TR Forums

I have a 100Mbps home network; I added a 1000Mbps switch in my office to handle traffic between 3 PCs (and more). Speeds aren't anywhere near what I expected. I've fiddled with setting to 1000 speed vs autosensing, I've looked at network utilization in Taskmgr.exe.

How can I measure network throughput? What kinds of things should I look at to ensure I'm maximizing my network speeds?

I've got XP SP2 on both desktops and a laptop, onboard NICs that support 1.0Gbps speeds, and the 1000Mbps switch. Using TeraCopy I am seeing ~7-8 MB/s file transfer speeds.

Thank you!

Make sure you have Duplex at auto-negotiation everywhere, when you start to use manual settings 'bad things'(tm) happen.

Or set them to full duplex on every single device and never forget about that and about all other devices you bring to your home even for just a 2 minutes, and never forget to set things back as they were after you disconnect and... you get the point

I picked up an inexpensive Netgear gigabit 8-port switch about a year ago, and I have been extremely pleased by how well everything auto-configured. I plugged the switch into one of the router's ports, plugged everything else into the switch, and it all sorted itself out and worked flawlessly right from the get-go. Two to five PCs (Vista, XP, ME, Ubuntu, Mandriva, etc.), a printer, a DVR, a PS3... it just works. When I was using 100bT, videos streamed across the network were occasionally jumpy. Now, they play without a hiccup.

Leave everything on autonegotiation and then look at what it negotiates to. If everything is good, it should report 1gbps, if it reports down at 100Mbps (or even 10Mbps) then you have connector or wire problems and need to fix those first.

To check network throughput, use Iperf:
http://dast.nlanr.net/Projects/Iperf/#download

If Iperf shows good speeds but your file transfers don't then your problem may be in your storage subsystem rather than your network subsystem.

The task manager network utilization is accurate.

I also have an 8-port gigabit at home. The server and my desktop connect to that and can transfer at over 100mbit speeds. Just set everything to auto and you should be fine.

BTW, isn't gigabit always full-duplex?

Usacomp2k3 wrote:

BTW, isn't gigabit always full-duplex?

Yes, and it also requires autonegotiation to be enabled. You can however set it so that 1Gbps and full-duplex is the only option offered on the autonegotiation to force it to gigabit, I wouldn't recommend that though - let it fail down to lower speeds if the line isn't good enough quality and use that as an indication to fix the line.

Note that it also requires all 4 pairs to be working in the Cat5, whilst 100Mbps and 10Mbps only use 2 pairs. I've had problems when the 100Mbps pairs were good but the other pairs were bad - an Intel NIC would renegotiate down to 100Mbps but an nVidia nForce onboard card would fail to connect (no link).

Moved to Networking from General Hardware.

Kevin

Additionally, be sure that the NIC you're attaching to the network really is capable of gigabit speeds before you start freaking out. A gigabit switch alone won't get you gigabit speeds. The reason I say that is because I thought one of my computers had an integrated gigabit Ethernet NIC. It didn't.

What were your transfer speeds before you put the gigabit switch in place of the 100Mb switch?

Thx all. Pardon my noob-iness in this area. What kind of MB/s transfer rate should I be seeing? I typically see ~7-8MB/s between the PCs on my 1000Mbps switch. It feels slower than transfers used to be in the office on 100Mbps connections (I currently work at home so I cannot benchmark).

Everything is set to auto-negotiate; when I set it to 1000 duplex speeds did drop. I also tried tinkering with jumbo frames (my D-Link switch supports those) but did not see any impact. Both PCs I'm using for testing now are reporting 1.0Gbps connections, and both are XP SP2.

