Re: gfortran, g95, and dual-core

Janne Blomqvist wrote:
On 2007-09-30, Charles Russell <NOSPAM@xxxxxxxxxxxxx> wrote:
Richard Maine wrote:

So computing time is not simply a matter of counting "flops"
Nope. Hasn't been for decades.
For my present hardware, using a benchmark for "flops" that simply multiplies large dense matrices, I have:

2/2003 desktop 100 Mflop/2.4GHz = 0.04 flop/Hz g77 with -fbounds-check
4/2006 laptop 40 Mflop/1.7GHz = 0.02 flop/Hz g77 with -bounds-check

Timing was not very accurate, done by stopwatch for portability of the benchmark program, and these were short runs. The laptop, though newer, is less efficient, presumably due to the factors you mention. It was a low-end machine.

If the benchmark results above are representative of your real
application(s), you ought to be able to improve performance by more
than 1 order of magnitude without any new hardware. Dense matrix
multiplication of large matrices gets very close to the peak
performance of the processor, when using an optimized library, largely
due to better utilization of the cpu cache compared to a naive
implementation of the matrix multiplication algorithm. That is, your
2003 desktop should be able to do something close to 4 Gflops on
matmul.

In my real applications the bottleneck is generally the repeated evaluation of expensive functions. Since I keep changing the models, this must be done in a generic way, without customizing for a particular structure of the model. To minimize the number of function evaluations I routinely use the PORT version of NL2SOL for optimization. For Monte Carlo (actually, Metropolis algorithm) there is not much I can do other than to scale the problem as best as I can. Convergence to a static distribution is fastest with about 80% rejections, so I suppose I could get a fivefold gain by parallel evaluation before the rejection step.
.

Relevant Pages

  • Re: gfortran, g95, and dual-core
    ... For my present hardware, using a benchmark for "flops" that simply ... multiplies large dense matrices, I have: ... implementation of the matrix multiplication algorithm. ...
    (comp.lang.fortran)
  • Re: Could my slow computer be caused by a hardware fault?
    ... If I am in Internet Explorer and go to Bookmarks and hold my cursor over the down arrow it moves at about 1 file per second. ... I am no expert but I am wondering whether it could be a fault on the hard drive or other hardware problem? ... beginning of the disk, and 40MB/sec near the end of the ... To benchmark a CPU, I use a program like this as a quick check. ...
    (microsoft.public.windowsxp.hardware)
  • Re: IBM releases #1 TPC-E benchmark!
    ... Last time he brough up this subject I tried to say that this are all "hardware" ... IBM Hardware releases a benchmark on DB2. ...
    (comp.databases.informix)
  • Re: How to build an infiniband cluster
    ... >your hardware. ... >The natural question then is what happens if I use MPI +OpenIB? ... The only benchmark number we have for OpenIB on our ... while the only Voltaire ...
    (comp.arch)
  • Re: Oracle Benchmark Results for Different Hardware Configurations?
    ... You may not find the exact configurations. ... In order for a benchmark to be published, ... hardware somes from the capital budget - capex (capital ...
    (comp.databases.oracle.server)