SI
SI
discoversearch

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor.  We ask that you disable ad blocking while on Silicon Investor in the best interests of our community.  If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.
Technology Stocks : Network Appliance
NTAP 109.28-2.0%Nov 26 3:59 PM EST

 Public ReplyPrvt ReplyMark as Last ReadFilePrevious 10Next 10PreviousNext  
To: riposte who wrote (10062)4/23/2002 1:12:57 PM
From: riposte  Read Replies (1) of 10934
 
How DAFS will change networked storage

storagemagazine.techtarget.com

Another interesting one from Storage Magazine.

How DAFS will change networked storage

by: Marc Farley
Issue: Apr 2002
One of the most interesting storage technologies expected to be introduced in 2002 is DAFS, or the Direct Access File System. The point of DAFS is to bring network file systems up to par with modern, high-speed data networks. The result could be as much as a 100% performance improvement over NFS. And with proper vendor support, it could be a big step in enabling database use on network-attached storage (NAS) boxes.

In a perfect world with gigabit-speed high-reliability networks, there is no reason to believe that a database application would not work with the appropriate-level NAS solution. As with many technologies, don't confuse implementation with the architecture. NFS and CIFS (the Windows equivalent) can be thought of merely as implementations of a generic network file storage system. In the same way DAFS can also be thought of as belonging to the same family of network file storage technologies, but with some significant evolutionary differences.

Most of the development work on DAFS was done by Network Appliance, which has made an effort to enlist others in the storage industry to participate in this work. Network Appliance has submitted the technology to the Internet Engineering Task Force (IETF) as a potential standard. Furthermore, the company has helped create an implementer's forum within the Storage Networking Industry Association (SNIA). For a storage company, that is fairly noble behavior, considering the trend in the industry to keep software innovations proprietary.

Not everyone in the storage industry is a DAFS advocate. The skeptical opinion of DAFS from most other NAS and subsystem companies is that it is not really an open standard, but an attempt by Network Appliance to grab a share of the storage market in data center transaction processing.

But many analysts believe that DAFS is a genuine evolution in NAS technology that could deliver significant benefits to users. While it is not at all clear that DAFS will succeed, there appears to be very little keeping other companies from adopting it. In fact, Network Appliance's competitors could take advantage of the groundwork and resources that Network Appliance has already spent developing the technology.

DAFS as an evolution of NAS
To understand DAFS' potential, one needs an open mind about the feasibility of NAS products supporting high throughput I/O for database and transaction processing systems. Historically, there have been legitimate concerns about the performance and reliability of NAS for database storage. The initial concept of NAS introduced by Sun as Network File System (NFS) in the mid-1980s assumed that networks were not reliable and therefore placed the burden of error recovery on the client and server systems involved. This approach followed the example of TCP/IP network technologies, which were developed with the assumptions that network resources were constrained and CPU cycles were plentiful.

But the relationship between networking bandwidth and processor capabilities has been reversed and network bandwidth is more plentiful than the CPU resources needed to operate the network. If we were to develop networking protocols with today's assumptions about network capabilities instead of using assumptions from 25 years ago, there is no way network protocol stacks would be processed the same way.

For example, today we are faced with overcoming the latency involved in processing the TCP/IP protocol stacks. Companies developing iSCSI storage technologies are dealing with that challenge - matching bandwidth and CPU resources - through the use of TCP offload engine technology (see "TCP offload engines finally arrive," March 2002).

DAFS is a different approach to the same problem. Instead of implementing protocol processing algorithms in specialized network processors, DAFS implements different networking protocols that assume the network is much more reliable and much better performance legacy Ethernet/TCP/IP networks. Just as legacy NAS differs from DAS and storage area networks (SANs) by implementing network filing protocols as opposed to block-based storage protocols, DAFS differs from legacy NAS by implementing a new network protocol stack.

