Score:0

Is there any hardware requirements for Linux kTLS in nginx?

in flag

I've followed closely the manual on NGINX blog. This works flawlessly on my machine, with an AMD Ryzen PRO 4750G. The output for a 5GB file using curl on https://localhost is twice as fast with kTLS, which is neat.

I've duplicated the same configuration options on my server that has a much less powerful CPU (Intel Atom D2550), but the system doesn't want to use kTLS, there is nothing in the debug log. So I'm wondering if there is some hardware requirements in the CPU, undocumented? The kernel documentation doesn't seem to mention any.

Score:1
br flag

If I understand your question properly then the issue isn't with the CPU but with the NIC in your systems - the NIC itself needs to support kTLS offload, not all do.

Adrien Clerc avatar
in flag
Yes, maybe, that's also an option. However, in my "local" test, it doesn't use the NIC, since it's on localhost, so it can only be bound to CPU features.
br flag
Well if it's on localhost then the CPU's performance becomes incredibly important - and a 4750G is a hell of a lot more performant than an Atom D2550 - that's the issue.
Adrien Clerc avatar
in flag
Yes, but if it wasn't clear enough: without kTLS on localhost, I got ~750MB/s. With kTLS on localhost, I've got 1200MB/s, and the debug log shows that it's used. It's not the case on my Atom, and I'm looking for a reason. But I'm about to test on my AMD 4750G with the real NIC (and not localhost)
Adrien Clerc avatar
in flag
I tested a connection from another client to my computer, and indeed, the kTLS benefit is here: the CPU is much less used on my AMD 4750G with kTLS.
I sit in a Tesla and translated this thread with Ai:

mangohost

Post an answer

Most people don’t grasp that asking a lot of questions unlocks learning and improves interpersonal bonding. In Alison’s studies, for example, though people could accurately recall how many questions had been asked in their conversations, they didn’t intuit the link between questions and liking. Across four studies, in which participants were engaged in conversations themselves or read transcripts of others’ conversations, people tended not to realize that question asking would influence—or had influenced—the level of amity between the conversationalists.