RE: [Xen-devel] [PATCH] Network Checksum Removal

["Ian Pratt" <m+Ian.Pratt@xxxxxxxxxxxx>]

What does the tx hw csum control actually turn on and off?

I'm surprised there's much benefit to csum offload on the tx side at all
as its almost always done as part of a copy.

I'd have thought the main benefit of csum offload was on the rx side, so
that packets received by the NIC are hardware csum'ed, passed through
the bridge, and then into the domU where the csum re-calculation is
avoided [it would normally need to be done before the TCP ack is sent,
and can't be done as part of a copy as the data won't be moved out of
the skb until the user app does a read].  The same rx csum check will be
avoided and hence provide benefit to domU <-> domU transfers.

In the figures below, which direction is the data stream heading? (I
presume it's a one way test, like ttcp?)

It's somewhat surprising that the dom0 bridge code is burning so much
CPU. xenoprofile results will be quite interesting to see what functions
are eating the CPU.

Ultimately, the best way of doing domU <-> domU networking will be to
allow point-to-point connections where netfronts are connected direct to
other netfronts if the hosts are on the same machine. However, the
priority for 3.0 is to optimise the normal front-back-bridge-back-front
path.

Thanks,
Ian

> -----Original Message-----
> From: Andrew Theurer [mailto:habanero@xxxxxxxxxx] 
> Sent: 25 May 2005 15:39
> To: Jon Mason; xen-devel@xxxxxxxxxxxxxxxxxxx
> Cc: Ian Pratt; bin.ren@xxxxxxxxxxxx
> Subject: Re: [Xen-devel] [PATCH] Network Checksum Removal
> 
> Tests for domU->dom0, domU->host, and domU->domU are completed:
> 
> 3.2 GHz Xeon with Hyperhtreading, 2GB (correction) memory
>                                                               
>                                                                       
> Benchmark: netperf2 -T TCP_STREAM
>                                                               
>                                                                       
> dom0, dom1, and dom2 on cpu0 (first SMT thread on first core)
>  domU to host
>   hw tx csum
>    msg-size: 00064  Mbps: 0186  d0-cpu: 49.38  d1-cpu: 44.35
>    msg-size: 01500  Mbps: 0917  d0-cpu: 62.13  d1-cpu: 37.87
>    msg-size: 16384  Mbps: 0933  d0-cpu: 66.63  d1-cpu: 33.37
>    msg-size: 32768  Mbps: 0928  d0-cpu: 66.96  d1-cpu: 32.66
>   sw tx csum
>    msg-size: 00064  Mbps: 0187  d0-cpu: 49.50  d1-cpu: 44.52
>    msg-size: 01500  Mbps: 0904  d0-cpu: 60.63  d1-cpu: 39.36
>    msg-size: 16384  Mbps: 0924  d0-cpu: 63.98  d1-cpu: 35.98
>    msg-size: 32768  Mbps: 0926  d0-cpu: 64.18  d1-cpu: 35.68
>       ^^about 2% reduction in cpu util on dom1^^
>  domU to dom0
>   hw tx csum
>    msg-size: 00064  Mbps: 0014  d0-cpu: 64.02  d1-cpu: 31.71
>    msg-size: 01500  Mbps: 1087  d0-cpu: 63.34  d1-cpu: 36.67
>    msg-size: 16384  Mbps: 1204  d0-cpu: 67.30  d1-cpu: 32.71
>    msg-size: 32768  Mbps: 1148  d0-cpu: 68.08  d1-cpu: 31.93
>   sw tx csum
>    msg-size: 00064  Mbps: 0014  d0-cpu: 64.88  d1-cpu: 32.39
>    msg-size: 01500  Mbps: 0948  d0-cpu: 62.20  d1-cpu: 37.80
>    msg-size: 16384  Mbps: 1063  d0-cpu: 64.73  d1-cpu: 35.27
>    msg-size: 32768  Mbps: 1012  d0-cpu: 65.71  d1-cpu: 34.30
>       ^^upto 13% throughput increase with cpu util down ~2% on dom1^^
>         Note the dismal performance for very small msg sizes
>  donU to domU
>   hw tx csum
>    msg-size:00064 Mbps: 0359  d0-cpu: 27.85  d1-cpu: 53.68 
> d2-cpu: 18.48
>    msg-size:01500 Mbps: 0594  d0-cpu: 47.42  d1-cpu: 21.77 
> d2-cpu: 30.78
>    msg-size:16384 Mbps: 0619  d0-cpu: 49.66  d1-cpu: 18.81 
> d2-cpu: 31.53
>    msg-size:32768 Mbps: 0616  d0-cpu: 49.58  d1-cpu: 18.68 
> d2-cpu: 31.74
>   sw tx csum
>    msg-size:00064 Mbps: 0361  d0-cpu: 27.81  d1-cpu: 53.58 
