On 08/05/2016 01:41 PM, Attila Kinali wrote:
Standard network cards will just trigger an IRQ at some point after
reception and enqueuing of the packet.
Perhaps drifting a bit...
In the broadest of handwaving terms, prior to Gigabit Ethernet, NICs
(Ethernet anyway) would post an interrupt for each arriving frame. In
the 100 Mbit/s Ethernet cards, (and perhaps FDDI and others) they
started avoiding transmit completion interrupts. With gigabit Ethernet,
they started avoiding receive interrupts through interrupt coalescing
settings. Different NICs did/do it differently, and some better than
others. Under Linux at least, one can use the ethtool utility to tweak
or even disable interrupt coalescing. That would likely be Just Fine
(tm) for a system dedicated to timekeeping, but the performance effects
on "normal" networking might be more than one desired.
With 10 Gigabit NICs, multiple receive queues come into play (again,
handwaving a bit). I would think that some of the later ones would be
sophisticated enough to enable sending NTP traffic through a specific
queue/IRQ but I don't know of any 10 GbE NICs with per-queue coalescing
settings. My experience with 10 Gbit/s NICs and the likes of say a
netperf TCP_RR test has been that except perhaps for the worst of them,
they will not arbitrarily delay a packet arriving at an "idle" time.
What *will* be a non-trivial bummer for such things as a netperf TCP_RR
test (think ping without "think time") will be power management in the
processor(s). Perhaps not visible in a MAN/WAN test setup, but
certainly visible in a LAN one.
rick jones
_______________________________________________
time-nuts mailing list -- [email protected]
To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts
and follow the instructions there.
