New wireless router - D-Link DGL-4300 owners chime in, please!

[Tommy44]

By the way, PC#1 was the PC at the top of my signature and PC#2 is the bottom PC. PC#2 has a few changes that I haven't updated in the sig yet. I believe it has 1 GB RAM and different hard drives than those listed.

[StarTraveller]

It seems I fell asleep after my previous post, hehe. But what a brilliant update to wake up to!

The numbers indicate a throughput of about 7.5 MB/s. It's not quite the 10 MB/s you normally expect at 100 Mbit, but as the sanity test showed even the Gbit interface maxes out around 8.75 MB/s due to overhead and the limitations imposed by the OS (Windows File Sharing is notorious for its overhead) and the hardware.

The important thing to me is that 7.5 MB/s is quite acceptable in terms of what I need - and due to the nature of the test, it's to be expected that the measured throughput forms a baseline with regards to performance. It's also very positive to note that GameFuel doesn't appear to affect throughput.

Thank you so much your assistance, Tommy! It looks like I'll have to order er D-Link DGL-4300 now, hehe.

[StarTraveller]

I put in the order, but they couldn't give me a delivery date on it right away. I have a naive hope that I can get it before the weekend, but it's probably unlikely.

To new toys!

[Tommy44]

Originally Posted by StarTraveller

Thank you so much your assistance, Tommy! It looks like I'll have to order er D-Link DGL-4300 now, hehe.

Sure no problem. Glad I could help out.

Do you think it would make a difference if I were to try it out on two PCs that both have GB ethernet cards?

Thanks for helping me get through this. I'm sure you are going to really like this router!!

[bjbgraphics]

Great work guys. StarTraveller, let us know when you get the router and how it performs for you. I'd be interested to hear your experiences with it.

BJB

[shaihulud]

Use iperf to test throughput of TCP and UDP. You are physically set with your network, only the software for the cleint and server hosts has to be installed. Testing with the method that is in use is not going to be conclusive at all. You will get figures, but this will let you know better in the terms of UDP streaming or a connection used with TCP. Windows sharing overhead contaminates any datums IMO, and is known for very limited throughput.

To note QoS/Gamefuel should not retain any bandwidth due to the detected upload but the dynamic fragmentation and packet queues can change the datums some what. This can be somewhat a test of the architecture due to the design of the IP3023 processor and ipOS. You can select the transmission to a lower state of allocable bandwidth and see how the MTU change will affect the possible bandwidth. Technically, MTU's of lower values do not necessarily limit the bandwidth, but can when the packets that normally would fit in a normal ethernet frame become excessive increasing the load on the host system. Smaller packets take more processing and FSB hits than larger sizes.


The instructions for the client and server are at the web page for iperf. To note it is open sourced and used often for simple throughput tests. There is a provided Windows install with the only requirements of Java to be installed.

http://dast.nlanr.net/Projects/Iperf/

Windows Installer: http://dast.nlanr.net/projects/Iperf2.0/kperf_setup.exe
User documentation: http://dast.nlanr.net/Projects/Iperf...ocs_1.7.0.html

My tests with NetI/O using only layer 2 from host (AMD 3500+, Intel 1000/MT desktop PCI adapter/82541GI) to host (P43.0 w/HT enabled, Intel 1000/CT CSA/82547EI) with no overhead is around 74MBps. With overhead it will obviously fluctuate and has been as high at 63MBps to as low at 9MBps. The average Windows share I will usually get is (based on performance monitor watching and no true baseline) 30MBps. Note, this is LAN to LAN.

NetI/O: http://www.ars.de/ars/ars.nsf/docs/netio

[StarTraveller]

Certainly! I'll post a couple of updates on everyday performance once I get it tweaked just the way I want it, hehe.

Tommy, you should be able to get more than 8.75 MB/s (in the LAN-to-LAN scenario) if both computers are fitted Gbit NIC's, but the overall benefit in general probably doesn't warrant getting one unless you have to buying a new NIC due to hardware failure or something like that.

I think Tom's Networking tested the Gbit switch part to perform in the area of 600 Mbit (just pulling it from memory) so if your drives can keep up then you would probably be pushing the limits of what Windows File Sharing can handle, hehe.

[StarTraveller]

Originally Posted by shaihulud

My tests with NetI/O using only layer 2 from host (AMD 3500+, Intel 1000/MT desktop PCI adapter/82541GI) to host (P43.0 w/HT enabled, Intel 1000/CT CSA/82547EI) with no overhead is around 74MBps. With overhead it will obviously fluctuate and has been as high at 63MBps to as low at 9MBps. The average Windows share I will usually get is (based on performance monitor watching and no true baseline) 30MBps. Note, this is LAN to LAN.

