AIM Independent Resource Benchmark results for kernel-2.5.44
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
From: "Jakob Oestergaard" <[email protected]>
...[snip]...
Quoted 4 lines
The correct way to terminate that loop is, like was already suggested, doing a comparison to see if the residual is "numerically zero" or "sufficiently zero-ish for the given purpose". Eg. "delta < 1E-12" or eventually "fabs(delta) < 1E-12".
Tim Witham at the OSDL told me that he ran some experiments with different convergent deltas:
zero Rate (ops/sec) Iteration Rate
10-6 331,300 1656.5 10-8 315,049 1575.0 10-10 302,000 1510.0 10-12 292,300 1461.5 10-14 285,400 1427.0
Anything smaller than 10-14 didn't converge.
--
John Hawkes
But then I fixed the code with the compile option. But
I agree I don't like floating point code that looks for equality with zero to end a loop. I would much rather have the code to for a fixed <= number as that prevents the issue with "how close to zero" on your system.
History lesson. In the very old days this benchmark
caused many a math library to be changed because different architectures had different ideas as to how close to zero you had to be before a floating point number was considered zero and some people ran a lot more iterations than other folks for the same test. The response was "Well that is an implementation detail and it does matter to folks as they write code to test for convergence." I guess they went to a cheaper school than I did. :-)
Of course that was when it took a 20 processor system
to get over 100 users. :-)
But if we are looking at this test overall maybe we
should be considering how to update it for systems that have changed so much over the years. For example the fakeh.tar file is only 192 KB. Heck that is the size of some single documents these days.
Tim
Quoted 22 lines · John Hawkes
From: "Jakob Oestergaard" <[email protected]>
...[snip]...Quoted 4 lines
The correct way to terminate that loop is, like was already suggested, doing a comparison to see if the residual is "numerically zero" or "sufficiently zero-ish for the given purpose". Eg. "delta < 1E-12" or eventually "fabs(delta) < 1E-12".
Tim Witham at the OSDL told me that he ran some experiments with different convergent deltas:
zero Rate (ops/sec) Iteration Rate
10-6 331,300 1656.5 10-8 315,049 1575.0 10-10 302,000 1510.0 10-12 292,300 1461.5 10-14 285,400 1427.0Anything smaller than 10-14 didn't converge.
--
John Hawkes
Signature · 5 lines
Timothy D. Witham - Lab Director - [email protected] Open Source Development Lab Inc - A non-profit corporation 15275 SW Koll Parkway - Suite H - Beaverton OR, 97006 (503)-626-2455 x11 (office) (503)-702-2871 (cell) (503)-626-2436 (fax)