[Bug 274994] Regression of iperf3 network throughput tests with erms "rep movsb" copyto loops

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=274994

            Bug ID: 274994
           Summary: Regression of iperf3 network throughput tests with
                    erms "rep movsb" copyto loops
           Product: Base System
           Version: CURRENT
          Hardware: amd64
                OS: Any
            Status: New
          Severity: Affects Only Me
          Priority: ---
         Component: kern
          Assignee: bugs@FreeBSD.org
          Reporter: rscheff@freebsd.org

Initially, the thinking was, that a missing patch 6210ac95a194 was responsible,
but even with 
that patch, the below results are reproducable;

The exclusion of the erms codepath was performed by patching copyout.c like so
- thus only
the copyin smap_std or copyin_nosmap_std would be used; however, smap is
disabled elsewhere too,
thus this collapses to a comparison between copyin_nosmap_std and
copyin_nosmap_erms:


>         switch (cpu_stdext_feature & (CPUID_STDEXT_SMAP | CPUID_STDEXT_ERMS)) {
>         case CPUID_STDEXT_SMAP:
>                 return (copyin_smap_std);
> + #if 0
>         case CPUID_STDEXT_ERMS:
>                 return (copyin_nosmap_erms);
>         case CPUID_STDEXT_SMAP | CPUID_STDEXT_ERMS:
>                 return (copyin_smap_erms);
> + #else 
> +       case CPUID_STDEXT_SMAP | CPUID_STDEXT_ERMS:
> +               return (copyin_smap_std);
>   #endif 
>         default:
>                 return (copyin_nosmap_std);
>         }




On Sapphire Rapids
Xeon Gold 5416S (processor ID 0x806f7) (4th Gen)
with erms "rep movsb":

> - - - - - - - - - - - - - - - - - - - - - - - - -
> [ ID] Interval           Transfer     Bitrate         Retr
> [  4]   0.00-20.00  sec  83.6 GBytes  35.9 Gbits/sec    0             sender
> [  4]   0.00-20.04  sec  83.6 GBytes  35.9 Gbits/sec                  receiver

replacing the erms with no-smap,no-erms copyto:

> - - - - - - - - - - - - - - - - - - - - - - - - -
> [ ID] Interval           Transfer     Bitrate         Retr
> [  4]   0.00-20.00  sec  86.7 GBytes  37.2 Gbits/sec  147             sender
> [  4]   0.00-20.04  sec  86.7 GBytes  37.2 Gbits/sec                  receiver


Xeon Gold 6438N (processor ID 0x806f7) (4th Gen) 
with erms "rep movsb":
- - - - - - - - - - - - - - - - - - - - - - - - -
> [ ID] Interval           Transfer     Bitrate         Retr
> [  4]   0.00-20.00  sec  1.55 GBytes   664 Mbits/sec    0             sender
> [  4]   0.00-20.05  sec  1.54 GBytes   661 Mbits/sec                  receiver

replacing the erms with no-smap,no-erms copyto:

> - - - - - - - - - - - - - - - - - - - - - - - - -
> [ ID] Interval           Transfer     Bitrate         Retr
> [  4]   0.00-20.00  sec  91.9 GBytes  39.5 Gbits/sec    0             sender
> [  4]   0.00-20.05  sec  91.9 GBytes  39.4 Gbits/sec                  receiver

On Ice Lake
Xeon Platinum 8352Y (processor ID 0x606a6) (3rd Gen)

with erms "rep movsb":
> - - - - - - - - - - - - - - - - - - - - - - - - -
> [ ID] Interval           Transfer     Bitrate         Retr
> [  4]   0.00-20.00  sec  33.6 GBytes  14.5 Gbits/sec    0             sender
> [  4]   0.00-20.04  sec  33.6 GBytes  14.4 Gbits/sec                  receiver

replacing the erms with no-smap,no-erms copyto:

> - - - - - - - - - - - - - - - - - - - - - - - - -
> [ ID] Interval           Transfer     Bitrate         Retr
> [  4]   0.00-20.00  sec  76.1 GBytes  32.7 Gbits/sec    0             sender
> [  4]   0.00-20.03  sec  76.1 GBytes  32.6 Gbits/sec                  receiver
> 


Similar performance degradation was observed on many older platforms using the
erms codepath, and the iperf3 testing utility, with typical performance impact
between 40-60%.

-- 
You are receiving this mail because:
You are the assignee for the bug.