Brian's Brain

When Intellon saw my recent hands-on networking cover story, the company was quite surprised with the disparity between my TCP and UDP benchmarking results on the various generations of HomePlug adapters I tested, as well as being concerned that competitor DS2's adapters had outperformed HomePlug AV on TCP. After strategic marketing engineer David Yount and senior field applications engineer Nael Atallah familiarized themselves with Iperf as well as with how DS2 and HomePlug AV adapters performed under the utility, they drove up from San Jose, CA yesterday to pay me a visit at my Sacramento home office.

In my earlier article, I wrote:

Benchmarking projects such as this one are fraught with potential peril. As I noted in an earlier article, up-front assumptions can heavily influence outcomes. “If I select a combination of equipment, software, and usage-model variables that are too specific, my results would be meaningful to only a narrow set of readers,” I wrote. “Choose a too-broad set of options, on the other hand, and I end up with a bewildering plethora of outcome data.” These statements apply equally well to this report.

Those of you who've already read the article and its online-only addendums will, I think, agree that my study was variable-rich:

• Four different 802.11 wireless technologies, also encompassing both the 2.4 and 5 GHz 802.11n frequency bands, and two different 802.11n hardware platforms
• Five different powerline networking technologies, also encompassing TCP- and UDP-optimized adapter settings where available, and
• Both 100 Mbit and 1 Gbit wired Ethernet (did I forget anything? ;-) )

Not desiring to make the project even more complicated, either for me in doing it or for you in deciphering it, I settled on a number of duly documented assumptions, thereby resisting the temptation to do additional tinkering. For example, as the above screenshot shows, the default TCP buffer length (which I used) is 8 Kbytes. Not explicitly revealed in the Iperf GUI is the additional fact that when not otherwise specified, the TCP window size matches the TCP buffer length...therefore also 8 Kbytes for my initial testing.

Atallah explained to me (TCP-knowledgeable readers, feel free to comment on what I'm about to say) that in typical operating system and application environments in which two network clients negotiate parameters like buffer length and window size as part of the connection setup, there ends up being at least a 10x window size multiplier versus the buffer length. And in fact, when I explicitly set the window size to 80 Kbytes, Iperf measured greater than 40 Mbps of TCP bandwidth between the Dell desktop PC in my office and one of Yount's Dell XPS M1210 laptops in the 'swamp cooler closet'. Further tweaking of both the TCP buffer length and window size got us beyond 45 Mbps. Our meeting was brief, so there wasn't time for a lot of experimentation; it's possible that additional manipulation of these and other Iperf parameters will further boost the results...although curbing my enthusiasm is the thought that it makes no sense to use values that don't mimic real-life configurations (again, readers, suggestions?).

Granted, this wasn't an apples-to-apples comparison:

• Intellon had upgraded the HomePlug AV adapters' firmware to the latest production version 'INT6000-MAC-3-0-3052-1120-20070816-FINAL-B' before the testing began, and
• Since I'd uninstalled Iperf from my MacBook, we used Yount's laptop as the roaming system

However, if you compare yesterday's data to my earlier results, I think you'll also be impressed with the improvement. We didn't test the powerline spur that runs between my office and the back yard 'hut', but I suspect the results there will proportionally increase by a similar factor. As a first-pass estimate, the 'swamp cooler closet' adapter is the one with a MAC address ending in '1c:8a' in the above screenshots (showing measured PHY performance via both Intellon and Linksys utilities).

Why did DS2's adapters outperform their HomePlug AV opponents in my earlier testing? Atallah told me that, based on his understand of competitor DS2's technology approach, it employs smaller powerline transmit packets than does HomePlug AV. HomePlug AV's powerline packet size is 64 Kbytes, and the adapters implement an accumulation-before-transmission function if each incoming TCP packet to be broadcasted over the power grid is smaller than 64 Kbytes. The overhead of this accumulation function, coupled with the small TCP buffer length and window sizes I employed, garnered DS2 an inherent advantage in my earlier study (at least in Intellon's opinion).

As a result, I've decided (at least for the moment) to swap out the DS2-based NETGEAR HDX101 adapters I was using prior to yesterday's testing in favour of HomePlug AV-based Linksys PLE200 units. I feel more confident about HomePlug AV's TCP capabilities than I did before. And as I mentioned in one of my online addendums, I'm of the opinion that HomePlug AV will probably be the dominant long-term powerline technology in the Americas (although DS2 and partners Buffalo, D-Link and NETGEAR are certainly putting up a good fight here, and DS2's UPA technology will also likely dominate Europe in the long-term).

I may revisit my evaluation of all three '200 Mbps' powerline technologies for print some time in 2008. Until then, I'll do bits and bytes of testing as my personal bandwidth allows, and I also welcome your thoughts.