[Date Index] [Thread Index] [Author Index]

Re: What is the fastest machine for Mathematica?

• To: mathgroup at smc.vnet.net
• Subject: [mg13762] Re: What is the fastest machine for Mathematica?
• From: Richard Easther <easther at het.brown.edu>
• Date: Mon, 24 Aug 1998 05:07:12 -0400
• Organization: High Energy Theory Group, Physics Dept, Brown University
• References: <6q3qkj$f19@smc.vnet.net> <6qp363$ajc@smc.vnet.net> <6r10id$fj5@smc.vnet.net>
• Sender: owner-wri-mathgroup at wolfram.com

Vladimir Radionov wrote:
> 
> Actually it's http://fampm201.tu-graz.ac.at/karl/timings30.html

I have been looking at these tests, and I have been wondering whether
anyone can shed any light on what the critical factors are that affect
Mathematica performance - ie memory, CPU type, cache, disk I/0,
floating point performance or integer....  Or do the tests listed on
this page not really give an accurate guide to performance?

In particular, I ran the benchmark on an UltraSparc with 4 300MHz
processors, 1 GB of RAM and 4 MB of cache per cpu (I think), and got a
benchmark score of less than 1.0, which is the sort of result seen with
a 166MhZ Pentium. 

There is a couple of reason why I might get sluggish performace from
this machine, as (for various reasons) the Mathematica files all live
on a remotely mounted disk, connected via a fairly slow network, so
disk I/0 is bad. Likewise, we are running the 3.0 kernel, and we are a
beta site for Solaris 2.7, both of which may (potentially) lead to
sub-optimal performance, I suppose. 

We are about to install the 3.0.2 release of Mathematica, and will put
it on a disk which is attached directly to the machine, and Solaris 2.7
is close to being finished, so we are in a position to isolate any
problems that are specific to our local setup. 

I do not have exact benchmarks on hand which compare the UltraSparc to
the Pentium, but for all other applications it "feels" a lot faster
than a Pentium 166....  Or has Wolfram simply put a lot more effort
into optimising the Mathematica code for Intel chips than it has for
the Sparcs?

Any comments?

Richard