Re: Memory performance significantly improved w/ 2.4.1ac11

Re: Memory performance significantly improved w/ 2.4.1ac11

[Arvid Ericsson]

Mark Hahn skrev:
> first, are you sure your clock is write?  the changes appear
> to be tiny ~2 MB/s, and might be explained by the fact that
> 2.2 and 2.4 have different implementations of gettimeofday.

The change I was happy about was the one between 2.4.0ac10 and
2.4.1ac11. I think that it's a good thing that a few percent more memory
bandwidth appears from nowhere... Well, but that's just me I guess.

> (I'm assuming you're using gtod (second_wall.c) rather than
> a times-based measure.  the latter will be far less accurate.)

*Scratching my head* I have no idea of what your're talking about.
Sorry. Im using stream_whatever.cpp and whatever might be in it.

> 
> are you also aware that more modern compilers (2.95.2, I think)
> have more specific CPU tunings than just -mpentium?

I just used the suggested optimizations.
 
> 
> by "cycle length = 2", do you mean "tCAS latency = 2"?

Not quite sure, that the BIOS says "sdram cycle length=2" is all I know.

> finally (and don't take offense), those are astonishingly low
> Stream scores.  it's been a while since I ran Stream on a p5-class
> machine, but jeeze!  my dirt-cheap duron/600/kt133/pc133 sustains
> 600 MB/s!

Oh-well, I guess I just have a even dirt-cheaper machine than you. I
haven't got any idea of how machines similar to mine perform. 

/Arvid

Memory performance significantly improved w/ 2.4.1ac11

[Arvid Ericsson]

[-- Attachment #1: Type: text/plain, Size: 274 bytes --]

Hi y'all!
 I thought someone might be interested in some stream results. Please
ignore me otherwise. I know you don't like attachments, but the text
plainly refuses to be cut and pasted in a proper way (posting from
Windows 95).
 Btw, Im not subscribed to the list. 

/Arvid

[-- Attachment #2: new streams --]
[-- Type: text/plain, Size: 1627 bytes --]

STREAM

Tyan 1592S (VIA vt82c586b rev 0x41, 512Kb)
AMD K6-266 (66Mhzx4)
egcs-1.1.2 (g++ -O3 -march=pentium -funroll-loops)

2x32Mb (60ns, EDO SIMM):
	    Function      Rate (MB/s)   RMS time     Min time     Max time
	    Copy:          54.6566       0.2928       0.2927       0.2928
2.2.14	    Scale:         54.2420       0.2950       0.2950       0.2952
	    Add:           59.6669       0.4024       0.4022       0.4035
	    Triad:         59.0813       0.4063       0.4062       0.4070

	    Copy:          55.5657       0.2887       0.2879       0.2909
2.4.0ac10:  Scale:         54.9947       0.2916       0.2909       0.2934
	    Add:           60.5573       0.3974       0.3963       0.4002
	    Triad:         59.9713       0.4013       0.4002       0.4040

1x128Mb (PC133, cycle length=2):
	    Copy:          85.8094       0.1865       0.1864       0.1870
2.2.14	    Scale:         84.3468       0.1897       0.1897       0.1898
	    Add:           91.5097       0.2623       0.2623       0.2623
	    Triad:         90.2346       0.2660       0.2660       0.2662

	    Copy:          85.8653       0.1864       0.1863       0.1866
2.4.0ac10   Scale:         84.4007       0.1897       0.1896       0.1898
	    Add:           91.6042       0.2621       0.2620       0.2623
	    Triad:         90.3296       0.2659       0.2657       0.2675

	    Copy:          88.4453       0.1817       0.1809       0.1840
2.4.1ac11   Scale:         86.4535       0.1856       0.1851       0.1874
	    Add:           93.0553       0.2583       0.2579       0.2599
	    Triad:         91.7694       0.2620       0.2615       0.2648