12 Feb
2020
12 Feb
'20
12:57 a.m.
---------- Forwarded message ---------
From: Adam Thornton <athornton(a)gmail.com>
Date: Tue, Feb 11, 2020 at 5:56 PM
Subject: Re: [COFF] Old and Tradition was [TUHS] V9 shell
To: Clem Cole <clemc(a)ccc.com>
As someone working in this area right now....yeah, and that’s why
traditional HPC centers do not deliver the services we want for projects
like the Vera Rubin Observatory’s Legacy Survey Of Space and Time.
Almost all of our scientist-facing code is Python though a lot of the
performance critical stuff is implemented in C++ with Python bindings.
The incumbents are great at running data centers like they did in 1993.
That’s not the best fit for our current workload. It’s not generally the
compute that needs to be super-fast: it’s access to arbitrary slices
through really large data sets that is the interesting problem for us.
That’s not to say that that lfortran isn’t cool. It really really is, and
Ondrej Cestik has done amazing work in making modern FORTRAN run in a
Jupyter notebook, and the implications (LLVM becomes the Imagemagick of
compiled languages) are astounding.
But...HPC is no longer the cutting edge. We are seeing a Kuhnian paradigm
shift in action, and, sure, the old guys (and they are overwhelmingly guys)
who have tenure and get the big grants will never give up FORTRAN which
after all was good enough for their grandpappy and therefore good enough
for them. But they will retire. Scaling out is way way cheaper than
scaling up.
On Tue, Feb 11, 2020 at 11:41 AM Clem Cole <clemc(a)ccc.com> wrote:
moving to COFF
On Tue, Feb 11, 2020 at 5:00 AM Rob Pike <robpike(a)gmail.com> wrote:
My general mood about the current standard way of
nerd working is how
unimaginative and old-fashioned it feels.
...
But I'm a grumpy old man and getting far off topic. Warren should cry,
"enough!".
-rob
@Rob - I hear you and I'm sure there is a solid amount of wisdom in your
words. But I caution that just, because something is old-fashioned, does
not necessarily make it wrong (much less bad).
I ask you to take a look at the Archer statistics of code running in
production (Archer large HPC site in Europe):
http://archer.ac.uk/status/codes/
I think there are similar stats available for places like CERN, LRZ, and
of the US labs, but I know of these so I point to them.
Please note that Fortran is #1 (about 80%) followed by C @ about 10%,
C++ @ 8%, Python @ 1% and all the others at 1%.
Why is that? The math has not changed ... and open up any of those codes
and what do you see: solving systems of differential equations with linear
algebra. It's the same math my did by hand as a 'computer' in the
1950s.
There is not 'tensor flows' or ML searches running SPARK in there. Sorry,
Google/AWS et al. Nothing 'modern' and fresh -- just solid simple science
being done by scientists who don't care about the computer or sexy new
computer languages.
IIRC, you trained as a physicist, I think you understand their thinking. *They
care about getting their science done.*
By the way, a related thought comes from a good friend of mine from
college who used to be the Chief Metallurgist for the US Gov (NIST in
Colorado). He's back in the private sector now (because he could not
stomach current American politics), but he made an important
observation/comment to me a couple of years ago. They have 60+ years of
metallurgical data that has and his peeps have been using with known
Fortran codes. If we gave him new versions of those analytical programs
now in your favorite new HLL - pick one - your Go (which I love), C++
(which I loath), DPC++, Rust, Python - whatever, the scientists would have
to reconfirm previous results. They are not going to do that. It's not
economical. They 'know' how the data works, the types of errors they
have, how the programs behave* etc*.
So to me, the bottom line is just because it's old fashioned does not make
it bad. I don't want to write an OS in Fortran-2018, but I can wrote a
system that supports code compiled with my sexy new Fortran-2018 compiler.
That is to say, the challenge for >>me<< is to build him a new
supercomputer that can run those codes for him and not change what they are
doing and have them scale to 1M nodes *etc*..
_______________________________________________
COFF mailing list
COFF(a)minnie.tuhs.org
https://minnie.tuhs.org/cgi-bin/mailman/listinfo/coff
12 Feb
12 Feb
1:58 a.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
Fwiw. I'm sure your observations are solid and do make some good sense for
you.
But I ask you to consider the words you used defending your choices are
almostly exactly what was said by my CS profs at CMU and UCB on the 1970s
but it was Pascal not C++ and Snobol not Python as the scripting front end.
