Re: [Xen-devel] RTL8139 support

[Keir Fraser <Keir.Fraser@xxxxxxxxxxxx>]

> Success!  Your suggestions worked first time, and my RTL8139 NIC now
> appears to be working fine.

Great! I cleaned up and applied your patch to the unstable tree.
 
> I tested the performance between my regular Linux kernel on Fedora and
> Xen 1.3.  An NFS (using default UDP) copy of a large file was almost 5%
> slower on Xen than Linux, and a scp (TCP) took just over 13% longer. 
> Presumably these are reasonable results given the driver's having to
> make extra copies in Xen?

There are no extra copies happening: in both setups, every packet that
is transferred is copied by the main CPU to or from an on-card
buffer. 

The extra slowdown on Xen is due to using IO_REGS (ie. inb/outb)
rather than memory-mapped IO (readb/writeb). If you want to get that
performance back you'll need to get the driver working without setting
USE_IO_REGS. It *ought* to work --- ioremap() exists in Xen and is
used by other drivers.

> The only unusual result was that Xen's time to flood ping 1000 packets
> over 100 Mbps Ethernet was 10306ms, less than Fedora's 19503ms for the
> same test.  For comparison an nVidia nForce in another machine achieved
> only 861ms, which makes both results for the RTL8139 look slow.

Well yes, the RTL8139 is a slow card! :-)

The nVidia NIC can DMA to/from host memory, so you're avoiding tying
up the host CPU executing PIO cycles to perform the transfer itself.

 -- Keir


-------------------------------------------------------
This SF.Net email is sponsored by: IBM Linux Tutorials
Free Linux tutorial presented by Daniel Robbins, President and CEO of
GenToo technologies. Learn everything from fundamentals to system
administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.sourceforge.net/lists/listinfo/xen-devel