Gigabit up and running...are these typical speeds?

[MIDIman]

Hi all,

I have two computers - both WinXP Pro SP1 - one has built-in Intel gigabit, and the other, an AMD nforce2-based system, has a new US Robotics gigabit card. Both are connected to an SMC gigabit switch (with jumbo frames support), all brand new cabling, cat6e.

All lights are green on the switch, and both are showing as 1000 in WinXP. The SMC's main port is going out to a Netgear wireless 10/100 router for internet access.

So, my first transfer from one computer to another, was a collection of ~200mb files totaling ~4.5gb. It clocked in at 7 minutes, using ~8% of the bandwidth according to WinXP's Task Manager. This was dragging the files from the AMD to the Intel machine.

Doing the other way around (from the Intel to the AMD), both computers show ~75% CPU usage, and it was a bit slower - ~7% of the network according to the Task Manager.


By my calculations, the first transfer is approximately:
4500mb / 7 minutes = 643mb/minute
<or>
4500mb / 420 seconds = 11mb/second


Is this typical for this kind of low-end gigabit solution? CPU loads seem a bit high, don't they?

 

[JackMDS]

You mean MB (B=Byte).

1MB/sec. is 88Mb/sec. (b=bit).

It can be considered as Excellent 100Mb/sec. but very slow Giga.

A Typical Entry Level Giga should be at least twice Faster.

Try to evaluate the connection with Qcheck, http://www.ezlan.net/faq#transfer

:sun:

 

[MIDIman]

OK - I made sure I had the latest networks drivers. In the network card's driver panel (for both computers), I did the following:

- Set/duplex speed to 1000/full, rather than auto

- Set Jumbo frames to 9KB on the Intel (its max is 14KB), and 7KB on the US Robotics (its max value)...the SMC supports up to 9KB. They were both default set to disabled.


This improved speeds up to around 20% from 8%, pretty much doubling the times/speeds give in my first post. But CPU is still a bit high (anywhere from 35% to 75%) on both machines.

Are there any other tweaks that can be made? The switch is unmanaged, so I don't think there are any manual settings that can be made with it.

 

[Genx87]

Found I am more disk limited than anything on my network. I can sustain a pretty solid 18-24% which is anywhere from 23-30MB\sec. The disks I have arent terribly fast but not slow either. Plus I have them setup with a lowend onboard RAID controller in RAID 1.

Before I would be pretty much maxed at 80-85%(10.5 MB\sec) utilization on my 100mbit network. At the very least I have more than doubled my throughput which IMO isnt bad for the money spent.

 

[MIDIman]

Originally posted by: Genx87
Found I am more disk limited than anything on my network. I can sustain a pretty solid 18-24% which is anywhere from 23-30MB\sec. The disks I have arent terribly fast but not slow either. Plus I have them setup with a lowend onboard RAID controller in RAID 1.

Before I would be pretty much maxed at 80-85%(10.5 MB\sec) utilization on my 100mbit network. At the very least I have more than doubled my throughput which IMO isnt bad for the money spent.

Great points and something to think about. I have all Seagate 7200.7's and .8's so I didn't think they'd be limited to this, and I generally keep them well defragged. But I hadn't heavily tested my 10/100 network from before, and I know I'm getting at least twice the throughput now, so yeah - for the $30 spent on the switch, $15 on the card, and less than $10 on all new cables, I'm happy with double the performance.

Just curious if there are any other tweaks that I might've missed.

 

[p0lar]

Try connecting the two machines directly using an interconnect cable (just a straight cat5e cable for gigabit, do not use a crossconnect) to eliminate the possibility that the switch is the culprit.

As well, try packet generation software instead of an operation that taxes your network and your disk IO. Isolation is a key element to debugging. I use iperf on several of the unix-like operating systems with good results; there is also a win32 build.

Good luck!

Edit: Windows 2000/XP TCP Performance Tuning

 

[Rilex]

What kind of transfer are you testing with? SMB/CIFS is going to be slower than say an FTP transfer.

I can transfer a 715MB file over 1500 MTU gigabit in about 2 minutes (from and to UDMA133 drives). The only thing Jumbo gets you is lower CPU utilization, not necessarily faster speeds.

 

[p0lar]

Originally posted by: Rilex
What kind of transfer are you testing with? SMB/CIFS is going to be slower than say an FTP transfer.

I can transfer a 715MB file over 1500 MTU gigabit in about 2 minutes (from and to UDMA133 drives). The only thing Jumbo gets you is lower CPU utilization, not necessarily faster speeds.

While I mostly agree with this, I would like to point out that, as a ratio of header to payload, the header for a jumbo frame is consumes ~0.5% of the total packet; whereas with 1500 byte packets, it's closer to ~3%. This may seem negligible, but it is worth a mention. Any retransmission can severely degrade performance as well while using jumbo frames.

 

[JackMDS]

In addition, Window XP was written more than 4-5 years ago, it is client software, and at the time Entry Level Giga was Not in existence. The OS is Not Optimized to Giga.

If you would load Windows 2003 on the same computers with the same hardware, you would get better performance.

Currently Giga is more a Server backbone thing rather then a Desktop solution.

Mobo providers "Jumped the Gun" and build Giga NIC into the Mobos since the effort and price differential is negligent as compare to 100Mb/sec. NIC, but the rest of the ?World? is not yet Giga ready.

:sun: