Riverbed's Steelhead swims through WAN bottlenecks

WAN accelerator eliminates the wait for inter-office file transfers

IT managers are always being asked to do more with less. When it comes to WAN links, less is sometimes all you have to work with. The Steelhead 2000 WAN acceleration appliance from Riverbed Technology is part file cache, part proxy, and part TCP optimization, providing a unique and extremely effective way of reducing the time spent transferring files from one office to another. In addition, the Steelhead is easy to install, and it does not require any LAN/WAN re-engineering. But like other WAN acceleration appliances, a unit must be installed on each end of your WAN link.

Steelhead sits transparently on the LAN between your firewall and your servers. As files and other TCP traffic pass through it, Steelhead automatically analyzes the stream. A technology developed by Riverbed called Scalable Data Referencing breaks the data into indexed chunks and stores it in a local disk cache. The power of Steelhead kicks in when you pass the same or similar traffic back through the appliances over the WAN. Steelhead recognizes the patterns, finds matches in the data store, and sends only a reference pointer for each chunk of data to the other side. There, the reference pointer is used to rebuild the traffic on the fly so that the TCP stream arrives intact. By sending only a small pointer and not the whole file, the appliance greatly increases perceived performance over the link and dramatically reduces the time spent waiting for the data.

Silent Runner

Installing the Steelhead is a no-brainer. Other than setting the IP addresses for the network interfaces through the unit’s browser-based GUI, there is not much to configure. In fact, Steelhead pairs will automatically detect each other over the network. By default, Steelhead will try to optimize all inbound and outbound TCP traffic, but you can select types of traffic to ignore. It also supports layer 4 switches and custom QoS mappings, allowing you to take advantage of your current traffic management policies.

Depending on the model, a Steelhead appliance can have anywhere from 40GB to 500GB of local disk space for caching segments. The Steelhead 2000 I tested came with 135GB of disk space and boasted a maximum throughput rating of 3Mbps. There is no practical limit to the size of the cache — in a perfect world your disk cache would be infinitely large so that you would never have to purge any stored data. Data stored in the cache persists even when the appliance is turned off.

Steelhead uses a dual-port Ethernet card. If the Steelhead software fails, or the unit suffers a hardware or power failure, the two ports will short between them, allowing network traffic to pass unimpeded. You can have two Steelheads in line with each other, with one configured as a slave. The slave unit will automatically take over should the master fail, but WAN users might experience a brief performance hit while the slave builds its segment cache from scratch. Riverbed currently provides no way to synchronize masters and slaves, but future transparent prepopulation is planned.

File caching  systems from companies such as Extreme Networks and Network Appliance have been around for quite some time. Where Steelhead improves on traditional file caches is in its use of indexed file segments. For example, most file caches can reduce the amount of traffic traveling over a WAN link only when the files requested have been cached on both sides and, typically, only if the files have not been changed or renamed. Frequent changes to files can significantly limit their effectiveness. 

Steelhead, on the other hand, does not care what the file is named, how or when it was updated, or how it was transferred across the WAN. Because the appliance does not cache “files” as such, but chunks of data it has plucked from traffic streams, it will always locate some segments of an already-requested file in its cache — eliminating round trips for entire files that have been only slightly changed or renamed. In fact, Steelhead ignores what application created the chunk. If a requested Word document contains a piece of data originally found in an Excel sheet, Steelhead will still use the data, thereby decreasing access time.

I tested this feature extensively in my lab to see just how effective Riverbed’s acceleration would be. For my WAN circuit, I used a Pentium III PC running FreeBSD and equipped with two 10/100Mbps network adapters. FreeBSD has a built-in service called DummyNet (part of IPchains) that allows you to set link speed, packet loss, and latency between two network interfaces. I used DummyNet as my WAN circuit engine.

More Files, Less Filling

Using various file types and transmission methods (file copy, FTP, HTTP, and Microsoft Outlook), I saw huge time savings when using Steelhead in-line with my WAN circuit. For example, in one test I simulated a T1 satellite WAN link by defining the circuit as 1.536Mbps with 3 percent packet loss and 800ms average latency (400ms in each direction). Over the raw link, it took more than 25 minutes to FTP an 8MB PDF document. With Steelhead in place and “hot” (meaning the data had passed through the cache at least twice before), it took only 12 seconds to FTP the same file.

To make sure Steelhead was smarter than a standard file cache, I renamed the file and transferred it again. Again the file was retrieved in roughly 12 seconds. Next, instead of using FTP to send the file, I simply copied it using Windows Explorer and saw the same results. In all of my tests, no matter how I tried to fool it, either through renaming the file or changing the transport, Steelhead always correctly identified the traffic and optimized it for a huge time savings.

Steelhead also improves WAN performance by reducing the amount of network chatter (the overhead data transmitted between systems simply to facilitate communications) caused by MAPI (Messaging API) and CIFS clients. Microsoft Exchange is very chatty, and the process of copying a file creates an enormous amount of network overhead. Steelhead understands these two protocols and intercepts the chat requests. Instead of sending all of the chats across the WAN, smaller references are sent over the WAN that tell the other Steelhead how to recreate the entire conversation. To the application, all of the conversations take place, preserving any user permissions or other dependencies. This substantially reduces WAN traffic but still maintains the end-to-end conversation. Riverbed said it will release support for other protocols over time.

To say I was impressed with the Steelhead would be an understatement. I have tested other file caching systems, and none have come close to delivering the overall time reduction provided by this product. Furthermore, Steelhead is self-tuning and virtually maintenance free. A true appliance, this box is as close to a hands-off device as you will find.

InfoWorld Scorecard
Setup (10.0%)
Interoperability (15.0%)
Reliability (15.0%)
Performance (50.0%)
Value (10.0%)
Overall Score (100%)
Steelhead 2000 9.0 9.0 9.0 9.0 9.0 9.0