LWN: Comments on "5.3 Merge window, part 2"

5.3 Merge window, part 2

neilbrown — Wed, 24 Jul 2019 00:43:53 +0000

> Or making it much worse by inducing bufferbloat.

Certainly possible. We would love to find a way to have the kernel auto-configure this number, but no brilliant ideas have come forward.

When you have parallel hardware (multiple cables, multiple DMA engines, multiple offload engines, multiple CPU cores), having multiple TCP connections can really help.
When you don't (or don't have enough), then multiple connections defeat congestion control and bloat buffers.

Determining the capabilities of the hardware (which could change) is, as yet, an unsolved challenge.

5.3 Merge window, part 2

Sesse — Tue, 23 Jul 2019 20:56:23 +0000

Or making it much worse by inducing bufferbloat.

5.3 Merge window, part 2

johill — Tue, 23 Jul 2019 20:13:33 +0000

I looked in qemu, and it sort of has this - if you emulate fully instead of using kvm, which is not feasible.

I think it *could* be done in qemu, remove the TSC, simulate an HPET that has the desired properties and use a virt-aware guest to do the idle handling? But it's tricky and probably requires touching not only qemu and the guest kernel, but quite possibly also the host kernel. Though then I wonder how virtualbox does it, maybe I should look at that.

Anyway, UML is good for me and actually simplifies my use case, so that's it for now. Seeing how much interest this has garnered, maybe somebody else will pick it up elsewhere :-)

Btw, I broke it in a build fix at the last minute, so you need https://patchwork.ozlabs.org/patch/1134739/ to actually use it.

For wifi tests in wpa_s/hostapd it cut down the time required from ~45m to ~7m on my machine, but I run with more kernel debug options now (which wasn't even possible before) in ~10m.

Next up is to add vhost-user support to UML (WIP) and then add a virtio-clock/idle driver to UML that uses it, so I can run multiple things (or instances of UML machines) on the same virtual time...

5.3 Merge window, part 2

re:fi.64 — Tue, 23 Jul 2019 19:23:35 +0000

Have you seen Mozilla's rr debugger as well? https://rr-project.org/

5.3 Merge window, part 2

unixbhaskar — Tue, 23 Jul 2019 08:27:29 +0000

"There is a new mount option (nconnect=) for NFSv4 filesystems that allows the administrator to request that multiple TCP connections be established to the server. Traffic is then load-balanced across those connections. "

This is certainly getting over the lagging issue.

5.3 Merge window, part 2

yann.morin.1998 — Mon, 22 Jul 2019 20:13:32 +0000

Hello,

> Why not init to 0xdeadbeef1337babe [...]

I think the following, from the referenced commit, does already address your concerns:

> If either SLUB poisoning or page poisoning is enabled, those options take
> precedence over init_on_alloc and init_on_free: initialization is only
> applied to unpoisoned allocations.

Regards,
Yann E. MORIN.

5.3 Merge window, part 2

Cyberax — Mon, 22 Jul 2019 19:25:44 +0000

Yes, this is super-nice.

You can also do that with VirtualBox (by specifying the "warp factor" via --warp-pct) but I'm not aware of any other virtualization environment that can do this.

5.3 Merge window, part 2

smurf — Mon, 22 Jul 2019 18:43:01 +0000

Ah. Init to zero. Why not init to 0xdeadbeef1337babe, or something along these lines? that could at least find us some bugs instead of papering them over. (Yes I know, that's not helpful for Android users – but it'd be helpful for kernel developers.)

5.3 Merge window, part 2

mtaht — Mon, 22 Jul 2019 18:11:42 +0000

Time travel mode looks potentially *wonderful* for finding bugs in long running daemons. There was a dhcpd6 once that only happened after 51 days of uptime, and I've hit numerous other bugs like that in my career.

+10!