Gigabit Ethernet burns up the wires
1 April 2005 | 0
Gigabit Ethernet is one of those technologies that sounds great on paper but ultimately is hard to justify, especially for desktop systems. Mention ‘Gigabit’ and ‘desktop’ in the same sentence and the first application that comes to mind is high-definition streaming media — a compelling use, but far from mainstream.
In fact, there are few examples of business-productivity benefits derived from deploying Gigabit Ethernet to the desktop. A pervasive Gigabit Ethernet infrastructure seems to be a luxury that IT departments simply cannot afford.
Or can they? The recent economic downturn has sent prices for many IT products tumbling, including Gigabit Ethernet adapters. You can now equip a PC with Gigabit-over-copper capability for under $50. And with Intel pushing OEMs to include Gigabit Ethernet components in its base PC designs, thereby displacing the ubiquitous Intel 82559 10/100Mbit/s connection, you will see an increased emphasis on connectivity performance as a selling point for enterprise desktop systems.
Yet the question remains, Is all this new silicon justified? Conventional wisdom, as defined by various ‘packet-blasting’ scenarios conducted using desktop Gigabit connections, says it is not. But our experience testing an all-Intel, Gigabit-over-copper system paints a different picture.
By taking the analysis up a level to the application layer, we were able to measure significant performance gains when compared to a traditional Fast Ethernet (10/100Mbit/s) deployment. And as you will see, our choice of applications was about as mainstream as you can get: client/server database, messaging, and business-productivity workloads running against Microsoft SQL Server, Microsoft Exchange Server, and a Windows 2000 IntelliMirror managed PC environment.
In an effort to weed out extraneous factors, we deliberately chose the most straightforward methodology we could employ given the tools at hand. Our first target was a transactional database scenario as employed by mainstream, client/server applications. The actual workload consisted of a single instance of our Database Performance Portal load-simulation object running against a Windows 2000 Server using Microsoft SQL Server 2000. By varying the table-size parameters, we were able to generate a broad set of test results spanning several different transaction types.
Similar to the Database Portal scenario, our second performance target involved client/server transactions against a Windows 2000-hosted resource. In this case it was a Microsoft Exchange 2000 mailbox store, and the tool employed was our MAPI (Messaging API)-based Workflow Performance Portal, which exercises the Exchange Server message store by conducting various message object creation, deletion, and copy transactions.
Once again, we varied the workload by changing the number of source-folder objects, creating a results table with transactions spanning approximately 25Mbyte of message and document content. When we factored in our Database Portal results we had nearly two-dozen transaction scenarios covering the most common client/server application execution models.
For our third and final scenario, we focused on desktop PC manageability. By comparing business-productivity application performance (using Microsoft Word, Excel, PowerPoint, and Internet Explorer) in a networkcentric, Windows 2000 IntelliMirror environment, we were able to evaluate Gigabit Ethernet’s performance impact on a highly managed desktop-computing scenario.
Of course, no discussion of methodology would be complete without a review of the hardware employed. Our client/server test beds consisted of 1.5GHz Pentium 4 and 733MHz Pentium III client PCs connected via switched CAT-5 cabling to a 733MHz dual-Pentium III server. Our 1Gbit/s backbone consisted of an Intel NetStructure 480T routing switch, whereas our 10/100Mbit/s backbone had a NetGear FS108 Fast Ethernet switch.
Active Directory PDC (primary domain controller) and Primary DNS (Domain Name System) functions were handled by a separate Dell PowerEdge server, while external Internet connectivity was provided by a FlowPoint 2200 SDSL (symmetric DSL) router.
Gigabit by the numbers
The first thing to notice about our transactional database performance results is the consistency of the performance curve. As we increased the number of records per transaction, the advantages of Gigabit Ethernet when compared to traditional Fast Ethernet became more and more apparent. Overall, the performance delta between 1Gbit/s and 100Mbit/s backbones averaged 41 per cent, with the most significant impact occurring during the most demanding workload scenario.
The results of our workflow tests, although not as dramatic, still generated a buzz in the lab. Overall, the performance delta across the five configuration scenarios averaged 22 per cent, with the best results achieved during the least demanding scenario.
At first glance, the results seem to contradict the performance curve from our database scenario. But further analysis reveals the true source of the bottleneck: Microsoft Exchange Server and its disk-bound transaction architecture.
In each case, the workloads started off with a very fast initial transaction time only to be quickly bogged down by disk write-related delays at the server. If we had been able to modify Exchange’s disk I/O behaviour — by queuing more lazy write data or checking for repetitive I/O operations, for example — we would no doubt have been able to achieve a higher sustained transaction throughput.
One area in which Gigabit Ethernet truly shined was in our managed PC scenario. With all application and user data running cached from a central server, the advantages of the Gigabit Ethernet connection became clear.
Our OfficeBench 3.0 test script, which uses Word, Excel, PowerPoint, and Internet Explorer to perform a set of common end-user tasks, ran as much as 67 per cent faster when executed via Gigabit Ethernet, with the Pentium III system reaping the greatest rewards.
Finally, there seems to be a direct correlation between desktop performance and Gigabit Ethernet return on investment: the faster the client PC, the greater the performance improvement. Migrating to a pervasive Gigabit environment is a desirable goal for any business-computing environment, but it makes even more sense as part of a broader desktop PC upgrade strategy.
With a transactional database throughput nearly 50 per cent higher than that of 100Mbit/s Ethernet, and with workflow application performance limited only by server I/O capacity, it is clear that Gigabit Ethernet can deliver tangible benefits to business desktop users. Although power users will reap the greatest rewards, even light-to moderate-use ‘knowledge workers’ will experience a measurable performance boost, especially when running in a highly managed PC environment.
THE BOTTOM LINE — Performance
Implementing Gigabit Ethernet on the desktop can bring clear performance benefits to business users. We identified the greatest gains in transactional database performance, but improvements in workflow and managed PC scenarios were also noteworthy.
Enterprise IT departments should seriously consider Gigabit Ethernet in their future infrastructure investments. Costs are dropping steadily, and because most Gigabit-over-copper products are backward-compatible with Fast Ethernet, there really is no reason to buy another 10/100Mbit/s component.
• IDG News Service