Network and Telecom Strategies Blog

Final Catalyst WAN Performance Optimization Notes

Posted by: Eric Siegel

OK, two final notes on WAN Performance Optimization from the Catalyst Conference.

(By the way, I'll be rewriting my 90-page, hyper-boring paper "Optimizing WAN Performance" starting in a week or so; it should be published in a couple of months. Ask me if you need a draft copy sooner than that. And post a reply to this blog entry if you have any suggestions for me or for anyone about WAN performance optimization!)

FIRST, be sure to test using realistic user scenarios. Don't just rely on vendor promises or magazine articles or your buddy's experiences or some simple "drag and drop across the Windows desktop" copies! I've been saying this for years, but it really bears repetition.

The classical example is assuming that all vendors handle Microsoft's remote file access protocol, CIFS (or its latest Vista version, SMB 2.0) in the same way. They don't. And there's a difference between string-based compression (e.g., Riverbed) and Wide Area File Services (WAFS)-based compression (e.g., Cisco WAAS and Expand Networks). If you just do a sequential copy (e.g., drag-and-drop across the Windows desktop from a remote file share to your local disk), all of the CIFS systems look good. BUT if you open a file from INSIDE Microsoft Word, Excel, or some other application (is that what your users do?), then you'll see major differences. An application doesn't usually do only sequential reads and writes; it'll do "random" reads and writes within the file, and it'll create temporary files. So the intelligence of the WAN performance optimization solution becomes relevant: it can greatly improve performance when using remote file shares, possibly using the inherent caching ability of CIFS (i.e., the "oplock" capabilities), but it'll also be more complex and vulnerable to possible failure if the programming is sloppy.

SECOND, don't underestimate the impact of Web caching. Even major vendors, such as Microsoft, don't set their cache tags for performance. You can make simply astounding improvements in performance by setting Web cache tags appropriately, and your application front end (AFE) or load balancer can even do that for you. Be sure to include an "Expires" header, or most caches will not cache your pages, or they'll have to check back with the server to check the expiration status of every single object (unless you're using a specialized Web accelerator). See my Burton Group paper "Optimizing WAN Performance" and http://www.mnot.net/cacheability/, http://www.squid-cache.org/, http://www.web-caching.com/, and http://webmaster.info.aol.com/caching.html for more information. (Also look at http://developer.yahoo.com/performance/rules.html, which Nick Fiekowsky recommends, for more caching hints and general performance suggestions for the Web.)

Building Applications for the WAN

Posted by: Eric Siegel

Our final session at Catalyst this year was a joint presentation between Network and Telecom Strategies and Richard Watson of Burton Group's Application Platform Strategies group. We talked about how programmers and network engineers can work together to improve application performance and problem diagnosis.

This session was based on the NTS report "Designing Applications for the WAN," which will be updated next week -- so go read it when the update comes out (it's out now!), but there are a two points that came out in the session that I think were rather surprising to us network folks and that should be emphasized here.

First, the applications developers are now using frameworks to enable them to develop applications more quickly. Those application frameworks completely hide ("abstract") the network. For example, an application developer will ask a "method" for a "data object." The developer doesn't even realize that the network is involved, or that there may be some SQL queries used by that method to create the data object. That's why when you have a crisis, and you put Wireshark or a Sniffer on the communications link and show the resulting trace to the developer, he or she has no idea what the heck is going on. I'll say it again: the developer didn't even realize an SQL query was generated by the method request, and has no idea whatsoever what that stuff in the query string, trapped by the trace, really means.

No wonder there's frustration and wasted time in diagnosis! The time has arrived for the network organization, the operations group, and the development organization to cooperate and purchase a diagnostic tool that will trace individual transactions through the production system for diagnosis. (An example is Quest Software's PerformaSure; there are others.) The programmers will understand what they're seeing, the network staff won't need to run around doing Wireshark or Sniffer traces for days on end, and the problem will be handled much more quickly.

And, as the second point, realize that performance is poor because of the "abstraction" of the network layer. The application framework has a lot of parameters that set timeouts, window sizes, buffer sizes, compression, whether to disable the Nagle Algorithm, etc. The programmers understand none of this, so they use the defaults. But the defaults aren't usually set by the vendor to ensure optimum performance; they're set to create the fewest trouble calls to the vendor regardless of the network characteristics. That may not be the optimum setting for your particular situation, and that's why the people who install and set up the framework should talk to the network staff to jointly decide how to set the networking parameters for the networks that the framework will be running over. That'll help quite a bit in creating good performance.

TCP stack Performance Tuning for Windows

Posted by: Eric Siegel

I held a Birds of a Feather (BoF) session on WAN performance optimization at Burton Group's Catalyst Conference this week in San Diego, and we had approximately 14 attendees, which was quite good. (Attendance at Catalyst is almost exactly the same as last year: approximately 1500 people! We're VERY happy.)

We started the BoF with a talk by Nick Fiekowsky of Wyeth; he also has a talk coming up on WAN optimization, but the BoF was a great opportunity for us to have a more unstructured discussion (it lasted over an hour) about some of the detailed technical issues.

