Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 26 Sep 1996 20:03:01 +0200 (MET DST)
From:      Stefan Esser <se@zpr.uni-koeln.de>
To:        Kees.Koster@nym.sc.philips.com
Cc:        hardware@freebsd.org
Subject:   Re: lmbench results for AMD 5k86-P100
Message-ID:  <199609261803.UAA08010@x14.mi.uni-koeln.de>
In-Reply-To: <2A363AB4B4F@NLNMG01.nym.sc.philips.com>
References:  <2A363AB4B4F@NLNMG01.nym.sc.philips.com>

next in thread | previous in thread | raw e-mail | index | archive | help
Kees Jan Koster writes:
 > I did a little benchmarking on my machine, as Stefan requested (are 
 > you on this list?). I haven't had the time to do the bytebench yet.

Yes, I'm on the list ...

 > Maybe later.

I'd appreciate it ...

 > Machine: Exp8661 mainboard, 512kb PB cache, 32Mb EDO,
 > AMD 5k86-P100 @ 100MHz (overclocking doesn't work :).

Well, they'd be selling 5k86-120 chips, if it did :)

 > Could someone shed some light on the 'bad MHz' messages below?
 > It worries me a little, because I had some touble before. Unix worked 
 > like it always has: sweet as honey, and MS-DOS kept tripping over its 
 > feet: random hangs and crashes, missed sequences on the soundcard, 
 > etc.

According to an article in c't magazine (currently the 
best computer mag in Germany, INOMHO), this is due to the
fact that the 5k86 uses random replacement in its primary
cache. It is 4 way set-associative as that of other x86 
CPUs, but twice as large (16KB) as that of the P5. There
is a version of "ctcm" that has been modified for the 5k86,
since they (c't mag) could not find the cache access time,
else (same as lmbench :)

 > Both Unix and MS-DOS work, altough MS-DOS still feels kind'a fragile.

I don't care for DOS, but it's a little surprising to hear,
that Unix runs better (more reliable) than DOS on some 
hardware. But I could consider this fact as pro-AMD5K :)

Since I wanted to know how the 5k86 compares to my 5x86, I 
performed a lmbench run on my system, too. I'm using the 
later 1.1 version of lmbench, which is not yet available
as a port (but it is trivial, and I'll send a diff to 
ASAMI Satoshi).

Seems that the 5x86/133 (which is rated as equivalent to the
Pentium 75) is 60% of the 5k86-100 (which ought to be as fast
as a P100) in those tests, were 2.1.5 and -current are not 
too different. (Ie.: ignore the 5x86's PIPE performance ... :)

My system was not idle, but I think the results are useful
anyway :)

 > 		    L M B E N C H  1 . 0   S U M M A R Y
 > 		    ------------------------------------
 > 
 > 	    Processor, Processes - times in microseconds
 > 	    --------------------------------------------
 > Host                OS  Mhz    Null    Null  Simple /bin/sh Mmap 2-proc 8-proc
 > 			    Syscall Process Process Process  lat  ctxsw  ctxsw
 > --------- ------------- ---- ------- ------- ------- ------- ---- ------ ------
   amd5x86   FreeBSD 2.2-C  132      13      4K     19K     40K  148   15   26
 > LikeEver  FreeBSD 2.1.5  100       6    3.8K   14.2K     24K  123   35   44
 > LikeEver. FreeBSD 2.1.5  100       6    3.9K   14.1K     26K  123   36   46
 > LikeEver. FreeBSD 2.1.5  100       6    3.9K   14.0K     24K  124   36   46
 > pentium    Linux 1.1.54   91       3    3.3K   15.4K     49K   33   25   42

The 5k86 is FAR slower in the context-switch tests. This might 
be due to changes between -stable and -current, or because of 
wrong assumptions in the case of the 5k86. 
Is there a cache or TLB flush in the context switch code ?

 > 	    *Local* Communication latencies in microseconds
 > 	    -----------------------------------------------
 > Host                 OS  Pipe       UDP    RPC/     TCP    RPC/
 > 					    UDP             TCP
 > --------- ------------- ------- ------- ------- ------- -------
   amd5x86   FreeBSD 2.2-C      76     353     751     460     938
 > LikeEver  FreeBSD 2.1.5     101     204     421     251     568
 > LikeEver. FreeBSD 2.1.5     103     214     427     268     594
 > LikeEver. FreeBSD 2.1.5     103     208     430     260     579
 > pentium    Linux 1.1.54     157     658    1030    1164    1591

Except for the pipe code, the 5k86 is a factor of 5/3 as fast ...

 > 	    *Local* Communication bandwidths in megabytes/second
 > 	    ----------------------------------------------------
 > Host              OS Pipe  TCP  File   Mmap  Bcopy  Bcopy  Mem   Mem
 >     			          reread reread (libc) (hand) read write
 > -------- ------------- ---- ---- ------ ------ ------ ------ ---- 
   amd5x86  FreeBSD 2.2-C 30    7     14     25     16     15   34    41
 > LikeEver FreeBSD 2.1.5 14  9.6   23.1   58.6     29     28   81    42
 > LikeEver FreeBSD 2.1.5 14  9.6   23.1   57.7     29     28   81    42
 > LikeEver FreeBSD 2.1.5 14  9.7   23.1   58.4     29     28   81    42
 > pentium  Linux 1.1.54  13  2.4    9.8    4.7     18     18   48    32

The 5k86 is special in that its memory read performance is much
better (by a factor of 2) than its write performance. Not sure why, 
but this may also be caused by the motherboards characteristics.
(What chip-set is that, BTW ?)

 > 	    Memory latencies in nanoseconds
 > 	    (WARNING - may not be correct, check graphs)
 > 	    --------------------------------------------
 > Host                 OS   Mhz  L1 $   L2 $  Main mem    TLB    Guesses
 > --------- -------------   ---  ----   ----  --------    ---    -------
   amd5x86   FreeBSD 2.2-C   131     9    147       341
 > LikeEver  FreeBSD 2.1.5   100     -      -         -      -    Bad mhz?
 > LikeEver. FreeBSD 2.1.5   100     -      -         -      -    Bad mhz?
 > LikeEver. FreeBSD 2.1.5   100     -      -         -      -    Bad mhz?
 > pentium    Linux 1.1.54    90    11    294       439   1254

Hmm, as you can see, a 486 class CPU under FreeBSD offers 
better memory and cache latencies than a Pentium under Linux :) :)
(And even with twice the first level cache size ...)

LMBENCH for sure is not the best CPU benchmark, but it seems
the 5k86 is at least comparable to a P5-100, and seems to be
a cost effective alternative for a server that does not need
to offer peak performance values ...

Thanks for posting the interesting 5k86 results !

STefan



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199609261803.UAA08010>