Swap4 SSE2/AVX2 benchmark : CPU Cache / RAM bandwidth

[Igor Pavlov]

The compiled code is different for linux and windows,
But crc code is written in asm. So it's same for linux and windows versions.
So you can compare MB/s results for all systems for that crc32 code:

7z b -mm=crc32:8 -md18 -mmt1 -bt >> crc.txt
7z b -mm=crc32:8 -md18 -stm2 -bt >> crc.txt
7z b -mm=crc32:8 -md18 -bt >> crc.txt

1) Windows
2) Windows + WSL (linux)
3) Linux

[HITCHER]

CPU microcode C6 ( i disabled Windows from loading CC)

[HITCHER]

with CPU microcode CC loaded again

[HITCHER]

CPU microcode C6, old linux from USB-stick

[Igor Pavlov]

So linux is about 1-1.5% faster. It's small difference. So probably the frequency is almost same.
I have no idea why 7-Zip can't detect the frequency in Windows for that system.
You can try to disable Turbo frequency in BIOS, and test both linux and windows in base frequency.

[HITCHER]

I not want to test all OS-configurations, it is clear, that linux is somewhat faster, because of less background tasks. But I can test what happens in windows with standard frequency 2,7 GHz, if it is detected correct.

[HITCHER]

I disabled EIST and Turbo modes, see picture.
That's funny, also was not correct detected, but did run 2,7 GHz while benchmark.

[Igor Pavlov]

So we lose about 5-6% of frequency somewhere.
I'll look frequency for another Skylake processor benchmarks, if they will be posted.
I don't know why it can't show correct frequency.
The code for frequency check in Bench.cpp is pretty simple:

#define YY1 sum += val; sum ^= val;
#define YY3 YY1 YY1 YY1 YY1
#define YY5 YY3 YY3 YY3 YY3
#define YY7 YY5 YY5 YY5 YY5
static const UInt32 kNumFreqCommands = 128;
static UInt32 CountCpuFreq(UInt32 sum, UInt32 num, UInt32 val)
{
  for (UInt32 i = 0; i < num; i++)
  {
    YY7
  }
  return sum;
}

[HITCHER]

Fedora36 with same BIOS settings.

[HITCHER]

CPU microcode F0, newly updated Fedora 36 Linux

[HITCHER]

I did that again, but with CPU microcode C6 ( i disabled Windows from loading CC).
Results are not consistent . Sometimes it shows correct frequency.

[Igor Pavlov]

In linux, THP option can give different results:
THP:always or THP:madvise
THP:always is fastest option for lzma compression.
But THP is not so important for Swap4.

[HITCHER]

7z b -mm=hash -bt >> cpu.txt
7z b -mm=hash -stm1 -bt >> cpu.txt

multithread frequency was 3,1 GHz, singlethread 3,3 GHz (see picture)

edit:
then i closed hwinfo64 and browser again, and it showed in CLI while benchmark 3 GHz for singlethread.

[Igor Pavlov]

It's something new.
Just after 4-threads load, cpu shows 3.3 GHz for 5-7 seconds for 1T load , but then it degrades down to 3.05 GHz.
Can it be some unusual power plan in Windows?

[HITCHER]

I watched hwinfo64 during the benchmark run, it was all the time at 3,3 GHz for singlethread run, except for the first 2-3 seconds i fluctuated.
Windows power plan is "balanced".

[Igor Pavlov]

You can try aslo "High Performance" windows plan.
Maybe it can solve the problem?

[HITCHER]

No, with that power plan, it still looks so:

[Igor Pavlov]

in another your first screenshot above there was:
Activer Takt: x31 (3094) for core 0, and x33 for another cores.

[HITCHER]

there i did '-mmt=1', so was probably not running on core0.

[Serhii Zalohin]

MB: Asus ROG Strix X570-E Gaming, BIOS 4602

Memory: 2x Crucial Ballistix BL32G36C16U4BL.M16FB 32 GB DDR4-3600 @ 3541.8 MHz DDR4 SDRAM 16-18-18-38 1T (CL-RCD-RP-RAS CR)

Memory frequency is not 3600 MHz, because BUS freq is set to 98.4 MHz, not 100. Because current ASUS motherboard BIOS say "Why not 100? Because FU, what's why."

[HITCHER]

Sure there is a way in BIOS to select other BUS frequency, or APU frequency,
you have to change some settings from Auto to Manual first.
But if it runs stable in this configuration, leave it as it is, benchmark results look good.

[HITCHER]

PhenomII x6 1055T, overclocked to FSB 240 (x14 (3360), x16,5 (3960)), RAM DDR2-800 CL4, 4x2 GB DR, mainboard chipset-graphics (shared mem), Linux

edit:
Added benchmark with factory default settings for CPU (swap4-8.txt).

[HITCHER]

Intel Atom 330, RAM 2x1 GB DDR2-667 CL5, Linux from USB-Stick

[HITCHER]

R9-3900x, 4x16GB DDR4-3000 CL16 DR, Fedora 36
CPU is bit picky on which core it can reach 4,6 GHz, in this setting and kernel only 0 and 1.
When 7zz started without affinity, it may start on any random core.

[HITCHER]

Raspberry Pi5b 8GB, Specs: https://www.raspberrypi.com/products/raspberry-pi-5/
Standard clocks, RaspiOS 64 Bit.

Also included other 7-zip benchmarks, because maybe not known so far.