Transfers as measured by TeraCopy:
PC1/RAID0 -> PC2 (copy job on PC1): ~11MB/s
PC2 -> PC1/RAID0 (copy job on PC1): ~11MB/s
PC1/RAID0 -> PC2 (copy job on PC2): ~11MB/s
PC2 -> PC1/RAID0 (copy job on PC2): ~9MB/s

PC2: HD501LJ -> HD753LJ: ~62MB/s
PC2: HD753LJ -> HD501LJ: ~52MB/s
PC1: RAID0 SATA1 -> ATA100: ~14MB/s
PC1: ATA100 -> RAID0 SATA1: ~20MB/s

notfred wrote:

To check network throughput, use Iperf @ http://dast.nlanr.net/Projects/Iperf/#download

Do I need to run that on both PCs, one as a server? What cmd-line switches do you use?

Usacomp2k3 wrote:

The task manager network utilization is accurate.

Mine was around 5% or so, IIRC, with 100Mbps, and with 1000Mbps it's around 8%.

What are those NICs? If it is onboard let us know the motherboard model.

Anything over 12.5 MB/sec is definitely going faster than a 100mbit will allow. This corresponds to roughly 12% on the utilization. Getting somewhere around 30% sustained for small-medium size files should be expected, I would think.

dkrum wrote:

Transfers as measured by TeraCopy:
PC1/RAID0 -> PC2 (copy job on PC1): ~11MB/s
PC2 -> PC1/RAID0 (copy job on PC1): ~11MB/s
PC1/RAID0 -> PC2 (copy job on PC2): ~11MB/s
PC2 -> PC1/RAID0 (copy job on PC2): ~9MB/s

PC2: HD501LJ -> HD753LJ: ~62MB/s
PC2: HD753LJ -> HD501LJ: ~52MB/s
PC1: RAID0 SATA1 -> ATA100: ~14MB/s
PC1: ATA100 -> RAID0 SATA1: ~20MB/s

So let me see if I understand this correctly. Your RAID 0 volume is transferring to an ATA100 drive on PC1 at 14MB/sec and your basic ATA100 drive is transferring to a RAID0 volume on the same system at 20MB/sec?

If that's the case, then I would expect the somewhat slow PC1 --> PC2 cross-network results above, since your volumes can barely transfer faster than 100mbps would allow, although the strict 11MB/sec threshold seems like a hard-limit imposed by a 100mbps cap. The numbers (14MB/sec and 20MB/sec in the same system from RAID0 to ATA100 drive) seem anomalously low.

What type of cable are you using? Clearly, something isn't configured as 1000mbps since your HD501LJ --> PC1 ought to be closer to 14-20MB/sec?

Usacomp, my understanding is that he is posting in-system transfer speeds between volumes. Everything else is hard-capped at 100mbps speeds when he goes for a network transfer.

You could always try this tool which does a memory to memory transfer test thus bypassing disks. I've used it in the past and it worked quite well.

In terms of general expectation of gige throughput I've seen servers do about 60MB/s in each direction simultaneous, that was database traffic i.e. lots of small messages back and forth. I would expect uni-directional traffic to get a bit higher than that though. Saying that I would be very surprised to see PC generated traffic get close to those figures - too many limiting factors in there.

dkrum wrote:

notfred wrote:

To check network throughput, use Iperf @ http://dast.nlanr.net/Projects/Iperf/#download

Do I need to run that on both PCs, one as a server? What cmd-line switches do you use?

You need to run it on both PCs. Run it as a server on one PC with "iperf -s" and on the client pc "iperf -c fred" where fred is the name of the PC running the server.

For gigabit on a PCI you should see ~700Mbps and PCI-E should be >900Mbps. If you are substantially less then you have a network problem. If you are close to these speeds but file copies are slow then you have a storage subsystem problem.

computron9000 wrote:

What type of cable are you using?

My cables are ~6' Cat 5e on PC2 and ~6' Cat 5 on PC1. Both systems report 1.0Gbps connections. My switch, btw, is a D-Link DGS-2208 (probably purchased at NewEgg).

Flying Fox wrote:

What are those NICs? If it is onboard let us know the motherboard model.

