Gig-E speeds as reported by Vista

[XMan]

On my main desktop, copying large files from my Windows Home Server it averages around 11.0MB a second. Jumbo frames are enabled on both systems.

Is that a decent speed, or do I need to do some troubleshooting? I'm using an SMC GS8 switch, Cat 6 cables from monoprice. The Home Server has an Intel Gig-E NIC, and my main PC has the onboard Gig-E by Marvell . . . I'm using a DS3.

 

[p0lar]

Originally posted by: XMan
On my main desktop, copying large files from my Windows Home Server it averages around 11.0MB a second. Jumbo frames are enabled on both systems.

Is that a decent speed, or do I need to do some troubleshooting? I'm using an SMC GS8 switch, Cat 6 cables from monoprice. The Home Server has an Intel Gig-E NIC, and my main PC has the onboard Gig-E by Marvell . . . I'm using a DS3.

Troubleshoot -- that is a good speed at 100mbit/s, but hardly noteworthy of gigabit.

I would check your disk I/O performance prior to blaming your gigE, even though several in this board have poo-poo'ed the Marvell interfaces and drivers. (myself included?)

 

[nweaver]

yeah, that is fine for 100Mb/s, but for gig, meh. Does your switch support jumbo frames as well? If not, then you are likely fragmenting like mad, and that will really slow stuff down. I've not done much "real world" gig testing yet, but I know I can get ~ 980 Mb/s on my spanned port using iperf (i have a tweaked script) and their UDP stuff.

 

[XMan]

The switch supports jumbo frames.

http://www.newegg.com/Product/...x?Item=N82E16833129025

I double-checked all of my settings on the PCs and everything seems to be right. I'm not expecting huge speeds - I mean, I only get 30MB/s transferring files on local systems between drives.

Any suggestions on settings to try is much appreciated! 😉

 

[nweaver]

it depends on disk IO, 30MB/s is 240Mb/s, so 1.5 times as fast as 100Mb/s ethernet. It's not great, but it's not bad, and at this point may be limited by disk IO.

 

[XMan]

Originally posted by: nweaver
it depends on disk IO, 30MB/s is 240Mb/s, so 1.5 times as fast as 100Mb/s ethernet. It's not great, but it's not bad, and at this point may be limited by disk IO.

I may not have been clear on that . . . on my desktop, copying a file from C: to D:, I get 30MB/s. Copying another big file from the server to the desktop gets around 11-12MB/s.

So I'm not expecting it to get anything like the desktop performance, but I guess I expected a bit more than what I'm getting.

 

[nweaver]

Originally posted by: XMan

Originally posted by: nweaver
it depends on disk IO, 30MB/s is 240Mb/s, so 1.5 times as fast as 100Mb/s ethernet. It's not great, but it's not bad, and at this point may be limited by disk IO.

I may not have been clear on that . . . on my desktop, copying a file from C: to D:, I get 30MB/s. Copying another big file from the server to the desktop gets around 11-12MB/s.

So I'm not expecting it to get anything like the desktop performance, but I guess I expected a bit more than what I'm getting.

I would think you could get 30Mb/s...if that's around your disk speed for rights.

There are multiple things that can contribute to this. Check that you are linking at gig speeds, replace home made/old cables with premade cables, turn your frames to default and try again, check for the newest NIC drivers. Check CPU usage on both machines while copying. Try a pull instead of a push or vice versa.

 

[p0lar]

Originally posted by: nweaver
There are multiple things that can contribute to this. Check that you are linking at gig speeds, replace home made/old cables with premade cables, turn your frames to default and try again, check for the newest NIC drivers. Check CPU usage on both machines while copying. Try a pull instead of a push or vice versa.

I agree, 30MB/s should be achievable. Some NICs (Realtek) have bizarre limitations significantly lower than a normal 8k jumbo frame. Try setting your frame size down to 4000 bytes to see if such a limitation exists. Also, be sure that you're not using crossconnect cables for any gigabit links and that your switch is actually reporting correct auto negotiation of speed and duplex.

 

[InlineFive]

Originally posted by: p0lar
Also, be sure that you're not using crossconnect cables for any gigabit links and that your switch is actually reporting correct auto negotiation of speed and duplex.

