Help me understand my iperf results.

[praecorloth]

So I've got a FreeNAS 8.0.3 box and performance has been so/so. It's been worse, when I first started this I didn't realize my NIC was 100Mbit. :) But since switching to gigabit, I've lived with vastly-superior-to-100Mbit-but-still-not-gigabit performance. I'm topping off just shy of 400Mbit on the reports. So given all of the fun testing I've done in the past, I went to iperf. I fired up the server side on a FreeNAS with a 16K frame, then I ran a client from my Windows box, and a Linux box. Both came back around the 650-700Mbit range. Not good, but not terrible. Linux as the server and Windows as the client came back at 950+Mbit. FreeNAS as the client came back at about 400-450Mbit. Hrm.

So FreeNAS has trouble both as the server and as a client, but definitely more so as a client. What could this mean? I did a netstat -i on the FreeNAS box and it didn't show any collisions. I'm fairly certain, though I can test when I get home, that there have been a couple of different cables between the FreeNAS box and the switch. I might even be able to try a different NIC. But...

It'll still be a Realtek NIC. I had heard a while back that there was a pretty terribad Realtek driver floating around the *nix world recently. Does anyone know how I would go about checking what driver I'm using? I could then go back and see if it's the same driver that people have been complaining about.

 

[cyberjock]

Until 2 years ago I have always just used the onboard NICs. As long as it worked I didn't care. I was never convinced that one NIC would really be that much faster than another one. Maybe 10%, but not much more. I got upset when I built my last windows server because it just wasn't performing as I had hoped. I could copy files from the RAID to the RAID and still get 500MB/sec! But trying to go across Gb LAN was lucky to get 45MB/sec. Searching forums some people swore Intel NICS were so much better. "Better than what" is what I always asked myself. Don't try to tell me an Intel NIC is really THAT much better than onboard. Onboard Gb NICs have been around for 10 years.. companies surely have figured this out by now, right?

As soon as I replaced the built-in NIC on my desktop and Server I went from about 450-700Mb/sec to 800Mb/sec miminum every time I do an iperf test.

Now, anytime a friend of mine has a central repository for his data, I ALWAYS tell him to buy an Intel NIC for his server and workstation. Never been disappointed. Let's face it though. Even if it didn't help at all, are you really going to notice you "threw" away $30 for an Intel NIC. I bet you throw away more money on hardware than that every year. ;)

My short answer: Buy an Intel NIC for your server and desktop, then tell me you're only getting 450Mb/sec ;) I'll probably think you're lying or you have something misconfigured.

Something else to keep in mind.. some NICs do NOT handle jumbo frames well at all. Some work great until you go above an MTU of 9000. You said you did the test with a 16k frame. It could be that your FreeNAS server has a NIC that doesn't handle large frames well.

 

[praecorloth]

Something else to keep in mind.. some NICs do NOT handle jumbo frames well at all. Some work great until you go above an MTU of 9000. You said you did the test with a 16k frame. It could be that your FreeNAS server has a NIC that doesn't handle large frames well.

I thought it was strange too, but that was the default from whichever one of the machines I started on, so that's what I kept throughout the testing. I'll try again with 1500. I'll also keep my eye out for an Intel NIC. What are your thoughts on dual headed NICs?

 

[cyberjock]

My windows server has dual headed NIC.. big waste for a WIndows machine. :/ Because of the way link aggregation works if you are using your server at home it's almost pointless.