> d2-cpu: 18.62
>    msg-size:01500 Mbps: 0584  d0-cpu: 46.22  d1-cpu: 23.18 
> d2-cpu: 30.60
>    msg-size:16384 Mbps: 0602  d0-cpu: 47.99  d1-cpu: 20.33 
> d2-cpu: 31.69
>    msg-size:32768 Mbps: 0603  d0-cpu: 47.67  d1-cpu: 20.59 
> d2-cpu: 31.74
>       ^^About a 2% throughput increase, and cpu down on d1
>         The cpu wasted on dom1 should be enough justification for
>         domU<->domU communication with point to point front end driver
>         communication.  
> dom0 on cpu0, dom1 on cpu2, and dom2 on cpu3 (dom1 and dom2 on same 
> core)
>  domU to host
>   hw tx csum
>    msg-size: 00064  Mbps: 0540  d0-cpu: 92.98  d1-cpu: 100.00
>    msg-size: 01500  Mbps: 0941  d0-cpu: 99.74  d1-cpu: 48.62
>    msg-size: 16384  Mbps: 0941  d0-cpu: 99.71  d1-cpu: 43.32
>    msg-size: 32768  Mbps: 0941  d0-cpu: 99.72  d1-cpu: 43.21
>   sw tx csum
>    msg-size: 00064  Mbps: 0545  d0-cpu: 93.47  d1-cpu: 100.00
>    msg-size: 01500  Mbps: 0941  d0-cpu: 99.76  d1-cpu: 51.43
>    msg-size: 16384  Mbps: 0941  d0-cpu: 99.69  d1-cpu: 46.58
>    msg-size: 32768  Mbps: 0941  d0-cpu: 99.72  d1-cpu: 45.39
>       ^^Finally at wire speed, but at a cost of 100% cpu on dom0
>         This cpu util seems excessive, maybe oprofile will show
>         some problems.  Notice dom1 has ~2% lower cpu.
>  domU to dom0
>   tx csum
>    msg-size: 00064  Mbps: 0390  d0-cpu: 97.92  d1-cpu: 100.00
>    msg-size: 01500  Mbps: 1571  d0-cpu: 97.36  d1-cpu: 54.83
>    msg-size: 16384  Mbps: 1582  d0-cpu: 96.20  d1-cpu: 49.93
>    msg-size: 32768  Mbps: 1596  d0-cpu: 96.32  d1-cpu: 49.63
>   sw tx csum
>    msg-size: 00064  Mbps: 0375  d0-cpu: 97.65  d1-cpu: 100.00 
>    msg-size: 01500  Mbps: 1546  d0-cpu: 96.36  d1-cpu: 52.99
>    msg-size: 16384  Mbps: 1598  d0-cpu: 95.88  d1-cpu: 47.48
>    msg-size: 32768  Mbps: 1641  d0-cpu: 95.89  d1-cpu: 46.37 
>       ^^very slightly better avg throughput, and lower cpu on dom1
>  donU to domU
>   tx csum
>    msg-size:00064 Mbps: 0287  d0-cpu: 84.97  d1-cpu: 100.0 
> d2-cpu: 75.46
>    msg-size:01500 Mbps: 1004  d0-cpu: 90.98  d1-cpu: 68.29 
> d2-cpu: 76.94
>    msg-size:16384 Mbps: 1018  d0-cpu: 89.78  d1-cpu: 60.82 
> d2-cpu: 78.12
>    msg-size:32768 Mbps: 1010  d0-cpu: 89.30  d1-cpu: 59.83 
> d2-cpu: 77.99
>   sw tx csum
>    msg-size:00064 Mbps: 0286  d0-cpu: 84.81  d1-cpu: 99.93 
> d2-cpu: 76.28
>    msg-size:01500 Mbps: 1018  d0-cpu: 91.30  d1-cpu: 67.27 
> d2-cpu: 75.08
>    msg-size:16384 Mbps: 1012  d0-cpu: 88.46  d1-cpu: 55.56 
> d2-cpu: 71.37
>    msg-size:32768 Mbps: 1017  d0-cpu: 88.33  d1-cpu: 54.96 
> d2-cpu: 70.96
>       ^^about same throughput, but ~4% lower cpu on d1
>         Again, point to point front end comms woudl be great here.
> 
> 
> IMO, I think the patch is a good thing.  There are other very major 
> issues with networking, like the massive cpu overhead for dom0.  I 
> wonder if we could have a layer 2 networking model like:
> 
> -Xen has have front end ethernet drivers only
> -dom0 has a Xen bridge front end driver, just to put eth0 (or 
> whatever 
> phys dev) on it.
> -no domain hosted bridge device or backend ethernet drivers
>  
> With this, Xen acts as a ethernet "switch", switching 
> ethernet traffic 
> in xen itself, without the help of a domain hosted bridge.  
> Packets are 
> forwarded to either a domain's front end driver, or the front end 
> bridge interface in dom0 (or any other driver domain).  With this we 
> may have better control of emulating offload functions, and we should 
> avoid some hops (and in may cases involving dom0) for the netwrok 
> traffic.  Comments?
> 
> -Andrew                                                       
>                                                               
>            
> 
> 
> 

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel