Solved: Gigabit Throughput

tonyraboza · ‎05-09-2008

Hi,

I have a Cisco 2960 switch. As a test, I connected two Windows XP machines on the Gigabit ports - no other machines are connected to the switch. And both the Windows XP machine and the switch confirmed that the connection is Gigabit. I shared a folder on one machine and transferred files (around 3 GB) over to it from the other machine. The maximum transfer rate I got was only around 190 Kbps? Is this normal? I did not expect to get full Gigabit transfer but I also did not expect it to be like this? What do you think of this? Anybody have a similar experience? Thanks.

Joseph W. Doherty · ‎05-10-2008

"When I tested on the Gigabit ports - I wasn't really expecting the full Gigabit bandwidth - but I think 190 Mbps is much too slow."

Not necessarily. At Gig network speeds, doing an actual file copy, sustained disk read/write speeds can come into play. So can NIC hardware performance, NIC driver performance, PC bus performance, OS overhead, disk fragmentation, file copy protocol, etc. There's also the bandwidth loss to network overhead, L2/L3/L4 headers, and L2 timing requirements (especially for gig Ethernet).

If you try a traffic generator application, I would expect you to see a boost in throughput.

View solution in original post

sean.cheney · ‎05-13-2008

Use iperf to test the throughput. It excludes IO.

View solution in original post

vosm · ‎06-03-2008

When we were testing out WAAS we also noticed that doing a copy in windows is a lot slower than an FTP or even going to DOS and copy the file. This might show you end user experience but don't count on it as how your network is actually performing.

View solution in original post

Johnny Nkunda Kabundi · ‎05-10-2008

Hi,

What is the speed of your NICs card connected to that two Windows XP Machines ?

Is it a FastEthernet NICs card or Gigabit ?

Did you define the speed to each gigabit port statically in the switch configuration ?

tonyraboza · ‎05-10-2008

The two Windows XP machines have Gigabit NIC cards and they had Gigabit connections when I checked on Windows XP.

No, I did not define it statically - their speed was Gigabit when I connected them to the Gigabit ports on the switch.

Thanks.

Joseph W. Doherty · ‎05-10-2008

190 K bits/s or 190 K bytes/s; and definitely not 190 M bits/s? If the latter, 190 M bits/s could be reasonable without jumbo frames and depending on disk speeds and other considerations, otherwise much too slow.

tonyraboza · ‎05-10-2008

Yes - you are correct - sorry about this. It is 190 Mbps.

When I tested on the Fast Ethernet ports of the switch, I was able to get about 86 Mbps when doing file transfer between the two Windows XP machine.

When I tested on my Linksys wireless router and my laptop (54 Mbps) - I got around 23 Mbps.

When I tested on the Gigabit ports - I wasn't really expecting the full Gigabit bandwidth - but I think 190 Mbps is much too slow.

Thanks for mentioning jumbo frames - I didn't know about it - but I'm reading up on it now.

Without jumbo frames - anybody achieved transfer rates of about 600 Mbps?

Thanks.

Joseph W. Doherty · ‎05-10-2008

"When I tested on the Gigabit ports - I wasn't really expecting the full Gigabit bandwidth - but I think 190 Mbps is much too slow."

Not necessarily. At Gig network speeds, doing an actual file copy, sustained disk read/write speeds can come into play. So can NIC hardware performance, NIC driver performance, PC bus performance, OS overhead, disk fragmentation, file copy protocol, etc. There's also the bandwidth loss to network overhead, L2/L3/L4 headers, and L2 timing requirements (especially for gig Ethernet).

If you try a traffic generator application, I would expect you to see a boost in throughput.

sirdudesly · ‎05-10-2008

you could try setting up a ramdisk and seeing what sort of transfer speed you get from that.

sean.cheney · ‎05-13-2008

Use iperf to test the throughput. It excludes IO.

tonyraboza · ‎05-15-2008

To everyone who replied to my post - thanks a lot! Learned so much. Anyway, I used iperf on the two WinXP machines and I was able to get about 749 Mbits/sec.

Again, my thanks to everyone.

aworrell · ‎05-29-2008

Hi Tony -

Just a few thoughts. I am not a switch engineer, but your situation doesn't sound normal.

Have you checked whether or not the switch connection is set to Full or Half Duplex?

If the PC is not set to autonegotiation, then half duplex is the default on many switches. Try setting everything to Full. This could be a big part of your problem.

Also, what are the MTU settings on your PC's?

Al Worrell

arash_compr · ‎05-30-2008

i had the similar case with 3com super stack, try to change your NIC speed from 10 half to 100 full duplex,if you got better at speed, keep 100 mbps for those two system, however don't think so its speed mismatch problem.

rstroobant · ‎06-01-2008

Hi,

I've had to deal with this question from clients in the past. Along with josephdoherty's comments above, you need to consider the latency imposed by TCP transaction delays. PCATTCP is useful for demonstrating this as you can run the test in TCP or UDP mode.

If you were to span one of your test ports to another port and attach a sniffer you will be able to see the TCP delays.

To test true network performance, products such as SmartBits and EXFO are typically used. Using PC's to test network performance is not an optimal method.

vosm · ‎06-03-2008

When we were testing out WAAS we also noticed that doing a copy in windows is a lot slower than an FTP or even going to DOS and copy the file. This might show you end user experience but don't count on it as how your network is actually performing.

dmann · ‎06-03-2008

We had a similar issue and upgrading the NIC drivers resolved the issue.