Source: Marc Staimer, founder, Dragon Slayer Consulting
TCP/IP wasn't designed to transport large volumes of stored data. The pervasive WAN protocol has no problem with small amounts of data that can be sliced and diced into tiny packets.
However, "TCP/IP doesn't handle congestion well. It does a lot of checking back and forth and resending, which just slows things way down," says Marc Staimer, founder, Dragon Slayer Consulting, Beaverton, OR.
As a result, when an organization tries to replicate large amounts of data between sites over, say, a T1 link (1.5 Mb/s) using TCP/IP the performance can't match what it gets when backing up that same data over the LAN at night. And if the organization wants to do anything else with that T1 link, forget it--the link is nearly saturated.
Today storage managers can choose among a growing number of appliances that promise to speed up the movement of large volumes of data over TCP/IP links. These products, network accelerators of various types (see "Network accelerators") allow organizations to transmit more data faster and use less bandwidth, which saves money and boosts performance. Some of the new appliances--dubbed remote network-attached storage (NAS) accelerators or WAN file servers--even promise to eliminate the need to maintain file storage at remote sites, allowing centralized NAS storage to meet the storage needs of remote workers as if they were local.
Finisar Corp., a manufacturer of optical telecommunications tools, figured that it needed two--or even three--T1 links between its Network Appliance Inc. (NetApp) NAS box at its Sunnyvale, CA, headquarters and a similar NetApp device at its facility in Malaysia to handle file replication and other data traffic between the sites. Faced with a communication investment running well into thousands of dollars a month, the company tested a network acceleration appliance from Peribit Networks, Santa Clara, CA. The company dropped a 100MB file on the NetApp box in CA, and using NetApp's SnapMirror function, fired it off to the NetApp filer in Malaysia.
"It took four minutes and 21 seconds to get it there compared to over nine minutes if we just used the T1," says Jon Hudson, Finisar's storage area network (SAN) architect. At that rate, the Peribit appliance, which compresses files, would give Finisar "a payback in weeks or a few months," he concludes.
A large Midwest shoe retailer faced a similar problem. "We were looking at having to buy a DS3 [45Mb/s] pipe to replicate data between our two data centers for disaster recovery," says the retailer's IT manager. The company maintains an AS/400 at each site and continuously replicates database transactions between them. In the event of a failure at either site, its hundreds of stores would be able to continue functioning and barely miss a beat.
Hoping to avoid the need for the costly pipe, the shoe retailer turned to NetCelera Networks, a company that provides a WAN acceleration appliance. The appliance terminates a TCP/IP connection and substitutes its own protocol to eliminate the TCP/IP overhead. In addition, it applies compression at Layer 5 of the ISO network stack. "We got 10:1 compression from NetCelera, which saved us from having to buy the DS3," says the manager.
NetCelera and Peribit are just two of a many vendors introducing products to boost the performance of TCP/IP for the transmission of large data sets over the WAN. Many of these products are basic network accelerators that compress traffic, effectively boosting performance of data transmission over the network or allowing the organization to use a smaller, less-costly communications link to achieve the same performance as before. Others don't use conventional compression at all. Instead, they transparently replace TCP/IP for the portion of the link between the primary and target site with an efficient proprietary protocol optimized for large data sets.
Even vendors just doing compression are achieving sizeable performance gains, typically 300%, says Peter Firstbrook, senior research analyst, at the Meta Group. Compression at the application level ordinarily slows down the server. The new products avoid this by moving the processing to an appliance. Other products look at the content of the data being transmitted and can achieve up to a tenfold performance gain with some content, he adds. For example, a product will look at a large data set, identify repeating patterns and assign a tiny token to represent a big chunk of data. The product then sends just the token--instead of the data--and reconstructs all of it on the other side from the token.
|How I-shared works|
A Tacit appliance in the data center intercepts TCP/IP and translates it into its proprietary protocol that works faster over the WAN. The appliance at the remote site also caches data and only sends changed files back to the data center, further speeding up WAN transmissions.
WAN file servers
The most radical of the WAN accelerators probably shouldn't be thought of as network accelerators. "We call them WAN file servers," says Arun Taneja, principal at the Taneja Group. Staimer categorizes them as remote NAS accelerators These products focus on the transmission of stored files and enable NFS to operate efficiently at long distances over a TCP/IP connection by taking TCP/IP out of the picture. "They substitute a new protocol and take out TCP/IP's error checking and latency. The result is almost as if everything were local at the data center," Taneja says.
