in terms of cost per gigabyte, Ousterhout estimates 2,000 servers could provide 48 terabyte of DRAM storage at $65 per gigabyte, meaning that RAMCloud-based storage would be a 50 to 100 times more expensive than hard disks.
Cost per access per second is another matter, however. Ousterhout claims that, when measured by this metric, DRAM is 10 to 100 times cheaper than disks.
Of course, price is not the only issue at hand; there are questions about durability, availability and scalability, most of which are addressed — but not entirely solved — in this white paper.
Nevertheless, the implications — if the model proves to be a success — are very interesting.
The team behind the project estimates that RAMClouds can provide storage with 100 to 1,000 times lower-access latency, and 100 to 1,000 times the throughput of disk-based storage systems. This, the RAMCloud team claims, would pave the way for “a new breed of data-intensive applications,” which, in turn, would mean a faster web for everyone.MORE