10Gb iPerf kit?

Saku Ytti saku at ytti.fi
Sun Dec 7 08:48:04 UTC 2014


On (2014-12-07 09:24 +1300), Pete Mundy wrote:

Hey,

> I've done loads of 1Gbit testing using the entry-level MacBook Air and a Thunderbolt Gigabit Ethernet adapter though, and I disagree with Saku's statement of 'You cannot use UDPSocket like iperf does, it just does not work, you are lucky if you reliably test 1Gbps'. I find iperf testing at 1Gbit on Mac Air with Thunderbolt Eth extremely reliable (always 950+mbit/sec TCP on a good network, and easy to push right to the 1gbit limit with UDP.

In my experience majority of people using iperf in UDP mode do not monitor for
packet loss. And once they start doing, they notice they can't go very fast.
For me 1 packet loss due to host issues is absolutely too much,I need flat 0,
and in optimal scenario, I'd get microsecond resolution jitter statistics.

Granted I've not tried on OSX, perhaps by default it has deeper buffers for
UDP. But on Linux, top of the shelf Cisco UCS server with 10GE interfaces
running RHEL, and 1Gbps is simply too much, you'll get packet loss. You can
observe the packets arriving, but they'll be registered as UDP errors on
'netstat -s', because your program isn't picking them up fast enough.

Things iperf could do to perform better
 - setsockopt for deeper buffers
 - recvmmsg to pick up multiple messages with single context switch
 - raw sockets to reduce overhead

But ultimately you're still going to be very far from 10Gbps, which is doable,
if you'll use something like DPDK.

-- 
  ++ytti



More information about the NANOG mailing list