Re: [mpls-linux-devel] Fedora x86_64 Kernel

[James L. <jl...@mi...>]

Wow. That is really bad.  I will start digging deeper.

On Tue, Nov 17, 2009 at 09:31:55AM -0500, Chris Robson wrote:
>
> Hello James
>
> Yes it has been a very long time but I've been determined to get time to  
> continue the MPLS-linux performance testing over 10GE.
>
> Finally I can report some results but they are not very good.  First, I  
> switched to netperf as it provides far better results and seems more  
> reliable than iperf.
>
> So here is the layout of the test configuration.
>
> 1.  Each host configuration: Dell 860 dual PCIe x8-lane/x4-lane with  
> Myricom Gen 2 PCIe 10GE NICs (cards are in x8-lane slot)
> 2.  Kernel version is:  2.6.29.6-217.2.16.f11 with MPLS-Linux patches
> 3.  IProute version is:  2.6.29-2.f11 with MPLS-Linux patches
> 4.  netperf command issued:  netperf -L {local LER IP} -H {target LER IP}
>      netserver -L {local LER IP}
> 5.  Network layout, back-to-back hosts acting as LERs such that:   
> LER<---->LER
> 6.  example MPLS configuration commands issued on both LERs
>      mpls labelspace set dev eth{interface to nexthop LER} labelspace 0
>      mpls nhlfe add key 0 instructions push gen {label} nexthop  
> eth{interface to nexthop LER} ipv4 {necthop IP}
>      mpls ilm add label gen {label} labelspace 0
>      ip route add {nexthop target subnet IP/mask} via {nexthop IP} mpls  
> {key}
>
> Test summary results:
>
> 1.  Baseline rates for directly attached interfaces:  9.3G
> 2.  Baseline rates for VLAN Tagged interface using "interface routes":  9.3G
> 3.  Baseline rates for non-MPLS routed subnets:  5.0G(plus-or-minus 1G)
> 3.  Routed IP thru MPLS LSP:  0.45G <------------- less than 1G, this is  
> not good
>
> Let me know if there is anything particular you'd like looking into on this.
>
> ...Chris
>
>

-- 
James R. Leu
jl...@mi...