Network assumptions with DAFS
Instead of assuming any type of network could be used, DAFS assumes Gbit transmission speeds and extremely low error rates. Starting with this new set of assumptions, the designers of DAFS had the advantage over their NFS and CIFS counterparts because they didn't have to build reliability into DAFS itself - instead it is provided by the network. DAFS is built on top of the Virtual Interface (VI) architecture designed by Microsoft, Compaq, and Intel several years ago as a high-speed, low latency network protocol. VI provides a way to perform remote direct memory access (RDMA) between applications running on different machines in a LAN or cluster. RDMA allows applications to read and write to memory addresses used by other applications in the network. As long as the applications agree on how to access and place data within each other's memory control, they can circumvent the system processes used to exchange data between systems on a network, dramatically streamlining the transaction.

Metaphorically, DAFS is a limousine in a world that has made a habit of taking buses and trains. In order to ride a bus or a train you have to work within the constraints of the schedules and routes provided by the mass transit system. The system of schedules and routes provided by the transit system are its protocol. By contrast, the transportation protocol you use when you drive an automobile is much different. You go where you want to, when you want, as long as you obey the traffic laws along the way.

NFS and CIFS are analogous to mass transit riders by virtue of their using existing shared facilities to transport storage data. The general-purpose algorithms and processes shared by all other applications have a detrimental impact on the performance of NFS and CIFS. DAFS, however, uses RDMA processes that are unique, dedicated and optimized for access to file storage. In other words, DAFS file requests to open, update, append, lock, or close are exchanged between storage clients and storage servers using RDMA.

Expected performance benefits of DAFS
Theoretically, this can result in enormous performance advantages. Researchers at Duke and Harvard Universities have analyzed the potential performance advantages of DAFS and have concluded that it provides, at a minimum, 100% better performance than NFS storage. Their conclusions state that they believe DAFS performance improvements would actually be greater than this due to the fact they were running state-of-the-art NFS implementations unlikely to be found in most commercial data processing environments.

In addition to leveraging VI and RDMA, DAFS includes additional features such as the ability for applications to send cache "hints" to the server to pre-fetch data it expects to need in the future. While cache hints are mostly a theoretical concept that will require careful development and testing work throughout the industry, the potential performance benefits of cache hints could make an astounding impact on I/O throughput. This idea of intelligent caching and cache hints is a by-product of using NAS file based storage, as opposed to block-based storage.

Opening the database storage market to NAS
If the point of DAFS is to let applications communicate directly with network file systems using RDMA, then it is obvious that some pretty big software applications are going to have to support DAFS before the technology will do anybody any good.

The big story that appears to be brewing with DAFS is whether or not it succeeds in being implemented with data center level database applications, particularly Oracle. While databases have traditionally used block-based storage, there could be significant advantages if they supported file-level interfaces. In essence, DAFS provides a fast, low latency and reliable file-based storage connection. Database data stored in a DAFS server's file system can then be managed using several types of data management products. In addition, by virtue of being a filing technology, DAFS has the potential to offload some of the internal storage processing that databases currently have to do.

In early December 2001, Network Appliance made a major effort to show their presence at the Oracle Open World exhibit in San Francisco. But more than the joint marketing image they presented with Oracle, they also had DAFS demoware showing Oracle databases using DAFS to read and write data on DAFS-supporting filers.

Of course, there are big difference between demos and real shipping products. It is clear that Network Appliance needs the support of Oracle or some other major software vendor to help make DAFS a reality in the market. Considering the somewhat tense relationship that has existed at times between Network Appliance and Microsoft, Oracle seems like the most potent software ally Network Appliance is courting to help bring DAFS to the market.

DAFS has a lot of potential, but for the near term, its success rests more with Oracle than it does with Network Appliance.

Next month, we'll look at what you should expect to do if you want to use DAFS

Report TOU ViolationShare This Post
 Public ReplyPrvt ReplyMark as Last ReadFilePrevious 10Next 10PreviousNext