[HN Gopher] AMD Zen 4 is faster with CPU security mitigations en...
       ___________________________________________________________________
        
       AMD Zen 4 is faster with CPU security mitigations enabled on Linux
        
       Author : simjue
       Score  : 83 points
       Date   : 2022-09-30 18:51 UTC (4 hours ago)
        
 (HTM) web link (www.phoronix.com)
 (TXT) w3m dump (www.phoronix.com)
        
       | brian_herman wrote:
       | Neat!
        
       | RobLach wrote:
       | That's a solid way to encourage better security practices :P
        
       | zazaulola wrote:
       | A very subjective opinion. CPU performance depends on many
       | factors. The article does not give the results of `perf`
       | performance tests. There are no given kernel parameters either at
       | compile time or set late `sysctl` parameters. How is the RAM
       | allocated to the CPU cores? How are IRQs used? Does RAM have
       | support for error correction? The only fact the article claims is
       | that by setting the `mitigations=off` flag in the system
       | configuration there is some misbalance which can make certain
       | tasks execute slower under certain conditions.
       | 
       | Pardon my English, it's not my native language.
        
         | rektide wrote:
         | I agree there could be more details about the specific system
         | used.
         | 
         | At the same time, I'd take this as more valuable & informative
         | than not, to be a reasonable indicator of what most users might
         | expect. There are quite a large number of different tests
         | features here that give quite good representation.
         | 
         | A lot of your questions don't seem concerning to me. Perf is
         | just one test, and we have plenty of other great real world
         | tests here. Kernel parameters are set at boot, and those
         | arguments are shown. RAM isn't allocated to CPUs: there is a
         | single I/O Die which all ram is connected to & cores access it
         | uniformly over Infinity Fabric. There's no tuning to IRQs
         | implied but if you have similar hardware it should reproduce
         | without fiddling, and more so, the IRQs should be working
         | similarly in both cases; the default config is more than good
         | enough to compare two cases with. I would not expect ECC to
         | make a vast difference, and it's notable that some level of ECC
         | already is built in to all DDR5 anyhow ("ECC" is now a matter
         | of whether the ECC extends to the CPU or is on-chip only). None
         | of these points seem strong enough that I'd consider discarding
         | or ignoring these findings.
         | 
         | This article has a pretty good starting place that paints a
         | pretty good general picture. It's a service & contribution &
         | informs nicely. For users who have more specific concerns, they
         | should follow up, and ideally, do as good a favor as Mr Larabel
         | did & blog their findings.
        
         | guerby wrote:
         | From the full result links it looks like it's ubuntu 22.04 with
         | just one kernel parameter added.
         | 
         | https://openbenchmarking.org/result/2209265-NE-ZEN4MITIG74
         | 
         | https://openbenchmarking.org/result/2209265-NE-ZEN4MITIG74&e...
        
         | sleepycatgirl wrote:
         | You don't have to apologize, your comment is very much well
         | written.
        
         | Delk wrote:
         | That was kind of what sprung to my mind as well. Since I'm not
         | really familiar with the mitigations or their low-level
         | details, I wouldn't be able to speculate on _what_ kinds of
         | factors might cause the performance effects to vary randomly
         | depending on specific circumstances, though.
         | 
         | Kind of like it turns out that the performance effects of
         | various compiler optimizations can vary a lot depending on the
         | exact memory layout of the compiled program, which might vary
         | from one compiler version to another or due to minor changes to
         | the program code. If you want generalizable results, the proper
         | way to benchmark might be to test with multiple random layouts
         | [1].
         | 
         | I don't know if there could be similar potential side effects
         | from vulnerability mitigations that might cause their
         | performance effects to vary in an effectively random way.
         | Perhaps the mitigations are different enough to not vary so
         | much. I don't really know. But the question kind of unavoidably
         | comes to mind.
         | 
         | [1] https://emeryberger.com/research/stabilizer/
        
         | brnt wrote:
         | > native language
         | 
         | (There is no such thing!)
        
       ___________________________________________________________________
       (page generated 2022-09-30 23:01 UTC)