MBps as in megabytes per second, right? I just want to be sure that I'm reading you right because first I read it as megabits per second and felt a bit worried about my order.

[shaihulud]

Yes, NetI/O is from socket buffer to socket buffer (LAN to LAN). This figure comes to 78380 KBytes/s TX (AMD 3500+, Intel 1000/MT desktop PCI adapter/82541GI) and 113380 KByte/s Rx (P43.0 w/HT enabled, Intel 1000/CT CSA/82547EI) at 32K packet size. Do division and the totals come to 74.748992919921875 MiB TX and 108.127593994140625 MiB RX. Look at the difference between the PCI (TX) adapter and the CSA (RX). UDP will always have a lower amount of throughput and possible errors, of course none were recorded during testing.

TCP
NETIO - Network Throughput Benchmark, Version 1.26
(C) 1997-2005 Kai Uwe Rommel

TCP connection established.
Packet size 1k bytes: 63399 KByte/s Tx, 109254 KByte/s Rx.
Packet size 2k bytes: 67729 KByte/s Tx, 109895 KByte/s Rx.
Packet size 4k bytes: 76873 KByte/s Tx, 112966 KByte/s Rx.
Packet size 8k bytes: 76828 KByte/s Tx, 112974 KByte/s Rx.
Packet size 16k bytes: 76815 KByte/s Tx, 99703 KByte/s Rx.
Packet size 32k bytes: 78380 KByte/s Tx, 113380 KByte/s Rx.
Done.

UDP
NETIO - Network Throughput Benchmark, Version 1.26
(C) 1997-2005 Kai Uwe Rommel

UDP connection established.
Packet size 1k bytes: 66250 KByte/s (0%) Tx, 99453 KByte/s (0%) Rx.
Packet size 2k bytes: 8224 KByte/s (0%) Tx, 8223 KByte/s (0%) Rx.
Packet size 4k bytes: 16450 KByte/s (0%) Tx, 16446 KByte/s (0%) Rx.
Packet size 8k bytes: 32896 KByte/s (0%) Tx, 32884 KByte/s (0%) Rx.
Packet size 16k bytes: 65784 KByte/s (0%) Tx, 65715 KByte/s (0%) Rx.
Packet size 32k bytes: 65790 KByte/s (0%) Tx, 86792 KByte/s (0%) Rx.
Done.

[StarTraveller]

Overall the measurements show a clear trend with regards to the effect of changing packet size, but for some of the measurements it seems that the test breaks from the general trend.

Did you run repeated tests or could those measurements be flukes where something was making noise on the network?

[shaihulud]

Yes, if I can find the paper by some people at Duke and Intel it explains very well how networking particulars effect the system, specially packet sizes. NOTE, this is not a large frame it is only a packet that is 32K in size and is segmented in the process to the ethernet frame size.

This test is for CPU, FSB, NIC, system buses used by host, network infrastructure, some resources and the interaction of them. There is no other contention in this test. So, something such as a hard drive will not pollute the capable throughput of the network.

The test does not deviate much at all, and is normal if it does a little. Remember this is from socket buffer, to DMA, to NIC buffer, to wire, and vice versa. This is a good, simple, and free test to use for the network between two hosts.

Maybe some others can try between two hosts on their network? State what you use as did I. For example, AMD 3500+, Intel 1000/MT desktop PCI adapter/82541GI. If you can place the chipset used by the system too. So in my example it should be AMD 3500+, nForce3, Intel 1000/MT desktop PCI adapter/82541GI. It is very simple to use. Just execute the executable to output the supported syntaxes.

EDIT: I have found the paper: http://www.cs.duke.edu/~jaidev/papers/ispass03.pdf

[SRTDodge05]

i dont know if this will help but i have my house running at gigabit and i can send and receive 700mb in 25sec. on 100mb i can send 700mb in 3-4min.

[Tommy44]

Originally Posted by TSgtJoe2005

i dont know if this will help but i have my house running at gigabit and i can send and receive 700mb in 25sec. on 100mb i can send 700mb in 3-4min.

You have GB on both networked PCs and a GB router (switch)? Your numbers seem low compared to the numbers I came up with. I had one GB PC and one 100MB PC connected to a GB router. I wonder what the difference is here...

[SRTDodge05]

all my computers are running gigabit, and all are hooked to a 16port switch. i just sent 705mb file from machine to machine at it took 15sec. my router is only 10/100 linksys wrt54g but im getting ready to get the dlink 4300.

[shaihulud]

Apparently, the based on IP5000 processor products will start to be available in June of this year, 2006. This was announced at E3.