In article <20030303171129.K3346@snowdrop.l8s.co.uk> you write:
>> rtk(4) appears to be very reasonably implemented to take advantage of
>> some useful ethernet chipset features. I don't know where you get the
>> idea that it is a hack designed to benchmark well.
>
>IMHO the 'early TX start' and 'early RX interrupt' are not 'useful
>ethernet chipset features'.  They are designed soley to improve
>latency at a cost of cpu cycles,

I don't see where the cost in CPU cycles comes in.  The NetBSD drivers
seem to automatically tune the offsets to eliminate underruns, so the
additional CPU usage only occurs for the few packets that actually underrun
while the driver's finding the right settings.

> this may be appropriate to the
>desktop, but are not relevant to servers.

Decent latency is useful for horrors like NFS as well.

-- 
Ben Harris                                                   <bjh21@netbsd.org>
Portmaster, NetBSD/acorn26           <URL:http://www.netbsd.org/Ports/acorn26/>