aside from running iperf back and forth, anyone think of any 'fun' things to try to really kick the tires for a nic driver?

multiple parallel iperf runs while doing multiple rsyncs (mix of small and large files)? Maybe toss some NFS traffic in for variety.

jbk: I have recently discovered that rge can stall out and not recover in a way that I've only noticed when doing line rate IP forwarding that comes in the interface and goes back out the interface

So

Maybe try to use the thing as a router

iperf does not seem to excite the same issue, I suspect being an in-kernel hairpin like that just makes for tighter timing etc

..... and after a while it stops passing unicast traffic (but broadcast traffice seems to get through)

i f$#@$@#$ hate this nic

jbk: I feel your pain

it's just so over complicated

and there is no hell hot enough for whoever decided on the design for their DMA engine

oh f me... the replacement drive is SMR.... no wonder why it's taken > 24 hours to resilver a bit over 1T of data

I'm sorry

yay.. managed to make it overnight without panicing (and still passing traffic)

oof (re: the SMR drive).

yeah :( lesson learned

put another way: fuck u *shingles ur magnetic recording*

the problem is there's no widely used filesystems (AFAIK) that really match how you want to use SMR

which basically (IIUC) your drive is divided into a number of regions, and basically you want to treat each region as 'append only'

and 'reset' the region to start over from its start when you want to rewrite it

so... use it like a tape drive?

i suppose sort of

i mean, if they're going to be around for a while, it might be worth thinking about such a fs...

though IIUC, there are _some_ similarities that could apply to flash as well (just apparently most flash devices have hw in front of the cells that do a better job of managing it on behalf of whatever is using it)

jbk: I'd think that something like the BSD log-structured file system might work ok with SMR if you could size the log segments to fit the SMR geometry.

openwall.com/lists/oss-security/2026/08/16/5 - AI can certainly build mile high walls of text no human brain can ever penetrate...

Well, metaslabs are actually pretty close.

I think one of them is saying that ZFS user authorizations do whoami such that you can direct them via namespacing?

but I'm far from sure

rmustacc: IIUC though, metaslabs use the 'free' space for accounting (shrinking as the metaslab utilization grows)... which would be less than ideal

(I suppose another problem is if you want to use mirror or raidz, the zone sizes would need to line up for it to work well..)