December 18, 2007

Identifying Performance Factors of Home Gigabit Networks

Filed under: Uncategorized — Anthony Park @ 12:41 pm

Ed Bukoski, Todd Ellermann, and I recently conducted an experiment to analyze how to get the best performance out of a home gigabit network. We used design of experiments (DOX) techniques to design and conduct the experiment, and held several factors constant to limit the combinations.

We were primarily interested in things which would beneficial to know before INSTALLING a home gigabit network, and held constant things which are put in place after the fact (ie. network cards, computers, software, etc.).

We analyzed the following factors:

1. Network cable length (10ft vs. 100ft)
2. Network cable wiring type (solid vs. stranded)
3. Network cable category standard (cat 5e vs. cat 6)
4. Network switch price level
5. Electro-Magnetic Interference (EMI) from a light switch
6. Electro-Magnetic Interference (EMI) from a power line

We set up our equipment in a large conference room, and ran 192 completely randomized scenarios.

Conclusions

None of these factors had a significant effect on network throughput performance. If factors are chosen to minimize and maximize network throughput based on the model equation generated in this experiment, the results would be as follows:

This experiment was especially interesting to me because I was considering whether or not to rewire my house with CAT 6 cable instead of the CAT 5E that is in place now. After these results, I’m not going to bother.

The full experiment description, design, data, analysis, and conclusions can be downloaded in this report.

tags: Media Center, Gigabit Networking, MVP

6 Comments

  1. What about jumbo frames? I found in the design of my home’s gigabit network that no single factor BUT Jumbo Frames mattered. Machines running 9k frame sizes could easily push 1000+ megabits.

    Comment by Scott Hanselman — December 18, 2007 @ 10:55 pm

  2. Jumbo frames is an example of one of the many configuration settings we did not look at. The idea with this experiment was to look at the base configuration of a home gigabit network, and see if any of these “hard to change” factors had a significant impact on network throughput.

    The next step would be to identify what types of traffic are expected to be transmitted across the network and tailor the configuration to the specific needs of your network. Things like MTU size (jumbo frames), QoS, and so on may work well for some types of data transfer but not others.

    I am trying to figure out how to study these next, but they are very specific to the type of traffic on the network.

    Comment by Anthony Park — December 18, 2007 @ 11:24 pm

  3. Perhaps I’m exposing my ignorance, but why is cat 6 in the “worst case” column, and cat 5e under “best case”?

    Comment by Ian Horwill — December 19, 2007 @ 5:50 am

  4. Two of the factors came out to be “counter-intuitive”. CAT 5E came out better than CAT 6, and 100ft cable came out better than 10ft cable.

    On page 13 of the full paper, we list the model equation generated from the analysis of the data in this experiment. The equation is as follows:

    Throughput = 800.29 + 3.88*A – 3.95*C + 0.58*E + 2.42*F + 4.67*E*F

    As you can see from this equation, all of the factors had a VERY MINOR impact on throughput performance. Factor C (cable category standard) has a value of -1 for CAT 5E and a value of +1 for CAT 6 cable. This means that changing from CAT 6 to CAT 5E produced an 7.9Mb (3.95*2) gain in throughput performance. This is such a small value that it is probably due to either experimental error or possibly better construction of the CAT 5E cable used in this experiment. The bottom line is this: cable category type didn’t really make a difference in network throughput performance in these experiments.

    Comment by Anthony Park — December 19, 2007 @ 8:31 am

  5. Only problem I see with this test scenario is that by holding the things constant that you did, you quite possible didn’t maximize your variables. For example by using jumbo frames and pushing the throughput of your network up alot, you probably would then notice a bigger difference between cat5e cat6 etc. IMO you need to first maximize the throughput settings of the constants (jumbo frames etc), and then hold them costant at this maximized configuration for your tests of the variables to be accurate.

    Comment by bpatters — December 19, 2007 @ 9:40 am

  6. I just realized that this part was not mentioned in the paper… We did modify settings to improve the initial performance of the throughput. We turned on jumbo frames, but this didn’t seem to make a difference in throughput performance with the test tool we were using.

    We used clean builds of Windows XP on identical machines, and installed all the latest patches and drivers. We installed the measurement tool, and ran loopback tests on the two machines to determine if there were any hindrences to performance. During this process, we found that turning off the Windows Firewall service provided an additional 20% network performance gain. Note: the Windows Firewall service was STOPPED – turning off the firewall in the control panel did nothing for performance. We continued to modify various settings until the loopback performance on the two machines was as high as possible (slightly above 1GB).

    Once the loopback tests were returning the highest possible throughput results, the experiment was conducted. All of the test scenarios were run in randomized order in one session (about 3 hours), and the initial settings were not changed once the experiment began.

    Comment by Anthony Park — December 19, 2007 @ 11:40 am

RSS feed for comments on this post.

Sorry, the comment form is closed at this time.