Sepherosa Ziehau published a PDF that shows the impact of Spectre and Meltdown mitigations on network performance. It’s as bad as anticipated. If this is a problem for you, remember the sysctl machdep.isolated_user_pmap turns these changes on and off.
9 Replies to “HTTP, Spectre, Meltdown”
Comments are closed.
Too bad that with all of these perf benchmarks occuring due to Spectre/Meltdown, no one is updating the painfully dated Dfly benchmarks
https://www.dragonflybsd.org/performance/
I and others have pointed this out multiple times – requests on random blog posts for “someone” to do a bunch of work isn’t going to do any good. I’ll delete these requests in the future, since you either aren’t reading or you don’t understand.
Ouch. All the work Sepherosa has done to improve networking and it gets wiped out, if not completely, at least quite a bit.
And it’s not just DragonFly – it’s all x86, everywhere.
Thanks Justin, I know that. I just seem to see Sephe’s work improving things reported here first. The results for this specific use case seem to mirror some of what Matt has reported in the past (at least in trend if not in specific number).
Is Spectre / Meltdown the reason why this site is slow on mobile devices?
From my iPhone, it’s painfully slow.
If your box is secured, e.g. your own web server and only you can login, I’d suggest to turn off both. However, you _do_ want to leave them to their default values for your desktops and laptops, etc.
Seph
Good to know!
Thanks :)
It seems the mitigations for Spectre have greater impact than for Meltdown, which I find surprising since Meltdown is a more glaring hardware bug. As different hardware reportedly experiences greater or lesser effects from the mitigations, it would be nice to describe the testing hardware more clearly. All in all, I view this as more a strike against the security promises of hyperconvergence and cloud than anything else.