IT always seems to be caught in the middle of the WAN-performance battle: On one hand, users never seem to be happy with an application's performance; on the other, the bean counters won't budget for bigger pipes. If more bandwidth isn't the answer to end-users' performance problems, then what is?
The solution to improving both response time and circuit utilisation is to make the connection more efficient and reduce the effects of latency. I've watched a whole new breed of WAN optimisation and acceleration appliances come to market during the past few years, and I've seen first-hand what a difference intelligent optimisation can make to WAN response time.
Once again, I had Riverbed's Steelhead WAN acceleration and optimisation appliances on my test bed, and true to form, the company provided excellent improvements with little tweaking. Release 3 of the RiOS (Riverbed Optimisation System) shows a more mature product, with the focus on improving existing features, the addition of greater application support, and much increased capacity. During its time on my bench, Steelhead proved again that it is more than capable of making oversubscribed WAN circuits appear faster than they were, while greatly reducing response time, improving on previous releases by as much as 20 percent.
I tested three Steelhead appliances: the 200, the 5520, and the 6020. The 5520 (offering 155Mbit/s of optimised throughput and 15,000 TCP connections) and the 6020 (boasting 310Mbit/s of optimised throughput and 40,000 TCP connections) are hardware platforms released this year, as are their cousins, the 3020 and the 3520.
I set up the 5520 as the branch-office appliance and the 6020 as the datacentre unit. (The complete solution requires at least two appliances.) The impish 200 (1Mbit/s of optimised throughput and 100 TCP connections) came into play during testing of optimisation of asymmetric routes (more on this later). My trusty Shunra VE simulated my various WAN conditions, and D-Link DGS-3324SR Gigabit switches tied it all together.
RiOS has always included optimisations for CIFS and MAPI aimed at reducing those protocols' chattiness. In RiOS 3, Riverbed has added NFS streamlining. As with CIFS, there is a lot of NFS protocol overhead that can degrade application response time significantly. RiOS 2.1 performed NFS optimisation via data and transport streaming; RiOS 3 does NFS application streaming, allowing for read-ahead and metadata pre-fetching. Unfortunately, I do not have any NFS performance numbers from previous releases, but Riverbed's internal testing shows a 10-fold improvement in RiOS 3. CIFS traffic gained an improvement of between three and six times compared with last year's testing.
CIFS traffic also received additional application-specific acceleration for programs which open a file multiple times during a single operation, for example SolidWorks, Visio 2003, and SolidEdge. Previously, Steelhead would not optimise subsequent requests for the locked file, erring on the side of data integrity. The new CIFS mode, called overlapping opens, allows Steelhead to continue to optimise CIFS traffic even though Windows has the file locked, improving response times even under these conditions.
This release also does away with the need for a software agent to reside on a server in order to pre-populate an appliance's cache and eliminate the "cold hit" penalty. RiOS's pre-population service will load data from any CIFS file share, such as Windows, NetApp, and EMC, into cache on a recurring schedule or on demand. IT can still use the RCU (Riverbed Copy Utility), available in the previous release, but the new pre-population service makes it obsolete.
Quality and quantity
Enforcement available in the Steelhead appliance yields better support of QoS (quality of service). Previously, RiOS could set QoS tags but could not impose QoS policy on traffic. With Release 3, not only can it enforce policies, but it can do so based on bandwidth and latency. This means IT can define QoS policies based on a class of traffic's priority, and not just the bandwidth required. This is really important when protecting latency-sensitive traffic such as VoIP and streaming video. RiOS can apply QoS to both TCP and UDP protocols, and to optimised and passed-through traffic.
Not all WANs, intentionally or not, return packets via the same path, creating asymmetric routes. For a WAN optimisation appliance to function correctly, the devices at each end of the circuit have to pass the same packets in and out. With asymmetric routing, a packet may go out one appliance but return through a different path, breaking the optimisation process.
RiOS can correct client-side and server-side asymmetric routing issues by forwarding the misrouted packets to a second Steelhead appliance (also located at the client location) that is in-line with the original traffic. Although this does require at least two Steelheads and knowledge of the WANs traffic flows, admins can still create optimisation policies regardless of how packets move between sites.
Seeing is believing
One criticism of WAN optimisation appliances is that the packets are mangled by the optimisation process and traditional network monitoring tools are blind to the traffic. To help admins know how their WAN is functioning, Steelheads can now export traffic flow information to NetFlow collectors. I tested this using Scrutinizer from Plixer International and was impressed by how much information about my WAN was being captured and recorded. All information, such as source and destination ports and IP addresses, was preserved.
The reporting engine in RiOS 3 is better than the last time I tested, with small but useful changes to graphs and charts. I can now view charts at one-second intervals and can choose a custom time interval. For example, I created a bandwidth optimisation report starting at midnight of the current day to 18:00 hours. Previously I could only select hour, day, week, and month as time intervals. A few new reports, such as WAN interface throughput and connection pooling, are also built in.
Riverbed's Steelhead family comprises some of the best, easiest-to-deploy-and-maintain WAN optimisation appliances on the market. I like that I can set up my optimisation scheme with a few simple mouse clicks; there aren't myriad choices and settings to define. The newest release improves on scalability and even on its already impressive data-reduction numbers. The NetFlow support is a welcome addition as are the new graphs and charts. All in all, Steelhead is still the target most other vendors aim for.
Riverbed continues to impress with expanded CIFS and NFS support, allowing for better performance with both Windows- and NFS-based servers. QoS can now be enforced in the appliance with policies based on bandwidth and latency. Reporting also saw upgrades, as did the pre-population service to help eliminate cold hit penalties. Detection and correction of asymmetric routes is a definite plus.