Re: Benchmarking Ultras, compilers and scalability

From: Bill Broadley (bill@math.ucdavis.edu)
Date: Sat Nov 08 1997 - 00:41:14 CST


In comp.benchmarks Achim Gratz <gratz@ite.inf.tu-dresden.de> wrote:
: A factor of two is indeed a good rule of thumb, except if you're
: memory bandwith limited. The 300MHz processor is actually an
: UltraSPARC-II, which deals somewhat more intelligently with
: outstanding requests to the system, so the scaling is a bit better as
: the difference in the UPA clock (100MHz instead of 83Mhz) would
: suggest. The new Ultra30 machines have an improved memory subsystem
: that should boost performance in the memory limited case (STREAM
: numbers, anyone?). Of course if you were previously trashing the

I cheated a bit. Because I have "only" 128 MB's so far it's hard
to accurately time the stream benchmark.

Anyone have a line or two of asmembly to sample the Ultrasparc cycle
counter? I have similiar code for intel, alpha, and parisc.

So to make up for this I ran a stream with size= 3 million (68.7 MB)
and had it do each loop 10 times. This made the timing much more
accurate.

Function Rate (MB/s) RMS time Min time Max time
Copy: 289.1567 0.1676 0.1660 0.1690
Scale: 282.3528 0.1723 0.1700 0.1730
Add: 339.6234 0.2130 0.2120 0.2140
Triad: 346.1539 0.2084 0.2080 0.2090

Function Rate (MB/s) RMS time Min time Max time
Copy: 289.1567 0.1665 0.1660 0.1670
Scale: 279.0699 0.1723 0.1720 0.1730
Add: 339.6228 0.2125 0.2120 0.2130
Triad: 349.5147 0.2075 0.2060 0.2080

One of the better scores I've seen btw.

-- 
Bill Broadley           Bill@math.ucdavis.edu               UCD Math Sys-Admin
Linux is great.         http://math.ucdavis.edu/~bill                   PGP-ok



This archive was generated by hypermail 2b29 : Tue Apr 18 2000 - 05:23:07 CDT