netdev
[Top] [All Lists]

Re: performance question: delay interrupts

To: Michael Richardson <mcr@xxxxxxxxxxx>
Subject: Re: performance question: delay interrupts
From: Donald Becker <becker@xxxxxxxxx>
Date: Thu, 24 Aug 2000 12:53:38 -0400 (EDT)
Cc: netdev@xxxxxxxxxxx
In-reply-to: <200008241517.LAA18336@xxxxxxxxxxx>
Sender: owner-netdev@xxxxxxxxxxx
On Thu, 24 Aug 2000, Michael Richardson wrote:

> >>>>> "Marcell" == Marcell Gal <cell@xxxxxxxxxx> writes:
>     Marcell> Alan says the Nicstar ATM card has HW support for [369]00 us INT 
> delay.
>     Marcell> Any ethernet cards known to have similar smart features?
>     Marcell> Is anything like this configurable for some ethernet drivers, has
>     Marcell> anyone done some successful or unsuccessful experiments with 
> this?
> 
>   The AMD PCnet32 will not post an interrupt unless you set a bit in the
> descriptor ring.

This is the common software Tx interrupt mitigation.  Most descriptor based
drivers use this.

> You can also have it provide an interrupt based upon a
> timer.

Timers are common, but not ubiquitous.
A usual issue with Ethernet adapter hardware timers is that the timers are
not consistent. They are usually referenced to variable time bases such the
PCI bus clock or the transceiver clock.  The transceiver clock is the worst
-- it changes in a non-obvious way (not 10x) between 10Mbps and 100Mbps.

>   The Alteon and SysKonnect GBE have similar features.

Hardware interrupt mitigation is a different beast: it can usually do things
such as "I just raised an interrupt a few 10s of microseconds ago.  I'll
defer this one for a bit" or "I'll wait to see that this next packet is not
for me before raising the Rx interrupt".  These cannot be efficiently
emulated with software.

Donald Becker                           becker@xxxxxxxxx
Scyld Computing Corporation             http://www.scyld.com
410 Severn Ave. Suite 210               Beowulf-II Cluster Distribution
Annapolis MD 21403


<Prev in Thread] Current Thread [Next in Thread>