40 yrs later Fortran has matured but as Stu Feldman famously said in 1977
- he did not what it would look like but in the future (year 2000) we will
still call it Fortran. He was right and we still do.
Hey I'm a die hard systems person and as I said C is my primary tool to do
my own job (I don't write in Fortran) but I do understand what pays my
salary (and has my entire career). My experience has always been that
Scientists (of both sexs) don't care about the computer language nearly as
much as they do about getting their science done and trusting the data sets
that describe it.
Don't take my word for it. Look at the URL I gave you. Check out the same
data from the other big sites around the world - USA, UK, EU, JP, CN, RU,
and IN. I think you'll discover similar graphics. Fortran has been pretty
much constant at #1. What has changed over the years is the languages
used to prep the scientific data before running it throught the primary HPC
tool.
We can take this off line if you like but imo the reason has been/continue
s to be - is economics of the existing code base is extremely difficult to
displace. It's not an old guy / new guy thing. It's when you get money
to do your science it's not cost effective to replace something you trust
just because you don't like the computer language.
More over i know of only one case we're a production code originally
written in Fortran was successfully replaced - in grad school years ago, my
old housemate Tom Quarles wrote SPICE3 in C to replace the original SPICE2.
but note Tom had to do a huge amount of work to maintain the performance.
I also know of a couple of other commercial Mech E CAD and one Oil and Gas
code that tried to go to C++ but fell back to the original Fortran.
Anyway best wishes but be careful. Noel's email has much wisdom. New is
not necessarily better and old fashioned is not always a bad thing.
Clem
On Tue, Feb 11, 2020 at 7:57 PM Adam Thornton <athornton(a)gmail.com> wrote:
---------- Forwarded message ---------
From: Adam Thornton <athornton(a)gmail.com>
Date: Tue, Feb 11, 2020 at 5:56 PM
Subject: Re: [COFF] Old and Tradition was [TUHS] V9 shell
To: Clem Cole <clemc(a)ccc.com>
As someone working in this area right now....yeah, and that’s why
traditional HPC centers do not deliver the services we want for projects
like the Vera Rubin Observatory’s Legacy Survey Of Space and Time.
Almost all of our scientist-facing code is Python though a lot of the
performance critical stuff is implemented in C++ with Python bindings.
The incumbents are great at running data centers like they did in 1993.
That’s not the best fit for our current workload. It’s not generally the
compute that needs to be super-fast: it’s access to arbitrary slices
through really large data sets that is the interesting problem for us.
That’s not to say that that lfortran isn’t cool. It really really is, and
Ondrej Cestik has done amazing work in making modern FORTRAN run in a
Jupyter notebook, and the implications (LLVM becomes the Imagemagick of
compiled languages) are astounding.
But...HPC is no longer the cutting edge. We are seeing a Kuhnian paradigm
shift in action, and, sure, the old guys (and they are overwhelmingly guys)
who have tenure and get the big grants will never give up FORTRAN which
after all was good enough for their grandpappy and therefore good enough
for them. But they will retire. Scaling out is way way cheaper than
scaling up.
On Tue, Feb 11, 2020 at 11:41 AM Clem Cole <clemc(a)ccc.com> wrote:
--
Sent from a handheld expect more typos than usual
moving to COFF
On Tue, Feb 11, 2020 at 5:00 AM Rob Pike <robpike(a)gmail.com> wrote:
_______________________________________________
COFF mailing list
COFF(a)minnie.tuhs.org
https://minnie.tuhs.org/cgi-bin/mailman/listinfo/coff
My general mood about the current standard way of
nerd working is how
unimaginative and old-fashioned it feels.
...
But I'm a grumpy old man and getting far off topic. Warren should cry,
"enough!".
-rob
@Rob - I hear you and I'm sure there is a solid amount of wisdom in your
words. But I caution that just, because something is old-fashioned, does
not necessarily make it wrong (much less bad).
I ask you to take a look at the Archer statistics of code running in
production (Archer large HPC site in Europe):
http://archer.ac.uk/status/codes/
I think there are similar stats available for places like CERN, LRZ, and
of the US labs, but I know of these so I point to them.
Please note that Fortran is #1 (about 80%) followed by C @ about 10%,
C++ @ 8%, Python @ 1% and all the others at 1%.
Why is that? The math has not changed ... and open up any of those
codes and what do you see: solving systems of differential equations with
linear algebra. It's the same math my did by hand as a 'computer' in
the
1950s.
