Ramcloud makes future data center without hard drive

In today's era, solid-state hard disk has become a hot technology in the enterprise. But the computer science team at Stanford University thinks there is a better solution. They have come up with scalable and high-performance storage methods called Ramcloud. This solution can store all the data in dynamic random memory. You can also aggregate memory resources across the data center.

The scalability and performance aspects of Ramcloud make it a key technology for high-performance computing. In particular, data-intensive applications, which are typically performed in compute-intensive applications.

The Ramcloud project was launched by John Ousterhout, a professor at Stanford University. He is also the inventor of TCL's scripting language. He is also chairman of the Tepco Cloud company, TEPCO Cloud, founded in 2002, whose main business is to provide high-performance software component tools. Ousterhout in the field of high performance computing, and research on distributed file system and high performance file system. In short, Ramcloud is a software platform that collects all application data in memory in a large number of servers in a data center or cluster. Ramcloud delivers 100-1000 times less latency and 100-1000 times higher throughput than disk-based storage.

The researchers believe that the current reliance on hard disk technology will not be enough to deal with data-intensive applications, a trend that is spreading to all aspects of enterprise computing. The researchers also announced in the article that if Ramcloud succeeds, it will replace the disk as the main storage technology in the datacenter.

The two most attractive features of Ramcloud are the ability to scale to thousands of servers and to maintain extremely low latency.

The Ramcloud delay can be as low as 5-10 microseconds. This is 1000 times times faster than a traditional disk, about 5 times times faster than flash memory. The researchers say that Ramcloud's low latency features are somewhat redundant for Web applications that have harsh requirements for low latency. But the development of this technology should be encouraged, and this will bring a huge performance advantage. At the same time, for some high-performance computing (HPC) applications, the low latency characteristics of ramcloud appear to be a duck in the fish.

However, the influence of network latency on Ramcloud low latency is an unavoidable problem. Although high-performance Ethernet vendors like InfiniBand, Myrinet, and Arista can delay data transfer across the Datacenter 10 microseconds. Today, however, most devices use TCP/IP Ethernet latency of 300-500 microseconds. Optimization of network latency is the key factor to maximize the performance of Ramcloud.

As for scalability, the researchers believe that the current commercial server and storage technology Ramcloud can be as high as 500TB. The cost of Ramcloud in the next 5-10 years could be reduced to $5 per gigabyte, and the scale could reach 1-10PB.

The chart represents the cost of three storage technologies for different size datasets and different access frequencies (Stanford University Computer Science department draws)

But in today's market, the cost of Ramcloud storage is about $60 per gigabyte. The DRAM system is slightly more expensive relative to disk and solid-state storage. DRAM storage systems are 50-100 times more expensive than traditional disks, 5-10 times more expensive than flash memory. But for high throughput I/O application dram can still reflect its value. The researchers say that if your code executes fast enough, you can break through the constraints of storing data access. From this point of view dram is actually 10-100 times lower than the traditional hard drive price.

The technology also inevitably faces problems, including accurate data models and APIs, how to optimize remote call latency, data persistence and availability, cluster management, multi-tenant applications, and support for atomic updates. However, these are problems that can be solved.

With the expansion of the data center, more and more data-intensive applications will be floated in the future. This laid the foundation for the popularization of Ramcloud. There is no timetable for a Ramcloud project to become a commercial product. However, the experience of ousterhout in Tepco Cloud can help Ramcloud to be applied in the commercial field as soon as possible.