As you may remember, I had been having all kinds of problems getting a D-Link DGS-1008D gigabit ethernet switch to work reliably. (You can read the previous episode in that saga.)
Unfortunately I had not been able to find anyone selling any other brand of switch than the D-Link ones. Through Google I was able to find some other options available in Taiwan. I finally was able to purchase a PCI FXG-08TXJ 8-port gigabit ethernet switch for TW$3100 (US$94). (You can read about my search over on Forumosa.com.)
Many of my readers probably don’t recognize the PCI brand name. It’s a Japanese company who primarily sell in Japan, but many of their products are also available in Taiwan. I had previously had an 802.11b PCMCIA card from them long ago. Their full name is Planex Communications Inc. The one drawback to their products is that most of their written materials are in Japanese. That’s not too big a problem considering the main alternative here is to get something all in Chinese.
The PCI switch itself is enclosed in an all-metal silver colored case. It feels nice and solid, whereas the D-Link has a plastic case and feels cheaply made. The PCI switch has the ethernet ports and all LEDs on the front panel and the power jack on the back panel. The D-Link had LEDs in front and ethernet ports and power in back. Personally I prefer both ethernet and LEDs in front. As a nice touch, they also include a small pigtail for the power brick so that you can only use one power outlet on your power strip.
Plugging everything in, it all worked right away with all computers linking at gigabit speeds. I’ve been using it for one week and there have been zero instances of link being dropped and no other oddities noticed. It just works. The switch supports up to 9.5K sized Jumbo Frames. After installing it I reconfigured the systems with Realtek 8169 cards to use 7K Jumbo Frames (the max size they support). My laptop has a Broadcom 5788 gigabit chipset which does not appear to support Jumbo Frames, so it was left at the standard 1500 byte frame size.
minquan: Solaris 10 6/06 P4-3.0ghz 1gb Realtek 8169 MTU7000
jinzhou: Solaris 11 b41 P2-450mhz dual CPU 384mb Realtek 8169 MTU7000
songjiang: Windows XP SP2 P4-2.53ghz 2gb Realtek 8169 MTU7000
noraneko: Windows XP SP2 PM-2.0ghz 1gb Broadcom 5788 MTU1500
server: iperf -s -m -M 100000 -w 1M -l 24k
client: iperf -c server -m -M 100000 -w 1M -l 24k -t 60
With minquan as server:
jinzhou: 268 Mbits/sec
songjiang: 578 Mbits/sec
noraneko: 493 Mbits/sec
With jinzhou as server:
minquan: 774 Mbits/sec
songjiang: 558 Mbits/sec
noraneko: 432 Mbits/sec
With songjiang as server:
minquan: 580 Mbits/sec
jinzhou: 265 Mbits/sec
noraneko: 332 Mbits/sec
With noraneko as server:
minquan: 612 Mbits/sec
jinzhou: 255 Mbits/sec
songjiang: 284 Mbits/sec
Performance testing was a bit interesting. The ancient 450mhz system did quite poorly as a client but quite well as a server. (Note: In iperf the definition of client and server are kind of backwards. The client sends data and the server receives.) I read elsewhere that the Realtek chipset is very CPU intensive compared to others, and indeed during heavy network usage I see that the CPU is pegged on jinzhou. Apparently the Intel chipsets are the most CPU efficient and also can more easily attain speeds above 800 Mbits/sec. They are a bit more expensive though. This system really needs a motherboard/cpu upgrade.
The laptop system (noraneko) does better but still nothing spectacular. The lack of Jumbo Frames support is probably what is holding it back. songjiang and minquan both turn in quite respectable benchmark results. Also of note, the Solaris scores tend to be better all other things considered. Keep in mind that these are just informal performance tests of raw network performance and don’t take into account actual performance which would be affected by disk speed and protocol efficiency.
All in all a pretty good result this time. I might replace the Realtek boards with Intel ones at some point, though they do pretty well for being 1/3 the price of an Intel board.