Re: [PATCH libgcrypt] Disable CPU speculation-related misfeatures

On 05/07/2025 14:37, Jussi Kivilinna wrote:
> On 14/06/2025 15:47, Guido Trentalancia via Gcrypt-devel wrote:
>> Three runs of timed "make check" (with suppressed output) on libgcrypt
>> version 1.11.1, with and without the safety feature enabled, provide
>> the following meaningful results:
>>
>> safe (avg): 54.26 seconds
>>
>> unsafe (avg): 52.94 seconds
>>
>> Therefore, the performance loss of safe versus unsafe is 2.5%.
> 
> Seems high given that "make check" is mostly single threaded test code
> running mostly in user-space and checking libgcrypt implementation.
> 
> Changing these prctl settings will affect the application where libgcrypt
> was loaded (directly or indirectly) and can also affect performance
> of other processes running on system (see discussion & push back on
> L1d flush change in linux kernel).
> 

Which means that just benchmarking libgcrypt test suite does not quite
cut it when estimating performance loss. Do you have break down how much
each prctl did affect performance individually?

-Jussi