-
gitomat
[illumos-gate] 17101 os-tests: unchecked function return 'fflush' -- Toomas Soome <tsoome⊙mc>
-
gitomat
[illumos-gate] 17102 rootnex: ignoring unreachable code. -- Toomas Soome <tsoome⊙mc>
-
gitomat
[illumos-gate] 17103 testoplock: unchecked function return -- Toomas Soome <tsoome⊙mc>
-
vetal
Hey)
-
vetal
Does current Mellanox Connect-X driver support RDMA?
-
jbk
AFAIK, none of the mellanox cards support RDMA on illumos
-
vetal
jbk: How do you think, is it hard to add RDAM support ? BTW, what was initial source for the drivers where it was ported?
-
dangergrrl
vetal, it probably depends on the expertise of the person adding it :) doesn't sound that difficult in concept
-
sommerfeld
RDMA is one of those networking features that has to punch through many layers of the stack to get bits from hardware to the application so it ends up likely as a more involved integration than most.
-
jbk
IIRC, the ib support we have is fairly old, but i don't think the rdma bits for it are coupled too tightly, but so it might be able to leverage the existing bits, but not sure how much work might be needed to get all of that up to more recent versions of any of the relevant standards
-
jbk
i'm not sure what source was used for the original drivers (and the person that did the work isn't online)
-
jbk
prior to nvidia buying mellanox, the programming guide was however freely available
-
jbk
so it wouldn't surprise me if that was the basis
-
jbk
I'm not sure how difficult it would be to get an updated one from nvidia
-
jbk
i suspect there's probably some sort of NDA involved, but might be able to just glean the differences by looking at say the freebsd drivers to see what incremental bits were needed to support the newer cards
-
jbk
my recollection on cx-4 -> cx-5 -> cx-6 was that it was mostly additive and there weren't too many radical changes (e.g. maybe a couple of new registers for the newer features/speeds, but most of the base functionality was identical)
-
rmustacc
There was no 'initial source'. The driver was written as its own thing.
-
rmustacc
Other things are used as references, but it's not like there's some common code or anything.
-
jbk
i'm assuming a programming guide was the source, but wasn't sure
-
jbk
(source for how it worked)
-
rmustacc
Yes, I just took vetal's thing to be code source.
-
rmustacc
And even the public one then was out of date.
-
richlowe
I feel like I've missed something
-
jclulow
richlowe: mellanox NIC driver stuff
-
richlowe
I feel like I'm happy I've missed something
-
jclulow
I usually am!
-
danmcd
Anyone got cycles to bringup an illumos driver for this?
-
danmcd
-
danmcd
(Bonus points if you can do some HW engineering and create a 4-chip/4-port version of this using a PCIe 4 x4 slot)
-
richlowe
I wonder what it is
-
richlowe
it seems like nobody new would enter the 10G asic market
-
richlowe
Marvel
-
rmustacc
Yeah, it's actually a much older part that I didn't realize was being refreshed.
-
danmcd
I wish these 10GBaseT parts would get older models shrunk/made-cheaper. I'm thinking HDC4 needs 4 10GBaseT ports that do 1/2.5/5/10. Kinda like the Intel X550-series but, well, cheaper. (I see the X722 has disappeared from most folks' radars and modern mobos that go there use X710... curious. :) )
-
rmustacc
It's because they dropped it from the PCH.
-
gitomat
[illumos-gate] 16906 bhyve should support AMD perf counters -- Patrick Mooney <pmooney⊙pc>