On Fri, Mar 19, 2010 at 10:35:11AM -0400, Alan Stern wrote: > On Thu, 18 Mar 2010, Johan Hovold wrote: > > > Hi, > > > > A while back I did some initial attempts at unifying the single-urb fifo-based > > and the multi-urb dynamically-allocated urb write implementations, but although > > the performance results were promising my design at the time was a bit too complex. > > > > I decided to postpone this and instead turn ftdi_sio into a kfifo-based > > generic driver first, but I soon realised that a single urb would not suffice to > > achieve maximum throughput. > > > > So I kept both schemes to avoid any regressions and instead made the > > prepare_write_buffer interface fit them both. > > > > Today I thought some more about the unification and came up with a new design > > that is both simple and clean, and that not only achieves maximum throughput > > but also does so at significantly lower CPU-usage for small writes (and only > > slightly higher for 2k ones). It currently uses two pre-allocated urbs (and > > buffers) so in contrast to the current multi-urb writes it does no allocations > > during runtime and also puts less pressure on the host-stack queues. > > > > This also simplifies the prepare_write_buffer interface. > > Wow! My suggestion sure got some quick results. The performance > results sound impressive. Thanks. As I mentioned I had already done some work on it, but sometimes not working on something for a while can actually help to make the final pieces fall into place. :) And I wasn't quite happy with the semantics of the first version of prepare_write_urb either. Performance increase is simply due to the fact that it is more efficient to send 256b of bulk data at a time than say 10b (or 1b), and for high baudrates you may also need more than one urb to keep the device buffers filled. Johan -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