FF, good to see you here! Yes, both are onboard.
* PC1= MSI RS480M2-IL (AMD 4200+ in there with 2GB RAM)
* PC2= GA-EP35-DS3R (Q6600 + 4GB RAM)

titan wrote:

Additionally, be sure that the NIC you're attaching to the network really is capable of gigabit speeds before you start freaking out. A gigabit switch alone won't get you gigabit speeds. The reason I say that is because I thought one of my computers had an integrated gigabit Ethernet NIC. It didn't.

They are advertised as such.

titan wrote:

What were your transfer speeds before you put the gigabit switch in place of the 100Mb switch?

Utilization was around 5%. I didn't have TeraCopy back then but when I dial it down it's ~5MB/s.

Usacomp2k3 wrote:

Anything over 12.5 MB/sec is definitely going faster than a 100mbit will allow. This corresponds to roughly 12% on the utilization. Getting somewhere around 30% sustained for small-medium size files should be expected, I would think.

When I get 10% or higher I think I'm doing well (sigh).

notfred wrote:

dkrum wrote:

notfred wrote:

To check network throughput, use Iperf @ http://dast.nlanr.net/Projects/Iperf/#download

Do I need to run that on both PCs, one as a server? What cmd-line switches do you use?

You need to run it on both PCs. Run it as a server on one PC with "iperf -s" and on the client pc "iperf -c fred" where fred is the name of the PC running the server.

For gigabit on a PCI you should see ~700Mbps and PCI-E should be >900Mbps. If you are substantially less then you have a network problem. If you are close to these speeds but file copies are slow then you have a storage subsystem problem.

Cool. Love this tool.
Two tests ran with PC2 as server: Bandwidth= 124/126 Mbps (Transfer= 154/156 MB)
Two tests ran with PC1 as server: Bandwidth= 101/99 Mbps (Transfer= 120/118 MB)

Hmmm.... network problem? Do I need Cat6 cables? My Cat5 and 5e cables are from the 90s

blue_bludger wrote:

You could always try this tool {dkrum: pcattcp} which does a memory to memory transfer test thus bypassing disks. I've used it in the past and it worked quite well.

Cool tool #2, thx!
2 PC1 transfers to PC2:
PC1 reports 15428, 15649 KB/s [would that be 15 MB/s ~ 121Mbps ??? I divided by 1024 to get MB then *8 to get Mb]
PC2 reports 13791, 14161 KB/s
Why each PC reports different throughput?

2 PC2 transfers to PC1:
PC2 reports 12047, 11401 KB/s
PC1 reports 12056, 11394 KB/s

Thanks all so far, this is valuable to me. Are we getting closer?!?! Can I get to gigabit speeds?!?!

You could try using CAT-6 to see if that improves any issues with crosstalk or EMI.

But, in general, I would not recommend using a CAT-5 cable from the 90s (or otherwise) for optimal gigabit performance.

The 5e cable is probably fine, but the CAT-5 cable may not be (could be your only issue). That's why I asked what cables you were using.

Are they short patch cables or longer runs?

Short answer: at least ditch the CAT-5 for CAT-5e. At best, replace both with proper CAT-6.

It may be cables, you could try removing the switch and putting each cable in their and trying the iperf test on each cable and see if one of the cables is bad. Also double check the sockets on the PCs and the switch, sometimes the little copper contacts get bent and if this is on one of the pairs that is needed for Gigabit but not for Fast Ethernet then that might explain the low speeds - PC1 as server looks too suspiciously exactly on Fast Ethernet speeds.

One final thought is do you have the AMD Dual Core Optimizer installed on the AMD system?
http://www.amd.com/us-en/Processors/Tec ... 06,00.html

I know people have reported various wierdness without it installed and I think some of those issues were networking issues.

Switched out the Cat 5 for 5e. iperf results:
PC1 to PC2: 127, 130 Mbps [PC2 as server measured it as 122, 129]
PC2 to PC1: 101, 105 Mbps [PC1 as server measured it as 102, 107]

5e direct between PCs:
PC1 to PC2: 132, 132 Mbps [PC2 as server reported same]
PC2 to PC1: 102, 103 Mbps [PC1 as server reported 103, 104 MBps]

Accidentally ran pcattcp on PC2 when PC1 was running iperf as server; pcattcp reported 10923 KB/s, and iperf reported 98 Mbps.

