I'm sorry. I was a little confused and jumped the gun on answering your
questions. I totally understand fixing processes to a specific CPU in that
scenerio. If you type in your CPU model i7 950 an Intel link called Intel
Arc will give you a bunch of the information I asked about like
hyperthreading and turbo boost. HT was brought back with the i7 920+ .
Maybe the kernel configuration is different between the two.
If you think any of my information might spark conversation (I'm pretty
sure I messed a few details up) feel free to post it. My ideas are GPL :)
On May 12, 2013 12:28 PM, "Britton Dodd" <brittman...@gmail.com> wrote:
>
> On May 12, 2013, at 12:15 PM, Thorsten Glaser wrote:
>
> Britton Dodd dixit:
>
> What virtualization subsystem are you running? I have a i7 920 at home
>
> and it's pretty fast.
>
>
> ARAnyM on bare metal Debian/i386.
>
> The only thing I can think of is maybe your running into a scenario
>
> where your VMs are pinned to threads that share the same CPU Core, and
>
> that might be inducing a bottle-neck somewhere.
>
>
> Hm. I can try pinning one on cpu#7 only, or not pinning them at all.
>
> (15 minutes later)
>
> Nope, I get still about 47 BogoMIPS for both the one pinned on cpu#7
> (the other running VMs are pinned to cpu#0 #1 #2 and #3) and the one
> unpinned, while the system was half idle.
>
> As an “ouch” on the side: the other VMs have ~70 BogoMIPS.
>
> Just out of curiosity, why pin the VM threads? The scheduler from what
>
> I've seen does a pretty good job with managing thread usage. Of
>
>
> There are no threads. ARAnyM is a single CPU (plus chips) which is
> emulated by a single user-space process, and not letting that flip
> between CPUs increases system performance.
>
> monitoring as well. And if 2 VMs are pinned to logical CPU0 and CPU1,
>
> they *might* be fighting for CPU time because of Intel HyperThreading.
>
>
> Does Intel still use HT? I thought it was a bad idea-thing of the past?
> Does the CPU I have use it?
>
> But on the other hand, if you see above, even if I pin it to just cpu#7
> I get the bad performance. Also, if I pin it on cpu#6 (the “other half”
> if I got the info right) I get the same (about 45-46 BogoMIPS).
>
> /sys/devices/system/cpu/cpu{0-7}/cpufreq/scaling_governor = performance
>
> In /proc/cpuinfo I also see the MHz up on the max. I’m out of ideas.
>
> And Turbo Boost *may* also be playing a part in this too..although
>
> I've rarely *if ever* seen it kick in on Linux…especially if your CPU
>
>
> I have no idea what that is…
>
> isn't exceeding the maximum TDP (thermal design power). If a CPU core
>
> or two needs extra speed, the i7 will temporarily adjust the
>
> multiplier up a couple of points on-the-fly, as long as heat output
>
> doesn't exceed the maximum.
>
>
> … or how to check for that.
>
> Other things like AHCI mode for SATA vs IDE emulation on SATA make a
>
> huge difference with respect to I/O too.
>
>
> Erm. I use NatFeat disc access. But this is purely about vCPU
> performance.
>
> I hope none of this offended you -- I'm just trying to help, as I
>
> really admire what you guys are doing with 68k equipment.
>
>
> No, not a problem. One thing I’d like to do is to take this back
> to the mailing list, so maybe others know. If that’s ok for you
> I’ll bounce these messages too (I won’t publish private mails
> without asking, normally).
>
> Thanks,
> //mirabilos
> --
> “Having a smoking section in a restaurant is like having
> a peeing section in a swimming pool.”
> -- Edward Burr
>
>
>
