Does 7-zip benchmark have somekind of affinity control ?

[SvenBent]

im currenlty sitting at a Intel COre i7 920 system (4C/8T)
i run 7-zip 16.02 benchmark 4 threads 32MB

With normal full affinity i get 50% CPU usage in task manager as expected.
If i set affinity to ever other cores then CPU usage drops til 25-33% sometimes even below 20%.
In my head the 4 thread still gets access to 4 cores and should still be at 50% usage in task manager.

So does 7-zip have some form of internal affinity or thread controls that conflicts with the windows affinity control ?

[Bulat Ziganshin]

cpu usage is computed as non-idle time of each hardware thread, summed up and divided by 8. so you can easily load the entire cpu by running 4 threads, but it will be counted as 50% usage by Program Manager

[SvenBent]

I'm aware of that all depending on thread utilization/efficiency for the cores and that taskmanager basically just show time scheduling of the CPU not really utilization, but that is not the subject here.

Having 4 CPU have threads show show 50% CPU usage in taskmanager with 8 (logical) cores
It shouldn't matter which of the 8 logical cores the threads are target to in regards to the CPU time scheduler. (Performance and physical core utilization is a different story)
Unless 7-zip does something with its threads to place it on specific logical cores it should not go down when i gives its 4 heavy threads access to only 4 logical cores instead of 8.
This is a new behavior as i used to use 7-zip to benchmark the increase in performance, from going from full affinity to every other Logical affinity on a CPU with few physical cores than logical cores.
However now the performance absolutely drops.

Bah i just tried to take some screenshots but now its back to "normal" and its getting the 10% boost from "Disabling Hyper threading" by affinity, and the usage it the same full affinity or reduced affinity.

[SvenBent]

now it does it again hmm

"full affinity.png" show the taskmanger reported CPU usage. its 50% when its at full affinity
"half affinity.png" show how it drops to 25% when i set 7-zip to run on 4 cores (out of .

It should still be 50%


my guestimate is that 7-zip doessome kind of internal affinity like adjustments

so lets say 7-zi[ puts its 4 threads to to core 0 1 2 3 4
hen i go in and give set affinity to only 1 3 5 7.
since out of its own chossen cores, 7-zip now only gets access to 1 & 3, so only 2 cores get utilized and it drops to 25% total utilization

[SvenBent]

A bit more info or "evidence"

running 7-zip with full affinity and selecting 4 threads. we should still see a kinda even utilization across the cores over times due to Windows modified round robin distribution.
but 7-zip seems to avoid 2 of the 8 cores entirely.

-- edit --
and it seems to change it from benchmark iteration to iteration

[SvenBent]

hmm Linpack show same behevaier how ever it uses 8 threades when loading/setting up its data array and swithched to 4 cores duringg crunhing.
It only discard 1 core so its 50% is sprread out on 7 logical cores.

if i set affinity to every other cores except ( not the discared one) it still stay on 50%
but if i reverse it so it gets every other one includign the discard one it drops to 33% usage which agian indicates there is somekind of blockage in the software to use the "discarded core" outside of the normal affinity control.

The same holds true if i monitor which cores they "Discard" and work affinity settings around it the utilization stays up

In both 7-zip or linpack case they dont discard cores when running 8(+)threads.

Its just seems like a weird behavior and im now no longer sure it 7-zip/linpack related or if its windows or HW platform

[SvenBent]

I just downloaded cinemark R15 becaise its yet another program that scales really well across cores and have an easy way to control threads
I set it to 4 threades, and gave it affinity to core 1 3 5 7 on my core i920 (4C/8T)

Instantly the CPU utilization dropped from 50% to 25 and lo and behold. only core 1 and 3 is being utilized at all
see first image (Cinemark4T4C.png)


The second image (Cinemark4T8C.png) show the CPU utilization when its running 4 threads full affinity.
the 4 threads get distributed around on the 7 of the 8 cores where as the last core has not utilization.
I started up pngout to compress an image and then 8th core would get some utilization so its not a system wide "discard". its locally to the multi threaded software.
but i dont think its the software itsellf since both 7-zip linpack and cinebench shows this weird behavioer


-- edit --
rinning 5 threads or 6 threads does not "discard a core", there is utilization on all 8 cores over time

[SvenBent]

in case somebody have the same issue. the reason is core parking.

its amazing that windows decide to keep a core parked even though you assign a CPU heavy thread to it by affinity.