GCC 10's LTO Will Make Use Of Available CPU Cores By Default

[c117152]

There's a dozen different ways to work it out, some better than others, but when it comes down to it, having 8c/16threads available means dropping a thread would often have less overhead than most setups you can come up with.

I'm not saying linux is doing something wrong. I'm just saying I don't see the point in doing any of this tweaking unless it's done on the distro level / compile server. For an end user on an x86, the IPC is quite bad nearing full utilization anyhow so you might as well drop a thread and save yourself the trouble.

[tchiwam]

Maybe implementing a -l to match a load like with make would make sense ?

I often use -l instead of -j with make and it helps a lot on my laptop.

[skeevy420]

Maybe implementing a -l to match a load like with make would make sense ?

I often use -l instead of -j with make and it helps a lot on my laptop.

Is -l based on all CPUs or per CPU? Like would 100 be everything in a 4 thread system or would it be 400? man and --help isn't very helpful there.

[skeevy420]

There's a dozen different ways to work it out, some better than others, but when it comes down to it, having 8c/16threads available means dropping a thread would often have less overhead than most setups you can come up with.

I'm not saying linux is doing something wrong. I'm just saying I don't see the point in doing any of this tweaking unless it's done on the distro level / compile server. For an end user on an x86, the IPC is quite bad nearing full utilization anyhow so you might as well drop a thread and save yourself the trouble.

One thing I've been experimenting with lately is using the BMQ scheduler and negative nice values with GUI programs. BMQ allows end users to set up (down?) to negative 10 without having to edit any system settings which makes it ideal for desktop nice values. So far I've been manually setting random programs to use negative 5, games at negative 10, and they don't seem to be as effected by programs being compiled with nice 0 and -j16 on my 8c/16t system. I need to look into running all of Plasma/KDE as neg 5 to see if that has more positive effects on desktop smoothness and responsiveness in any ways.

[tchiwam]

Is -l based on all CPUs or per CPU? Like would 100 be everything in a 4 thread system or would it be 400? man and --help isn't very helpful there.

-l is based on the load of the system, if you have 8 logical cores (threads), you can say -l8 on 2 build projects and they will fight between them and the load might end up a little above 8. I usually do make -l (Nrealcpu-1) and leave it in the background.

Over simplified guideline :
1.0 load is one CPU at 100% load
...
8.0 load is 8 CPU at 100% load

[skeevy420]

-l is based on the load of the system, if you have 8 logical cores (threads), you can say -l8 on 2 build projects and they will fight between them and the load might end up a little above 8. I usually do make -l (Nrealcpu-1) and leave it in the background.

Over simplified guideline :
1.0 load is one CPU at 100% load
...
8.0 load is 8 CPU at 100% load

Thanks. man and --help weren't very descriptive of how the float worked.

[c117152]

One thing I've been experimenting with lately is using the BMQ scheduler and negative nice values with GUI programs. BMQ allows end users to set up (down?) to negative 10 without having to edit any system settings which makes it ideal for desktop nice values. So far I've been manually setting random programs to use negative 5, games at negative 10, and they don't seem to be as effected by programs being compiled with nice 0 and -j16 on my 8c/16t system. I need to look into running all of Plasma/KDE as neg 5 to see if that has more positive effects on desktop smoothness and responsiveness in any ways.

bmq? I'm still on 4.19 and haven't switched from cfq and deadline to mq-deadline on most boxes... A machine freezing is me getting another coffee break so I don't see a problem worth fixing

[tchiwam]

Thanks. man and --help weren't very descriptive of how the float worked.

Always best to just try it

[skeevy420]

Always best to just try it

To add to this, been experimenting with "nice -5 ionice -c 3 makepkg -sif" today and I really haven't been getting any slowdowns with firefox (nice --5) watching How the Universe Works on Hulu while doing a wine compile. Normally I notice a bit of scroll lag if I'm posing while doing both of those, when dropping Yakuake down, copying folders (copied KSP set it up with RSS+RO+Principia) and I haven't noticed any of the usual UI lag and slowdowns that I normally get. Crazy how noticeable of a difference that addendum to makepkg made.

Haven't edited makepkg.conf to test out -l yet due to all my nice experimenting.

[pal666]

it's unclear how it is better than -flto=`nproc`
and better teach poor ninja jobserver protocol and use -flto=jobserver like real men