RNA Networks, a start-up company based in Oregon, has launched what it claims to be the first high-performance networked virtual memory system.
RNA on Monday launched its first two products: the Memory Virtualization Platform (MVP) and RNAmessenger. MVP, which is the overall controlling software, works by pooling, or aggregating, the available memory across multiple systems on a network. All its systems then become a shared pool of a network's resources available to all servers in the datacentre, RNA said in a statement.
The servers on the network can then access the pool as they require resources, or contribute to the pool, or both, the company said. By pooling resources in this way, organisations should be able to improve the utilisation of their datacentres.
RNA said its Memory Virtualization allows applications to take advantage of the shared memory pool without requiring any changes to existing applications. With this capability, organisations can potentially use a very large amount of memory to improve performance and increase efficiency.
RNAmessenger is the company's first product aimed at online applications, such as gaming and trading, that have high-volumes and low latency; or in other words, applications that process many small transactions very quickly, the company said.
"In early customer evaluations, we've repeatedly seen how frustrated enterprises are with the limitations of the current memory 'island' model," said Clive Cook, chief executive of RNA, in a statement.
Cook explained the benefits of his company's approach in a blog post on the VMBlog.com site. "The application's working data set is frequently larger than the available physical memory in the server," he wrote. "Today's single-server memory capacity ranges from 1GB to greater than 64GB. Memory is a captive resource to the CPU it is connected to, yet the working data set of many applications is well beyond this."
"Using Memory Virtualization, the entire working data set can be loaded into memory for the processor to access directly, without going to disk," Cook added.
Passing messages over the network instead of to local memory would increase latency but, the company told ZDNet UK: "The comparison is better made using the MVP memory pool compared to storage or adding local memory. The system resource constraint is the available memory on each server."
Cook said it is expensive to scale software for individual servers and "even when you do, it is captive to that server and not a shared resource". The problem then, he said, was that the data set is larger than local memory, and "in most cases this is the situation, you need to swap data from storage to memory". So with MVP, "you keep the working data set in memory, significantly reducing latency", Cook said.
If the software does not need to go to the disk so often, it should speed things up. Cook claims his software will give a 10x to 30x performance increase.
According to the company, RNAmessenger software supports 32-bit or 64-bit hardware and runs on IBM's Power, and Intel Itanium, x86, x64 processors. It will cost from $7,500 (£5,200) per system, depending on configuration. This is comparable to the cost of around 128GB of physical RAM, but the company told ZDNet UK: "From a cost perspective, by pooling memory you take advantage of the resources already installed. Large blocks of memory bought standalone can in some cases exceed the cost of a low-end server, especially when you are getting into the 64/128GB range."