Doesn't the IEEE 802.3ab standard specify auto-MDIX rendering the UTP cable type a moot point? Please correct me if I'm wrong....

 

[p0lar]

Originally posted by: InlineFive

Originally posted by: p0lar
Also, be sure that you're not using crossconnect cables for any gigabit links and that your switch is actually reporting correct auto negotiation of speed and duplex.

Doesn't the IEEE 802.3ab standard specify auto-MDIX rendering the UTP cable type a moot point? Please correct me if I'm wrong....

It does, but that didn't stop early vendors from failing to conform or taking matters into their own hand. I've seen it go both ways -- in fact, I have seen certain older Intel cards actually negotiate down to 100mbit/s when using a crossconnect cable. Intel, of all people!

 

[nweaver]

Originally posted by: p0lar

Originally posted by: InlineFive

Originally posted by: p0lar
Also, be sure that you're not using crossconnect cables for any gigabit links and that your switch is actually reporting correct auto negotiation of speed and duplex.

Doesn't the IEEE 802.3ab standard specify auto-MDIX rendering the UTP cable type a moot point? Please correct me if I'm wrong....

It does, but that didn't stop early vendors from failing to conform or taking matters into their own hand. I've seen it go both ways -- in fact, I have seen certain older Intel cards actually negotiate down to 100mbit/s when using a crossconnect cable. Intel, of all people!

dig up my post about an intel chipset having MAJOR issues linking to a Cisco 3508...changing to a different intel chipset fixed the issue, but I was amazed that we would have 50+% packet loss between intel and cisco gear.

 

[p0lar]

Originally posted by: nweaver
dig up my post about an intel chipset having MAJOR issues linking to a Cisco 3508...changing to a different intel chipset fixed the issue, but I was amazed that we would have 50+% packet loss between intel and cisco gear.

Neither Cisco nor Intel fail to surprise me anymore, and to put it mildly, I make a living on Cisco (God only knows how much money I have lined their pockets with). That being said, I'm not a staunch supporter of that particular series (3508), so I might have leaned either way on that call. 😉 I still feel that the Intel 8255x (fxp), were some of the greatest adapters ever released for general consumption (ironically, those are what wound up in the PIX, though I think that was nary a Cisco design element at that point). Ok, I'm rambling.... sorry OP! 😀

 

[XMan]

Originally posted by: nweaver

Originally posted by: XMan

Originally posted by: nweaver
it depends on disk IO, 30MB/s is 240Mb/s, so 1.5 times as fast as 100Mb/s ethernet. It's not great, but it's not bad, and at this point may be limited by disk IO.

I may not have been clear on that . . . on my desktop, copying a file from C: to D:, I get 30MB/s. Copying another big file from the server to the desktop gets around 11-12MB/s.

So I'm not expecting it to get anything like the desktop performance, but I guess I expected a bit more than what I'm getting.

I would think you could get 30Mb/s...if that's around your disk speed for rights.

There are multiple things that can contribute to this. Check that you are linking at gig speeds, replace home made/old cables with premade cables, turn your frames to default and try again, check for the newest NIC drivers. Check CPU usage on both machines while copying. Try a pull instead of a push or vice versa.

1. All machines are linking at Gig speeds

2. All the cables are Cat-6 rated from Monoprice

4. NIC drivers are latest

5. CPU usage is not much at all on my Core 2 Duo - 2-3% total while copying.

6. Pull / push doesn't seem to make a difference

Aaaaand . . . aha! I tried the third thing on your list. I dropped jumbo frames down to 4088 from 9014 and speeds jumped up to 25.5MB/s.

Why would that be exactly? Wouldn't larger frame sizes give you higher speeds?

 

[spidey07]

If you are not correctly setting the jumbo frame size it can significantly harm your throughput. See what the switch and NICs support, set it to the minimum between all devices involved.

 

[nweaver]

Also, I just realized that this is Windows Home Server. Perhaps that's part of the issue?

sounds like a frame issue.

 

[p0lar]

What other options for frame size do you have?

 

[VisionxOrb]

I used to have horrible throughput on my vista machines around 10-12 MB a sec on my gig network until I installed the performance,compatiblity and reliability packs that MS just recently released. Now I get 38-39 MB a sec and im not using jumbo frames.