Quick Thoughts and Benchmarks on Moving to Gigabit Ethernet

Netgear GS108 8-port gigabit switch

So gigabit ethernet is finally ubiquitous enough that I (unintentionally) have two machines that both have it. My existing ethernet switch still works, but don't have enough ports on it so I sometimes end up having to unplug stuff and switch cables which is inconvenient.

I've been hesitant to invest in a new switch. My existing switch which is a 10/100 Netgear blue box switch originally gave me problems. About a year after my original purchase, the switch started malfunctioning and then stopped working. It was still under warranty, so Netgear replaced it, but I remember they were very slow about it (3 weeks?). Meanwhile, I was without a switch and ended up buying a temporary switch (a Linksys) to hold me over while waiting for Netgear to replace it. To it's credit, both the replacement Netgear switch (and also the Linksys) still work today. These switches are well past their warranties now.

However, from what I've heard and read, consumer gigabit switches have reliability problems similar to the original switch I bought. It doesn't seem to be uncommon for a gigabit switch to just burn out fairly early in its life. And it sounds like customer service isn't any better and you still wait 3 weeks for a warranty turn around.

Well, Netgear was having a rebate offer on their switches this month on some of their blue box line. I did a little bit of reading and asking around, and the impression I got was that their gigabit switches might actually not self-destruct any more. So I decided to buy a new switch, a GS108 which has plenty of ports for my needs.

Unfortunately, I later read the reviews on Newegg, and many were alarmingly bad. Many seemed to fail within a couple of months if not less, and the customer service still sucks. Oops. Well, too late. So I'm keeping my fingers crossed that this thing will continue working.


Anyway, I decided since my 100Mb switch was still working, I would do a few benchmarks before replacing it and see how the performance compared. 

My setup would be to benchmark between a Windows XP box (Mobile AthlonXP 2.2 GHz / Nvidia NForce2 with Marvell Yukon Gigabit built-in card) an iMac (2.4 Core Duo 2, Mac OS X 10.5.1).

Obviously, there are a lot of variables that affect performance. I didn't have a lot of time to do a comprehensive test, but I did do enough to at least get a loose sense of the differences.

So to measure performance, I would do file transfers under Samba (a.k.a. Windows File Sharing). I know there is a lot of baggage that comes with Samba, especially intermixing Windows and Mac, so this isn't an ideal speed test. (Something like FTP might be better.) But I typically use Samba to exchange files between these two systems, so this is an actual real world case for me that I care about. 

I try two different file tests as well, a large-single file that is 1.07GB, and a large set of small files that is 73.3MB. I forgot to count the number of files so I don't know the number. The files were mostly jpeg files that were usually no larger than 300K, but often a lot less. Though I wasn't terribly interested in the small file transfer performance, so I started skipping the test so I'm missing some data points.

To try to notice any accidental asymmetry in the operation due differences caused by Mac/Windows, CPU speeds, file systems, hard disk differences, Samba implementations, etc, I do the file copy in both directions. I initiate the copy to and from in Windows alone (using the old File Manager). I thought about repeating the test doing the copy from Mac alone just to compare the differences, but I didn't have the time. I did a single test with the large-single file under the 100Mb switch, and the Mac to Windows copy was 37 seconds slower. I don't know if that was an anomaly since I didn't repeat the test or do further ones. In the numbers I report below, a few instances, the numbers were really slow and far off, so I repeated the test and the numbers usually improved so I throw the bad numbers away.


100 Mbit

Direction

Test File(s) Classification

Data Size

Time (seconds)

Throughput (Mbit/sec)

Mac to Windows

large-single

1.07GB

119

73.7

Windows to Mac

large-single

1.07GB

126

69.6

Mac to Windows

small-multiple

73.3MB

17

34.5

Windows to Mac

small-multiple

73.3MB

17

34.5



Gigabit

Direction

Test File(s) Classification

Data Size

Time (seconds)

Throughput (Mbit/sec)

Mac to Windows

large-single

1.07GB

57

153.8

Windows to Mac

large-single

1.07GB

114

76.9

Mac to Windows

small-multiple

73.3MB

10

58.6

Windows to Mac

small-multiple

73.3MB

-

-



Gigabit with Jumbo Frames

Direction

Test File(s) Classification

Data Size

Time (seconds)

Throughput (Mbit/sec)

Mac to Windows

large-single

1.07GB

44

199.2

Windows to Mac

large-single

1.07GB

35

250.4

Mac to Windows

small-multiple

73.3MB

10

58.6

Windows to Mac

small-multiple

73.3MB

-

-


So a few of observations. First, I was surprised that I got up to about 70% of the theoretical bandwidth on the 100Mb switch, especially compared to the gigabit switch which was 25% and below.

Second, there was at least an all-around improvement going to gigabit even if the results didn't even come close to the advertised 10x factor. And enabling jumbo frames did make a noticeable improvement on the large file test.

Still, the under 25% utilization is kind of depressing. I don't know what is to blame since there are so many variables. in play.


I can't complete this entry without it being known that I encountered yet another Leopard bug while doing this. In trying to enable Jumbo frames, I needed to switch the configuration mode value from "Automatic" to "Manual". This seemed to confuse the OS into thinking the ethernet device is disconnected/unplugged, even though the cable was already plugged in and never removed. I was still able to surf the net, but local Bonjour name services seemed to stop. So I had to reboot. (Apple Bug: 5654237)


Copyright © PlayControl Software, LLC / Eric Wing