Miles Xu wrote:
Darren Reed wrote:
Andrew Gallatin wrote:
Miles Xu wrote:
Paul Durrant wrote:
Andrew Gallatin wrote:
Isn't the overhead to map/unmap the buffer fairly high? Why don't
you keep the buffer mapped?
That was what I was getting at. If the pool is shared amongst
drivers then the buffer has to be unmapped before being recycled
(in case it's allocated by a different driver) thus blowing away
most of the advantage of loanup.
It was also my concern. The original design is not as the current
one. The partial experiments I conducted on e1000g showed no
measurable impacts in terms of throughput. The current design is
also temporary. I may change the design as the experiments continue.
You really need to test on a 10GbE NIC/driver, as I can't imagine that
1GbE NIC will put substantial load on any modern hardware.
That depends on the nature of the traffic.
For a FTP/HTTP transfer, I wouldn't argue with you.
But as the packets get smaller and more numerous, the scales start to
move , especially if you are routing.
Well, for e1000g, nge and some other NIC drivers. When the packets get
smaller to a particular threshold such 512 bytes, the driver actually
allocates a piece of memory and copy the contents from the DMA buffer.
Darren has a point. If your buffer pool stuff is good enough,
then you might see improved performance from loaning for
buffer sizes smaller than the threshold.
Drew
_______________________________________________
networking-discuss mailing list
[email protected]
