TECHNOLOGY FOR CHANGE

Memory Virtualization Adds a New Layer of Efficiency

Ericka Chickowski | Date: 07-06-09 | Comments

One Portland startup hopes to improve the way enterprises manage server memory through pooled resources.

In the quest to wring out greater efficiencies from the IT infrastructure, many organizations have found relief in a number of different virtualization technologies: on the endpoint, in the storage infrastructure and at the server level. But even with the advantages offered by the most common virtualization techniques, the server utilization rate of the typical enterprise is still limited by memory constraints.

�It�s a well-known, although sometimes little-discussed problem,� says Andy Mallinger, vice president of marketing for Portland, Ore.-based RNA Networks. �Server virtualization certainly helps with server consolidation and reduces some degree of server sprawl, but you can't get your mission-critical applications on a virtualized server. There are just too many constraints. If you think about it, if you take a server and you put 10 instances of the virtual machine on there and that server has 4 or 8 gigabytes of memory, you have to over-commit that memory to all of those virtual instances and you really run into performance problems.�

As Mallinger puts it, many times servers are swapped out of data centers not because they�re in need of a processing boost or a bump in storage, but because they just don�t have memory parity, especially as server virtualization improves processor utilization.

Enter RNA, which seeks to add memory virtualization as yet another subset of virtualization that can give organizations a boost in performance, improved power efficiency and better overall utilization rates.

�We believe that this is part of the very big trend of virtualization of servers, storage virtualization. There has been a lot of innovation and a lot of benefits driven out of those two types of virtualization, but memory has been left out of the party, so to speak,� Mallinger says.

The first of its kind in the field of memory virtualization, RNA�s products work by separating out the physical resource of memory within individual servers, pooling them together logically and then making that pool available to all of the servers so that the data sets they can hold are many times larger than what they could if they only depended on their individual memory resources.

�What I see, especially in the market that I support, which includes financial markets and high-performance computing, is that for certain classes of problems the amount of memory that you have in an individual commodity server is no longer sufficient,� says John Barr, research director for The 451 Group. �And so having some way of presenting a larger amount of memory across multiple systems as a single resource for a program running on one is an excellent idea.�

Initially conceived several years ago by a number of veterans in the high-performance computing field, RNA came out of stealth mode in February 2009. According to Mallinger, the company has a handful of customers, with still more in the pilot phase. The company is backed by $7 million in venture capital from Menlo Ventures and has already garnered interest from analysts such as Barr and Nik Simpson, senior analyst for Burton Group.

Simpson says he believes RNA�s biggest market opportunities are for organizations that maintain very large Internet database stores that have to be "sliced and diced and searched in multiple different ways."

�At the moment what [organizations with these stores] do is distribute that over many small servers and keep a small portion of the database locked in memory on each server,� he says. �What this would enable them to do is to use one larger server to do the actual slicing and dicing and to store the database engine or the Website environment in memory. The idea is that you're going to be able save power and save cost and improve performance when you're working with large data sets like that.�

RNA technology also provides relief for large financial organizations trying to optimize their financial messaging systems, Barr says.

One of RNA�s first customers, a large hedge fund that chose not to be disclosed for competitive reasons, reported that it experienced a 700 percent improvement in messaging volume using RNA technology.

�They were sending about 6,300 transactions per second when they started, they had a goal of getting up to about 10,000 per second; they would have been thrilled with that,� Mallinger says. �In the end, we reached well over 50,000 transactions per second.�

While the numbers look impressive in specific scenarios, such as financial markets, cloud computing scenarios and in industries doing complicated modeling such as oil exploration, both Barr and Simpson warn that RNA is not necessarily a general-purpose technology.

�For classes of application where you're reading a lot of data, it�s fantastic,� Barr says. �But for classes of application where you're doing as much writing as reading, then the performance penalty of writing data to systems which use memory on another system becomes an issue.�

Additionally, Simpson wonders whether the promise of flash memory in solid-state disks may eventually make this approach obsolete.

�The question is whether technologies like flash memories in the form of SSDs or in other form factors actually inside the host rather than being shared across a network might not [be] the better solution in the long run,� he says. �If you interface in such a way directly onto the motherboard or onto the server, then your latency is probably going to be less than going out over the network to use memory on another server.�

In the meantime, though, for the right situations, both analysts agree that the approach is worth a look.