There is not 'tensor flows' or ML searches running SPARK in there.
Sorry, Google/AWS et al. Nothing 'modern' and fresh -- just solid simple
science being done by scientists who don't care about the computer or sexy
new computer languages.
IIRC, you trained as a physicist, I think you understand their thinking. *They
care about getting their science done.*
By the way, a related thought comes from a good friend of mine from
college who used to be the Chief Metallurgist for the US Gov (NIST in
Colorado). He's back in the private sector now (because he could not
stomach current American politics), but he made an important
observation/comment to me a couple of years ago. They have 60+ years of
metallurgical data that has and his peeps have been using with known
Fortran codes. If we gave him new versions of those analytical programs
now in your favorite new HLL - pick one - your Go (which I love), C++
(which I loath), DPC++, Rust, Python - whatever, the scientists would have
to reconfirm previous results. They are not going to do that. It's not
economical. They 'know' how the data works, the types of errors they
have, how the programs behave* etc*.
So to me, the bottom line is just because it's old fashioned does not
make it bad. I don't want to write an OS in Fortran-2018, but I can wrote
a system that supports code compiled with my sexy new Fortran-2018 compiler.
That is to say, the challenge for >>me<< is to build him a new
supercomputer that can run those codes for him and not change what they are
doing and have them scale to 1M nodes *etc*..
_______________________________________________
COFF mailing list
COFF(a)minnie.tuhs.org
https://minnie.tuhs.org/cgi-bin/mailman/listinfo/coff
3:01 a.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
I'm following this with some interest, though I don't have a horse in
this race, hell, I'm not even at the race track. But still, interesting.
What little Fortran background I have suggests that the difference
might be mind set. Fortran programmers are formally trained (at least I
was, there was a whole semester devoted to this) in accumulated errors.
You did a deep dive into how to code stuff so that the error was reduced
each time instead of increased. It has a lot to do with how floating
point works, it's not exact like integers are.
For the record, I *hate* floating point, for numerical work in C I have
always tried to have "domains", these 32 bits cover the domain from
0..2^32-1, these 32 bits cover the domain from 0..2^32-1 * 2^32, etc.
It's more work because you have to translate across domains but it
is exact. Floating point lets you cheat but the price of cheating is
the accumulated error.
Maybe I'm just going off of my 35 year old training, but it seems to me
that Fortran people pretty much live in floating point, and get what you
have to do to make that work. C / python / systems people "understand"
floating point but don't understand what you have to do to make that work.
They think they do, but they mostly don't, they just haven't hit the
accumulated error problem and wouldn't recognize it if they did.
What do you guys think?
On Tue, Feb 11, 2020 at 08:58:50PM -0500, Clem Cole wrote:
Fwiw. I'm sure your observations are solid and
do make some good sense for
you.
But I ask you to consider the words you used defending your choices are
almostly exactly what was said by my CS profs at CMU and UCB on the 1970s
but it was Pascal not C++ and Snobol not Python as the scripting front end.
40 yrs later Fortran has matured but as Stu Feldman famously said in 1977
- he did not what it would look like but in the future (year 2000) we will
still call it Fortran. He was right and we still do.
Hey I'm a die hard systems person and as I said C is my primary tool to do
my own job (I don't write in Fortran) but I do understand what pays my
salary (and has my entire career). My experience has always been that
Scientists (of both sexs) don't care about the computer language nearly as
much as they do about getting their science done and trusting the data sets
that describe it.
Don't take my word for it. Look at the URL I gave you. Check out the same
data from the other big sites around the world - USA, UK, EU, JP, CN, RU,
and IN. I think you'll discover similar graphics. Fortran has been pretty
much constant at #1. What has changed over the years is the languages
used to prep the scientific data before running it throught the primary HPC
tool.
We can take this off line if you like but imo the reason has been/continue
s to be - is economics of the existing code base is extremely difficult to
displace. It's not an old guy / new guy thing. It's when you get money
to do your science it's not cost effective to replace something you trust
just because you don't like the computer language.
More over i know of only one case we're a production code originally
written in Fortran was successfully replaced - in grad school years ago, my
old housemate Tom Quarles wrote SPICE3 in C to replace the original SPICE2.
but note Tom had to do a huge amount of work to maintain the performance.
I also know of a couple of other commercial Mech E CAD and one Oil and Gas
code that tried to go to C++ but fell back to the original Fortran.