Nick made some great new observations about TCP performance on Windows; they're very interesting and also seem to have a large impact on performance. So here, with his permission, are some of them:

First, as Nick says, "Latency doesn't end at the RJ-45." When you're computing the required buffer size (receive window, RWIN) for long fat networks (i.e., large bandwidth and large latency), we always think of just the latency on the path and the bandwidth. I hadn't previously thought about the time that the received packet spends inside the receiving station's TCP buffers, waiting to be processed and then waiting to be read by the application. But that's an important factor in the latency; after all, the RWIN can open until there's space in the receive buffer, and the receive buffer won't have space until the application has read something out of it!

So Nick looked at what actually happened, and he discovered that although the bandwidth delay product (and therefore the receive buffer size) was approximately half a megabyte for a particular path, there were still long pauses waiting for acknowledgements to be returned to the sender -- throughput was lower than it could be. He had to increase the receive buffer size to almost THREE MEGABYTES before the sniffer (actually, Wireshark) trace showed that the bandwidth was being fully utilized. Wow.

And that led to Nick's other major point: it's important to optimize Windows performance to keep it from stalling while processing those incoming packets:

• "Non-Paged Pool" (NPP), where the buffers are located, is a precious resource. Nick says that in a 32-bit Windows OS, the NPP = RAM/16. That's not very large, especially if you need large receive buffers. (Look at the Task Manager to see how much of the NPP is being used. Or you can just notice that the TCP stack is requesting a small RWIN when a new connection comes in, because it's under memory pressure.) The only way to increase that is to increase the total RAM in the processor. In the 64-bit Windows OS, the NPP allocation is handled much better, TCP auto-tuning in Vista, Windows 7 and Windows Server 2008 makes better use of NPP, and finally the 64-bit OS also makes the entire system run better because system processes are faster. So the 64-bit OS is a big improvement here.
• Defrag the hard drive to be sure that fragmentation isn't slowing down the processes and the stack.
• Use updated NIC drivers. Nick says that recently he's noticed that there's a new emphasis on performance in the drivers, possibly being driven by iSCSI and datacenter requirements. Also increase the number of transmit and receive descriptors in the driver configuration.

Nick had a lot of other recommendations, such as the big improvement made by ensuring that Web cache expiration tags are properly used, and I'll try to get to them in a later blog.

Meanwhile, thanks, Nick!

Backup for WAN performance optimizers at remote offices

Posted by: Eric Siegel

(This is the first of a few blog entries that I'm going to post to share some ideas that came out of the Birds of a Feather (BoF) session on WAN performance optimization at Burton Group's Catalyst Conference this week in San Diego.)

A primary advantage of WAN performance optimization appliances is that they cut down the bandwidth needs of a remote branch office -- usually by 50% or more. And they reduce response time to make remote file shares, etc., actually usable instead of being a miserable, irritating experience that destroys productivity.

But what happens when the optimizer breaks? Bandwidth requirements spike upwards, and response time instantly becomes lousy.

My suggestions in the past involved QoS in the branch office boundary router to ensure that critical traffic would still flow despite the worsened bandwidth restrictions, and recommendations that the enterprise make arrangements with vendors or resellers to have spares available in-country (to avoid delays in customs and in transit) and configuration information available on a USB flash drive or by downloading as an email attachment. There would be a day or so of worse-than-normal performance as the compression directories are rebuilt, and then things would be OK.

But Nick Fiekowsky, while sharing his slides at the BoF, had a great idea: use the new software client optimizers, such as Blue Coat's Proxy Client, Cisco's WAAS Mobile, Expand Network's Mobile Accelerator Client, and Riverbed's Steelhead Mobile to serve as emergency backups. These software optimizers run on the users' laptops and provide the function of the optimization appliance. They're not as good, because (except for the new Expand Networks Mobile Accelerator Client! More later...) they don't share their compression directories and caches. But they're a lot better than nothing, and they also provide the ability for the user to use acceleration from home or on the road. This is especially great if the software optimizer system licenses are inexpensive or can be turned on and off easily from the central server location, with billing only when they're used.

My updated recommendation is therefore to install the WAN accelerator vendor's mobile client in branch office desktops. If the office's WAN accelerator fails, the desktop clients will step in to reduce WAN traffic and provide an acceptable user experience. It is necessary to have enough license capacity to cover all the client devices that may become active when a single branch office's WAN accelerator fails.

Closing in on Catalyst

Posted by: Michael Disabato

The preparation for Catalyst includes reviewing all the presentations that will be given. While this is a lot of work (NTS reviews 18 presentations in our track and about 20 more in other tracks), We get an excellent sense of where the conference will be headed. We have consistently improved Catalyst through one simple tactic: We listen to our attendees. We hear you when you talk about the things that keep you up at night. We look through the corridor of time to the future and see what will be exciting to you when it's ready for prime time. We present that in a forum where you can meet and greet your peers and the presenters.

If you have not made plans to attend Catalyst 2009, head over to the Burton Group web site, take a look at the agenda, and make your reservations today. We are indeed covering the right issues with the right speakers, and this is the right time to hear what they have to say.

Michael