[c-nsp] RAM thing

[Saku Ytti]

On (2014-02-17 23:47 +0100), Lukas Tribus wrote:

> I find that hard to believe. All routers, even old ISR G1s have ECC to detect
> such failures. Depending on the platform and configuration, those boxes may
> crash/failover due to parity errors, right?

Sure about that? TX Ring and RX Ring have ECC memory? Do you know full path
from ingress to egress and know for a fact that each place is ECC?

> Also on hardware based platforms I doubt the RAM ever sees a packet.

It's going to be stored somewhere, cut-through is never solution you can use
only. Where it's stored, how is is it copied from there from RX ring.

> I don't have the impression that silent corruption of traffic is what this
> is about. RAMs fail all the time, but our boxes do detect it (or at least
> crash). Of course, hardware malfunctions can lead to packet corruption, but
> thats uncommon compared to RAM failure I believe (otherwise we would see a
> *lot* of payload corruption on the Internet which is really not the case).

If you run MPLS network, monitor your egress PE core facing interfaces for IP
header errors. If you see them, where did they come?
They didn't come from client PC, your ingress PE would have seen and dropped.
Yet you've bound to find egress PE with IP header errors, your core is
corrupting them. It may not happen often. But when you do see it, you're only
seeing them, because it by chance happened in area you're aware of, so it's
certainly happening more, since you have limited view on corruption.

-- 
  ++ytti