Re: NetBSD/sparc64 and GbE performance

[Julian Coleman <jdc%coris.org.uk@localhost>]

Hi,

> But I seriously doubt that the cas(4) hardware is faster or even as
> fast as a wm(4) card.

Having done some work on the gem(4) driver and looked at the cas(4)
documentation, perhaps I could offer a few comments:

  with a gem card in a 143Mhz U1, I could get 100Mb/s (I didn't have a
  1Gb/s card to test against at the time).  The CPU wasn't 100% busy.
  I could probably re-run tests with an E3500 and a 1Gb/s network.  This
  would be comparable with the current state of our cas(4) driver (see
  below).

  I received a comment that Apple got better throughput on the same gem
  hardware than we did.  However, I'm not sure if that points to the gem(4)
  driver, or our stack.  Our cas(4) driver is very similar to the gem(4)
  one.

  our cas(4) driver is currently very primative, and does not support any
  of the hardware features of the chip.  It has transmit and receive TCP
  checksum calculation, TCP stream coalescing, and packet batching (multiple
  small packets to a single receive buffer).  It also has the ability to
  match TCP streams to up to 4 interrupt lines, so that different CPU's can
  handle different streams.  (We are fairly unlikely to support this,
  unless someone reworks the PCI interrupt code to handle it).  We also
  have a problem with recognising the card on non-sparc hardware.

  the card has an external PHY to support copper and fibre.  There are
  single port fiber and copper cards available, as well as 4-port copper
  cards (PCI and PCI-X).

However, the card is an old design by modern standards (2003 or earlier).
So, some more modern chips/cards are likely to give better performance.

Thanks,

J

-- 
  My other computer also runs NetBSD    /        Sailing at Newbiggin
        http://www.netbsd.org/        /   http://www.newbigginsailingclub.org/