Messages in this thread | | | From | Wander Lairson Costa <> | Date | Tue, 31 Oct 2023 08:36:14 -0300 | Subject | Re: [PATCH net-next 0/2] net: Use SMP threads for backlog NAPI (or optional). |
| |
On Tue, Oct 31, 2023 at 7:14 AM Sebastian Andrzej Siewior <bigeasy@linutronix.de> wrote: > > On 2023-10-16 16:53:39 [+0200], To Jakub Kicinski wrote: > > On 2023-10-16 07:17:56 [-0700], Jakub Kicinski wrote: > > > On Mon, 16 Oct 2023 11:53:21 +0200 Sebastian Andrzej Siewior wrote: > > > > > Do we have reason to believe nobody uses RPS? > > > > > > > > Not sure what you relate to. I would assume that RPS is used in general > > > > on actual devices and not on loopback where backlog is used. But it is > > > > just an assumption. > > > > The performance drop, which I observed with RPS and stress-ng --udp, is > > > > within the same range with threads and IPIs (based on memory). I can > > > > re-run the test and provide actual numbers if you want. > > > > > > I was asking about RPS because with your current series RPS processing > > > is forced into threads. IDK how well you can simulate the kind of > > > workload which requires RPS. I've seen it used mostly on proxyies > > > and gateways. For proxies Meta's experiments with threaded NAPI show > > > regressions across the board. So "force-threading" RPS will most likely > > > also cause regressions. > > > > Understood. > > > > Wandere/ Juri: Do you have any benchmark/ workload where you would see > > whether RPS with IPI (now) vs RPS (this patch) shows any regression? > > So I poked offlist other RH people and I've been told that they hardly > ever test RPS since the NICs these days have RSS in hardware.
Sorry, Juri is in PTO and I am just back from sick leave and still catching up. I've been contacting some QE people, but so far it is like you said, no stress test for RPS. If I have some news, I let you know.
> > Sebastian >
| |