Subject: Re: performance drop with alpine linux upgrade
To: alice <alice@ayaya.dev>, alpine-user <alpine-user@lists.alpinelinux.org>
References: <196e1b55-ade9-8a4b-889b-006fef7f63c1@libero.it>
 <CMUCC7OZ4Y9K.10UWLB8RC8NP8@sumire>
 <9cb23b83-bf70-dd44-6249-a69e30e6f3e0@libero.it>
 <CMUG2DUUFPDZ.21TDKCEIN8UCS@sumire>
From: Riccardo Mottola <riccardo.mottola@libero.it>
Message-ID: <7e8bd9af-3697-ef04-91e6-8b04324f5d3f@libero.it>
Date: Tue, 20 Sep 2022 13:16:34 +0200
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101
 Firefox/68.0 SeaMonkey/2.53.13
MIME-Version: 1.0
In-Reply-To: <CMUG2DUUFPDZ.21TDKCEIN8UCS@sumire>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit

Hello Alice,

I discovered something unexpected. The impact of mitgations is extreme!


alice wrote:

> maybe there's a chance that if intel_pstate loads first it blocks the
> others? (just guessing). both of the others are also available as
> modules, maybe something funny happens with pstate. perhaps
> modprobe.blacklist=intel_pstate does something (or similar things in
> that vein)

I tried playing with that, no relevant changes. The module wasn't even
loaded (I suppose the error is about that).

>> I cannot identify this CPU here:
>> https://en.wikipedia.org/wiki/List_of_Intel_Celeron_processors#Netburst_based_Celerons
>>
>> since 2.2Ghz appears to have 128K cache. It could be this one:
>> https://ark.intel.com/content/www/us/en/ark/products/27318/mobile-intel-celeron-processor-2-20-ghz-256k-cache-400-mhz-fsb.html
>>
>> still i don't think too much relevant to have the exact detail.
>>
>>
>>> add mitigations=off to the kernel cmdline, via whatever booting setup
>>> you have. usual warnings/caveats apply, but there you go. i doubt they
>>> would be that impactful though..
>> I doubt too, on more modern CPUs I have found an impact of 1% to 10%
>> usually... but maybe on this old cache-limited CPU?
>>
>> I have syslinux, I don't know hot do add that option to
>> |/etc/update-extlinux.conf|, but I will add it manually for a boot and
>> test the difference.
> the default_kernel_opts line can have it


That was it. mitigations=off makes the system almost as fast as before,
it takes now about 680minutes instead of... 7000!! (still slower than
the previous 500 minutes, but also gcc was upgraded, so other things
could take in account the slowdown).

It means one or more of the disabled mitgations has an extreme impact
(roughtly 1/10th performance) in the specific task of heavy compilation,
python and make scripts.
Being a celeron, this is a non-hyper-threaded CPU, but it was one of the
first CPUs with a very deep pipeline and high clock speed (2.2GHz).
Maybe some mitigations aren't even needed, however it would be
interesting to know which one(s) makes this huge impact, could be useful
for others. But to do that I need somehow a smaller testcase, I can't
wait a night vs. a fortnight everytime :)

Riccardo