sigh... src.illumos.org/source/xref/illumos…/gssd_pname_to_uid.c?r=b9175c69#288

what's up?

should nfs and smb performance be more or less equal when sharing the same dataset?

i'm getting like max 35MB/s over SMB but hitting like 400MB/s over NFS

client is a bhyve vm

i would expect it to be closer?

smb does some internal windowing, depending on the client

shouldn't matter much when the latency is very low

dumb clients do 16k ping pong

I'm running Hipster as a VM guest and trying to find documentation on how to adjust the display resolution above the 1280x720 maximum right now. Any pointers would be helpful. Thank you.

I fiddled a bit on the client side with the options but can't get it faster at all -_- guess I'll need to mess around with nfs4_acls to get the permisisons on NFS OK :(

tsoome: so if namebuf.value == NULL, s becomes an uninitialized pointer?

yes

well, it is uninitialized before we set it:)

hmm, is that case possible? I mean it should be initialized regardless, I'm just wondering if that's a possible corner case being hit

I mean I guess it is if it's checked and conditionally assigned, hah

I'd hope this is userspace code at least?

yes it is

How far back does this go? illumos.org/man/8/ksslcfg I'm messing around with nginx kTLS ssl_sendfile()

copec: I think we retired that actually? illumos.org/issues/14767

→ BUG 14767: retire kssl (Closed) | code.illumos.org/c/illumos-gate/+/2207

Yeah, it was looking old. Is implementing ssl_sendfile() on anyone's list?

Not that I know of

If I had a business riding on it, I would do it / pay to have it done. No complaints here. I've just been slowly watching the kTLS stuff matriculate into mainstream and was curious.

A business riding on TLS throughput specifically?

Sure, at first I saw stuff from Netflix about their cache boxes, but now I see stuff from all sorts

You have NVMe drives and want to saturate your network connections, that copy from kernel to userspace and back can make or break doing that

youtu.be/36qZYL5RlgY?t=425 Is recent

Of course Netflix is probably the most extreme case, but since everything is TLS based these days, kTLS becoming a standard thing to do is inevitable

this is something funny: src.illumos.org/source/xref/illumos…ote.c?r=bbf21555&mo=5209&fi=203#236

good grief

tsoome: I'm glad your on a system-wide source trawl and get to see the horrors, too

it feels like I have company

:D

copec: I think that assumes, still, bulk TLS transfer of data that requires no transformation; i.e., static file service

that is an awful lot of stuff

tsoome: ouch

I also don't really understand how it interacts with, say, wanting to reuse a single connection for many things like you do with some of the more advanced HTTP stuff

that I don't know

I get that it helps Netflix, but they also control both the server and the client, and have a pretty specific need

but I think we have a long history of ignoring stuff like this because it's not some magically general solution

and probably shouldn't just go "eh" off the bat

To be clear I'm super inf

in favour of someone to whom it is relevant, working on it

I just believe understand why nobody is

(at least visibly) doing so

e

andyf: did we ever convince your relative(?) to make interrupting wsdiff work properly?

there are whole families working on this?

lol

tsoome - that's horrific

richlowe - I know that an attempt was made, but I don't think it was successful