let me try to translate my mail to english...
* Henning Brauer [2007-09-07 12:49]:
traditionally, a network inteerface card interrupted once per packet ...
> received (or successfully sent, but let us ignore the tx path for now).
and let me explain why that makes so much of a difference. rocessing
the packets is one thing, context switching between the interrupt
handler (which takes teh packet from the nic's rx ring and puts it into
the ipintrq), the soft interrupt handling (where the packet is taken
from the ipintrq and running through ip_input, then maybe ip_forward and
_output and put into the (potentially other) nic's output queue, well,
that switching is not exactly free, performance wise, to put it nicely.
so instead of interrupt, one packet into ipintrq, processing at softnet,
oh wait interrupt, peel another out of the rx ring, etc yadda yadda it
is much cheaper to process a bunch of packets at once.
the downside is that you potentially increase latency, since after
packet reception teh nic does not interupt immediately and you don't
process immediately (for some value of immediately), but only after
enough othe rpackets are there. to cope with that effect the time I
mentioned above is there, so that a apcket doesn't sit in teh rx ring
for longer than XXX msec, wether otehrsd arrive or not.
--
Henning Brauer, hb@bsws.de, henning@openbsd.org
BS Web Services, http://bsws.de
Full-Service ISP - Secure Hosting, Mail and DNS Services
Dedicated Servers, Rootservers, Application Hosting - Hamburg & Amsterdam
| Andrew Morton | -mm merge plans for 2.6.23 |
| Benjamin Herrenschmidt | Re: [PATCH] Remove process freezer from suspend to RAM pathway |
| Greg KH | [GIT PATCH] driver core patches against 2.6.24 |
| Mel Gorman | [PATCH 6/8] x86_64 - Specify amount of kernel memory at boot time |
git: | |
| Jarek Poplawski | [PATCH] pkt_sched: Destroy gen estimators under rtnl_lock(). |
| David Miller | [GIT]: Networking |
| Gerrit Renker | [PATCH 15/37] dccp: Set per-connection CCIDs via socket options |
| Jarek Poplawski | Re: Soft-Lockup/Race in networking in 2.6.31-rc1+195 ( possibly?caused by netem) |
