Hi all,

On 6/13/20 12:24 PM, Patrick Erdmann wrote:
> Just to be sure, try this line as kernel parameters:
>
> noibrs noibpb nopti nospectre_v2 nospectre_v1 l1tf=off
> nospec_store_bypass_disable no_stf_barrier mds=off tsx=on
> tsx_async_abort=off mitigations=off

I tried all "noibrs noibpb nopti nospectre_v2 nospectre_v1 l1tf=off
nospec_store_bypass_disable no_stf_barrier mds=off tsx=on
tsx_async_abort=off" and here some results.

Laptop with i7 (4x parallel build):

ascii "standard": 30-31 minutes
beowulf "standard": 41 minutes
beowulf "fast": 34 minutes


Laptop with T7200 (2x parallel build):

ascii "standard": 63 minutes
ascii "fast": 61 minutes


From this we see that at least some mitigations were already "slowing
down" ascii already and at least on another CPU I gained some speed
back, but this makes the comparison even with beowulf!

We see that "fast" still is slower than previous ascii by about 10%
(which we can estimate do be more, since I did not test the same
configuration).

Something is still slowing down... it is not that my life changes with
3minutes, but I hate to know tht just a plain "upgrade" robbed me 10-15%
of speed. By the way, informal tests (not on the same hardware) show
that this speed gain is the difference between an i5 and an i7, so like
to say my CPU was downgraded :-P


What could be slowing down? some other kernel "settings" not related to
mitigations.. cache handling, file system, scheduling?


Riccardo