WAN file servers enable an organization to radically re-architect the way it does storage at remote sites. Typically, a branch office might have a file server and disk array or a NAS appliance. Periodically, files would be replicated between the main site and the remote location. Additionally, someone at the remote site would back up the local files to tape, a task that often isn't done.
With a WAN file server appliance in the remote office and another at the central site--such as I-shared from Tacit Networks, South Plainfield, NJ--the organization can eliminate storage at the remote site altogether (see "How I-shared works"). Users access files over the WAN and store their work at the central site as if they were using local NAS storage. The appliance terminates the TCP/IP connection just before it leaves the site and inserts its own communications protocol and file system. When it hits the location on the other end, the appliance restores the TCP/IP connection and the native file system, such as NFS or the common Internet file system (CIFS). The process is transparent to the applications and users on each end. To them, it looks like NAS going out over TCP/IP, but the performance rivals local NAS.
Tacit doesn't eliminate TCP/IP; its I-shared appliance simply ignores it. Tacit uses the TCP/IP link, but terminates/restores TCP at each end. In between, Tacit is able to avoid the problems that slow down conventional TCP/IP network accelerators through so-called difference recognition, which allows it to transmit only changes (deltas) after the initial transmission. In addition, read/write caching holds files locally while allowing users to write as well as read and it reads ahead to anticipate which file the user will want next. Data streaming avoids repeated request-and-wait cycles. Finally, like the conventional accelerators, it also does compression.
"Basically, you are bringing everything back to the central site," says Taneja. "You install a black box on each end. There are no files at the remote site [except in the cache]. You can do this because the protocol is just so much more efficient. And you don't have to worry about protecting remote files because you back everything up centrally."
Although the idea sounds good, "remote NAS is gaining little traction," says Dragon Slayer's Staimer. Most companies are reluctant to install what they think is exotic and proprietary. Instead, they do what Finisar does, install a standard NAS appliance and use a conventional approach to network acceleration to improve performance.
Exotic or not, Anadarko Petroleum Corp., The Woodlands, TX, turned to Tacit to not only deliver data to remote locations, but also to run Windows applications over the WAN link as if they were local. It chose Tacit's product after experimenting with more conventional network accelerators from Peribit and Expand Networks.
Anadarko's problem is slow satellite links. "We get a realistic speed of 128Kb/s and a latency of 1.1 seconds," says Barry Fairbanks, Anadarko's network advisor. The company wanted a way to perform basic tasks such as copy a stored file, access e-mail and open a Word document at the distant oil-drilling platform. "Tacit had significant advantages when you wanted to open a non-local file or open any Windows file with our bootup," Fairbanks explains. Anadarko's bootup procedure involves considerable authentication and authorization checking with a central server. "Tacit lets all this happen fast," he says.
Anadarko uses Tacit and Expand products. "Tacit is best where you want to copy Windows files or you have to involve Active Directory," Fairbanks says. On the other hand, Tacit, which is optimized for file systems, does nothing for e-mail, which uses different protocols.
Rewards Network Inc., in Chicago, took another approach and opted for the Expand Networks solution to replicate data between data centers in New Jersey and Miami. The company, which runs a restaurant cash-back customer rewards program, uses a T1 link to replicate data in real time between an AS/400 in each data center. Before it implemented the Expand solution, replication was extremely slow and some transactions were lost.
"We considered upgrading the link, but that was very expensive," says Frank Del Campillo, Real Rewards' senior network architect. An upgraded communications link would cost the company an extra $5,000/month. The cost of two Expand appliances--one in each data center--ran about $20,000, the cost of just four months of higher bandwidth. Del Campillo says the company gained the equivalent of a tenfold increase in network capacity.
Generally, none of the appliances are terribly expensive; most start below $10,000, up to $45,000. An appliance needs to reside at each end of the connection, which means buying two appliances at a minimum, three to support two remote offices, etc.
All the vendors have made their products transparent to the storage system, servers, applications and even the network. A few vendors include management utilities, "but these appliances don't need much management. Basically, you just drop them in," says Peter Sevcik, president, NetForecast, Charlottesville, VA.
The advent of SANs introduced storage professionals to local networking. With distributed files, long-distance file replication and disaster replication, the storage group now must overcome the challenges of time, distance and the cost of using a WAN protocol not intended for storage. New accelerators promise to change that.