Performance: Opteron 150 v.s. Athlon 64 x2 4400+ for LAMP server?

On Sun, 12 Aug 2007 16:16:30 +0000, Wes Newell wrote:

On Sun, 12 Aug 2007 11:53:39 +0000, General Schvantzkoph wrote:

On Sun, 12 Aug 2007 05:08:30 +0000, Wes Newell wrote:

On Sun, 12 Aug 2007 01:11:46 +0000, General Schvantzkoph wrote:

AMD Processors don't have much margin, if you want to overclock you
should use Intel. The Core2 processors have a lot of headroom. The
Core2 processors are also 30% faster then AMD processors on a clock
for clock basis. My recommendation would be a 6750 which is priced at
$212 on Newegg. I'm running my 6800 at 3GHz with stock cooling. It's
been running 24/7 since last November without a problem.

AMD holds the top 11 spots in the price/performance index, and 14 out
of the top 16. the top spot is held by AMD X2 3800+, and there isn't a
core 2 duo chip in the bunch. All AMD X2 cpu's top all the core 2 duo
chips except one. The E4300 beats only the X2 6000+ in this chart. As
for overclocking, the AMD X2 3800+ will overclock close to 50% if not
do it. And running your 6800 at 3GHz isn't much of an overclock since
it's default speed is 2.93 GHz to start if I'm not mistaken and sells
for about $900. And I don't know where you came up with that 30%
faster per clock, but that's just pure BS.

I've benchmarked GCC compiles, Xilinx place and routes, and NCVerilog
simulations. The Core2 is minimally 30% faster then the Athlon 64 X2 on
a clock for clock basis, on NCVerilog it's twice as fast. These are
real applications not bull**** synthetic benchmarks or graphics heavy
games.

I went through about 10 of the benchmarks on tomsharware and none of
them came even close to 30% faster. You can go through the charts here.
C2D and X2 both at 3000MHz.

http://www23.tomshardware.com/cpu_2007.html?
modelx=33&model1=921&model2=872&chart=435

I just reran some tests. I did a GCC make with -j 1 and -j 2, and
NCVerilog with recordvars on and recordvars off (recordvars saves
simulation state so it does a lot of IO). The two processors are an X2
4400+ (2.2GHz, 1M caches) and an E6700 overclocked to 3Ghz (4M shared
cache). I've normalized the clocks so the ratios are 1 for 1. The
smallest difference is on a single threaded GCC, 12%, the largest is
NCVerilog with recordvars off, 61%.


X2 Core2
Cache 2x1M 4M
Clock 2.2 3

Times in seconds,
Make J 2 24.44 15.32
Make J 1 44.71 29.24
NCVer, trn 190.88 96.21
NCVer, no 31.93 14.51

Clock Normal 1.36 1

Normalized times in seconds,
Make J 2 17.92 15.32
Make J 1 32.78 29.24
NCVer, trn 139.97 96.21
NCVer, no 23.41 14.51

Ratios vs Core2
Make J 2 0.85 1
Make J 1 0.89 1
NCVer, trn 0.68 1
NCVer, no 0.61 1

Ratios vs X2
Make J 2 1 1.169
Make J 1 1 1.121
NCVer, trn 1 1.454
NCVer, no 1 1.613

On Aug 13, 1:04 am, General Schvantzkoph
wrote:

I just reran some tests.

Interesting!

Core2 seems faster; how much of that can be attributed to the larger
L2 cache would you think?

Another interesting thing is that - in this particular test - they
both utilize the two cores quite efficiently; at the same level one
would expect from a more traditional SMP system. Most applications
appears not take advantage of this, though - which makes the whole
multiple core proposition much weaker. At the moment, it seems as if
the only way to make full use of several cores is to run a
virtualization platform, allowing several instances of the OS to run
in parallel, presumably one on each core. It does seems like a rather
hefty overhead to make use of the two cores?

On Mon, 13 Aug 2007 03:34:47 -0700, Peter wrote:

On Aug 13, 1:04 am, General Schvantzkoph wrote:

I just reran some tests.

Interesting!

Core2 seems faster; how much of that can be attributed to the larger L2
cache would you think?

Another interesting thing is that - in this particular test - they both
utilize the two cores quite efficiently; at the same level one would
expect from a more traditional SMP system. Most applications appears not
take advantage of this, though - which makes the whole multiple core
proposition much weaker. At the moment, it seems as if the only way to
make full use of several cores is to run a virtualization platform,
allowing several instances of the OS to run in parallel, presumably one
on each core. It does seems like a rather hefty overhead to make use of
the two cores?

LAMP servers are running a number of applications simultaneously, MySQL,
Apache and PHP, so there is plenty of opportunity to utilize both cores.
I'm pretty sure that MySQL is multithreaded as is Apache so a LAMP server
ought to be able to take advantage of a four core chip as well.

On my benchmarks the big cache is probably the largest contributor to the
performance difference, it certainly is for NCVerilog which is very cache
sensitive. The Core2 has some other important advantages, it cracks 4
instructions at a time instead of three, and it has better branch
prediction. Small improvements in branch prediction have a very large
impact on performance.