Yep, installed AMD Dual Core Optimizer last year, thx.

I am starting to think if it's not the 5e slowing me down maybe network configs on the PCs. I would have thought maybe switch but direct connect is yielding same throughput #s (although last time I did this stuff I was told had to use cross-over cables; I have one but it's cat 5).

Made a mistake earlier on PC1 NIC. I am using a PCI card-- Network properties reports it as Realtek RTL8169/8110 Family Gigabit Ethernet NIC.
* 802...vlan tagging: disable
* Flow control: enable
* Jumbo frame: disable
* Link speed: auto-negotiate (can choose 1000, etc)
* Network addr: not present
* Offload checksum: Tx/Rx checksum
* Offload TCP LargeSend: enable
...and wake on lan stuff

On PC1 I'm running:
* Client for MS Networks
* Eacfilt driver (not sure what that is, properties button is greyed out)
* File and printer sharing for MS networks
* QoS packet scheduler
* Network monitor driver
* tcp/ip

IIRC PC1 is configured as the, oh what's it called, master controller or something? The host of the networks clients. Did this 2 years ago or so.

BTW, motherboard is MSI RS480M2-IL (AMD 4200+ in there with 2GB RAM)

Task manager, network properties, PC Wizard 2008 all report a 1.0Gbps connection.

PC2 is using onboard NICs: GA-EP35-DS3R Rev 2.1 (Q6600 + 4GB RAM). Network properties reports it as Realtek RTL8168/8111 PCI-E Gigabit Ethernet NIC.
* 802...vlan tagging: disable
* Flow control: enable
* Jumbo frame: disable
* Link speed: auto-negotiate (can choose 1000, etc)
* Network addr: not present
* Offload checksum: Tx/Rx checksum
* Offload TCP LargeSend: enable
...and wake on lan stuff

PC2 is running:
* Client for MS Networks
* File and printer sharing for MS networks
* QoS packet scheduler
* tcp/ip

PC3 is a Dell laptop with running XP SP2, and onboard NIC is Broadcom NetXtreme 57xx Gigabit Controller. It is on a long Cat5 cable (~15 feet?) connected to the 1.0Gbps switch. Network configs looks the same, all are set to auto.

PC3 -> PC2: 162, 173, 402, 392, 408, 422 Mbps
PC3 -> PC1: 118, 121, 110 Mbps

Did PC3 -> PC2 once via wireless G: 20 Mbps

Can't get PC3 to receive; it's a company asset with Norton, and I suspect it's preventing certain TCP traffic.

PC4 is in another room and is wired to my router (Linksys WRT54GL at 100 Mbps wired). (The switch is in my office and PC1 and PC2 are connected to it.) PC4 is connected at 100 Mbps and transfers to PC2 at 91 Mbps (measured twice).

Looks to me like something is wrong with PC1. Try forcing 1000 on PC1 instead of auto-negotiating.

15 feet is not a long CAT-5 run. Anything around 250-300 feet is considered long, since that approaches the maximum distance for CAT-5 runs. (just an FYI)

dkrum wrote:

IIRC PC1 is configured as the, oh what's it called, master controller or something?

Primary Domain Controller (PDC)?

Maybe try something like an Ubuntu 8.04 live CD on each one in turn or maybe even on both PCs and see if iperf gives better results. If it does then you know you have a software problem. If it doesn't change the speeds then it looks like a definite hw problem.

computron9000 wrote:

Looks to me like something is wrong with PC1. Try forcing 1000 on PC1 instead of auto-negotiating.

