LAN network performance

Unanswered Question

Hi All,

I have been asked to look into the performance of my LAN as it seems "slow". I have started some initial testing with iperf from my desktop to servers and from servers to servers. I think I understand the output, but I just wanted to make sure I do with someone has used it before. for instance when i run a test between my desktop and a server i get something like the following:

E:\temp>iperf.exe -c <server address>

------------------------------------------------------------

Client connecting to <server address>, TCP port 5001

TCP window size: 63.0 KByte (default)

------------------------------------------------------------

[108] local <pc address> port 51286 connected with <server address> port 5001

[ ID] Interval Transfer Bandwidth

[108] 0.0-10.0 sec 969 MBytes 813 Mbits/sec

In my head, this means that I am getting roughly 81% of my gigabit line speed. is this correct?

If this is correct, what could be my next steps into trying to figure out where the other 19% is going?

TIA,

R

I have this problem too.
0 votes
  • 1
  • 2
  • 3
  • 4
  • 5
Overall Rating: 3 (1 ratings)
Loading.
wochanda Fri, 04/27/2007 - 16:03

Hi,

I'm not too sure how netperf works, but there are a few things I see that can make this value lower than expected:

1. Overhead

Depending on how iperf makes the calculation, it may not take into account ethernet or IP overhead. If it is calculating how long it takes to send the raw data across, you need to factor in a 14-byte ethernet header and 20-byte IP header. If it is sending 1500-byte packets (best case), the overhead is at about 2%. This would put your actual throughput at 830Mbps.

2. Server speed

Every packet you're sending to this server needs to be processed and acknowledged in order to count. Chances are your server cant keep up with a gig worth of traffic, and instead is only processing about 813Mbps.

3. Switch/router oversubscription

There are shared points on every part of a network, such as interface buffers on routers/switches, backplane bandwidth, uplink ports, etc. If any of these is congested, you wont get a full 1Gb through.

813Mbps is better than most people would expect out of a gigabit switched network, and it certainly disputes any claims of a 'slow network'.

Will

Thank you very much for your reply. One thing I have noticed is that if I do the same test the other way, from server to workstation...I get about half of what I do in the first test. roughly 500Mbps. Is there some reason it would be slower one way and not the other? Would any of that have to do with how fast my pc can process info coming from the server?

jasonrandolph Mon, 04/30/2007 - 13:05

Well, this will partly have to do with the PC in question being used, partly the switch being used, type of blade (if applicable), and even the cable type.

Cat 5e cable, while capable of operating at gigabit speeds, isn't ideal. There's a reason why the Cat6 spec exists.

You could be limited in speed by port over-subscription (# of ports per ASIC, 1 ASIC = 1 gigabit of traffic) on the switch.

The size of the packets in question applies as well.

avmabe Mon, 04/30/2007 - 09:35

R,

I highly recommend you not test/troubleshoot reports of "slow network" with iperf. You are just creating traffic that could cause more problems...

That being said, this issue plagues all of us in the networking industry and the first step is to ask WHAT is slow and why the user thinks that. It could be their application... It could be a database.... It could be hundreds of things, none of which are related to the LAN.

I suggest you ask why they think it's slow. Go through your switches and look for logs that look out of the ordinary. Check the actual port the users plug into and watch out for 100/half-duplex... I've seen that one 1000's of times. Check your STP architecture and make sure STP isn't flapping.

Blasting packets will not find the issue... That is reserved for special troubleshooting cases that are not required here...

Actions

This Discussion