Hal Murray <[EMAIL PROTECTED]> wrote:
> Some Ethernet adapters have a bug/feature similar to RS-232 chips.  The
> idea is to batch interrupts to reduce overhead.  Ethernets do it by
> only making one interrupt for several packets as compared to several
> bytes for the RS-232 chips.

> I'd expect gigabit cards/chips will be more likely to have that
> feature than old "slow" 100 megabit chips.

Virtually all the GbE NICs I've ever encountered have such a feature.
And literally all the 10 GbE NCIs I've encountered have it.

You can see such behaviour in a NIC/driver which does it "badly" (IMO)
with a netperf TCP_RR test.  For example if you only ever see
8000-12000 single-byte transactions per second with a "contemporary"
system.

A while back I did a writeup on the tradeoff the NIC/driver strappings
were making.  A version of it can be found at:

ftp://ftp.cup.hp.com/dist/networking/briefs/nic_latency_vs_tput.txt

Not all NIC/driver combinations do it as badly as others.  I've
encountered at least one, perhaps two 10GbE NICs which seem to at
least pass the TCP_RR sniff test and get both good TCP_RR performance
and good CPU util on TCP_STREAM with defaults.

rick jones
-- 
oxymoron n, commuter in a gas-guzzling luxury SUV with an American flag
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to post, OR email to rick.jones2 in hp.com but NOT BOTH...

_______________________________________________
questions mailing list
[email protected]
https://lists.ntp.org/mailman/listinfo/questions

Reply via email to