Yep, I was thinking that's what the #s are indicating. Setting it to 1000/full duplex, iperf xfers are 130 & 126 Mbps. A couple test file transfers (using TeraCopy) came in at 10 MB/s (wouldn't that be 80 Mbps?). Hmmm.... wish I had an extra PCI gigabit NIC lying around to see what that might do.

notfred wrote:

Maybe try something like an Ubuntu 8.04 live CD

Oooohhh, live CD. I've recently tried creating a boot CD but got stuck trying to get other utilities onto it. You have a link that makes this an easy process? This brain is getting fried. Back in the day I relished this kind of tinkering.

computron9000 wrote:

15 feet is not a long CAT-5 run. Anything around 250-300 feet is considered long

Thank you, appreciate the that.

titan wrote:

Primary Domain Controller (PDC)?

Master browser.

(BTW, iperf defaults to 63KB TCP window size when run as client and 8KB when run as server. Don't know if that means anything. I've also played around with various settings-- buffer size, tcp window size, nodelay, mtu, udp-- and specifying my own MPG files-- everything is around 130 Mbps for PC1 -> PC2.)

dkrum wrote:

computron9000 wrote:

Looks to me like something is wrong with PC1. Try forcing 1000 on PC1 instead of auto-negotiating.

Yep, I was thinking that's what the #s are indicating. Setting it to 1000/full duplex, iperf xfers are 130 & 126 Mbps. A couple test file transfers (using TeraCopy) came in at 10 MB/s (wouldn't that be 80 Mbps?). Hmmm.... wish I had an extra PCI gigabit NIC lying around to see what that might do.
...
(BTW, iperf defaults to 63KB TCP window size when run as client and 8KB when run as server. Don't know if that means anything. I've also played around with various settings-- buffer size, tcp window size, nodelay, mtu, udp-- and specifying my own MPG files-- everything is around 130 Mbps for PC1 -> PC2.)

Bear in mind that there is some overhead with wired networks, too. You'll never reach 100 or 1000 Mbps exactly. There are several headers attached to each data packet. So, ~80 Mbps sounds about right. Maybe more, maybe less depending on what options you have attached to each packet.

What's the possibility that you're not the only making transfers to and from the server when you're doing the tests? That might be part of your problem. Have you looked at the switch to see if one of the ports is flashing a lot? Maybe there is a bad NIC somewhere on the network that is interfering with the performance.

You really have no other option than to remove variables one by one until you improve performance.

Different transfer medium (disk, memory, DVD, etc.)
Different NIC (onboard, PCI, etc)
Try forcing 100mbps and testing. Then force 1000mbps and test. Same results or different?
Update local drivers
Replace / Reinstall OS
And so on.

Eventually, you'll see what effects the speed, but I disagree with the post above that suggests 80mbps is acceptable speed for 1000mbps LAN transfers. 10MB/sec sustained is slow, even for old, crappy hard drives and there is more than ample bandwidth on such a small 1000mbps network.

Can you post the model of your switch?

dkrum wrote:

Oooohhh, live CD. I've recently tried creating a boot CD but got stuck trying to get other utilities onto it. You have a link that makes this an easy process? This brain is getting fried. Back in the day I relished this kind of tinkering.

Ubuntu's install CDs boot into a live CD desktop by default, so creating the boot CD is a no-brainer.

As far as getting the utilities installed goes, if nothing else you should be able to boot from the live CD and run utilities off of other external media (e.g. flash drive).

Just burn the Ubuntu install / live CD and boot from that. Once it is up you may need to "sudo apt-get install iperf" and then you can just do the "iperf -s" or "iperf -c fred".

computron9000 wrote:

Eventually, you'll see what effects the speed, but I disagree with the post above that suggests 80mbps is acceptable speed for 1000mbps LAN transfers. 10MB/sec sustained is slow, even for old, crappy hard drives and there is more than ample bandwidth on such a small 1000mbps network.

I can see where the misunderstanding took place. My bad. I was talking about having around 80Mbps transfer rate on a 100Mbps Ethernet is about right. I was further stating that you can never reach 100% of the specified capacity because of the overhead generated by individual packet headers and header options. I should have clarified, or even just changed the order of my sentences a bit, which specification I was talking about.

That makes more sense.