[OpenIndiana-discuss] ZFS remote receive

[Richard Elling]

On Nov 1, 2012, at 1:24 AM, Jim Klimov <jimklimov at cos.ru> wrote:

> On 2012-11-01 01:47, Richard Elling wrote:
>> Finally, a data point: using MTU of 1500 with ixgbe you can hit wire speed on a
>> modern CPU.
> 
>> There is no CSMA/CD on gigabit and faster available from any vendor today.
>> Everything today is switched.
> 
> Ok then, I'll stand corrected by the practice, although my networking
> education speaks against this statement. We were taught that since GbE
> retains (at least formally) compatibility with older ethernet, and with
> halfduplex in particular, it is capable of hubbing, csma/cd, etc. These
> provisions are of course suboptimal (useless and harmful to performance)
> on point-to-point fullduplex links like host-to-host and host-to-switch.

It does retain half-duplex compatibility. But when it is FDX, it does not use the
HDX timing. I recall seeing a HDX GbE hub for sale once... but have never 
seen one in real life.

> It may be a matter of particular modernized OS defaults however - to
> disable those obsolete beasts. But they should be there, and when Jumbo
> was introduced (for GbE, none other) - these beasts were known to cause
> these sorts of problems.

Jumbo preceded GbE, but late in the deployment of FastEthernet, so there
weren't many FastEthernet switches that supported Jumbo frames, thus
limiting their deployment. Everything converged with GbE --  inexpensive
switches and Jumbo frame support.

> Likely, interrupt coalescing, NIC offloads and CPU horsepower played
> their roles as well. And buffering, I've rather forgot about this.
> (Also note that buffers blindly thrown at any problem at all layers
> of the stack might only hide the performance degradations and disable
> protocol adaptability to reduced network abilities, as often happens
> with WiFi during interference, for example - draining the several
> hundred packets of the buffer over 1Mbit before even noticing that
> a problem exists, can take a good part of the second, if not more
> than one).

Indeed :-)
 -- richard

--

Richard.Elling at RichardElling.com
+1-760-896-4422