Anyway best wishes but be careful. Noel's email has much wisdom. New is
not necessarily better and old fashioned is not always a bad thing.
Clem
On Tue, Feb 11, 2020 at 7:57 PM Adam Thornton <athornton(a)gmail.com> wrote:
---------- Forwarded message ---------
From: Adam Thornton <athornton(a)gmail.com>
Date: Tue, Feb 11, 2020 at 5:56 PM
Subject: Re: [COFF] Old and Tradition was [TUHS] V9 shell
To: Clem Cole <clemc(a)ccc.com>
As someone working in this area right now....yeah, and that???s why
traditional HPC centers do not deliver the services we want for projects
like the Vera Rubin Observatory???s Legacy Survey Of Space and Time.
Almost all of our scientist-facing code is Python though a lot of the
performance critical stuff is implemented in C++ with Python bindings.
The incumbents are great at running data centers like they did in 1993.
That???s not the best fit for our current workload. It???s not generally the
compute that needs to be super-fast: it???s access to arbitrary slices
through really large data sets that is the interesting problem for us.
That???s not to say that that lfortran isn???t cool. It really really is, and
Ondrej Cestik has done amazing work in making modern FORTRAN run in a
Jupyter notebook, and the implications (LLVM becomes the Imagemagick of
compiled languages) are astounding.
But...HPC is no longer the cutting edge. We are seeing a Kuhnian paradigm
shift in action, and, sure, the old guys (and they are overwhelmingly guys)
who have tenure and get the big grants will never give up FORTRAN which
after all was good enough for their grandpappy and therefore good enough
for them. But they will retire. Scaling out is way way cheaper than
scaling up.
On Tue, Feb 11, 2020 at 11:41 AM Clem Cole <clemc(a)ccc.com> wrote:
--
Sent from a handheld expect more typos than usual moving to COFF
On Tue, Feb 11, 2020 at 5:00 AM Rob Pike <robpike(a)gmail.com> wrote:
_______________________________________________
COFF mailing list
COFF(a)minnie.tuhs.org
https://minnie.tuhs.org/cgi-bin/mailman/listinfo/coff
My general mood about the current standard way of
nerd working is how
unimaginative and old-fashioned it feels.
...
But I'm a grumpy old man and getting far off topic. Warren should cry,
"enough!".
-rob
@Rob - I hear you and I'm sure there is a solid amount of wisdom in your
words. But I caution that just, because something is old-fashioned, does
not necessarily make it wrong (much less bad).
I ask you to take a look at the Archer statistics of code running in
production (Archer large HPC site in Europe):
http://archer.ac.uk/status/codes/
I think there are similar stats available for places like CERN, LRZ, and
of the US labs, but I know of these so I point to them.
Please note that Fortran is #1 (about 80%) followed by C @ about 10%,
C++ @ 8%, Python @ 1% and all the others at 1%.
Why is that? The math has not changed ... and open up any of those
codes and what do you see: solving systems of differential equations with
linear algebra. It's the same math my did by hand as a 'computer' in
the
1950s.
There is not 'tensor flows' or ML searches running SPARK in there.
Sorry, Google/AWS et al. Nothing 'modern' and fresh -- just solid simple
science being done by scientists who don't care about the computer or sexy
new computer languages.
IIRC, you trained as a physicist, I think you understand their thinking. *They
care about getting their science done.*
By the way, a related thought comes from a good friend of mine from
college who used to be the Chief Metallurgist for the US Gov (NIST in
Colorado). He's back in the private sector now (because he could not
stomach current American politics), but he made an important
observation/comment to me a couple of years ago. They have 60+ years of
metallurgical data that has and his peeps have been using with known
Fortran codes. If we gave him new versions of those analytical programs
now in your favorite new HLL - pick one - your Go (which I love), C++
(which I loath), DPC++, Rust, Python - whatever, the scientists would have
to reconfirm previous results. They are not going to do that. It's not
economical. They 'know' how the data works, the types of errors they
have, how the programs behave* etc*.
So to me, the bottom line is just because it's old fashioned does not
make it bad. I don't want to write an OS in Fortran-2018, but I can wrote
a system that supports code compiled with my sexy new Fortran-2018 compiler.
That is to say, the challenge for >>me<< is to build him a new
supercomputer that can run those codes for him and not change what they are
doing and have them scale to 1M nodes *etc*..
_______________________________________________
COFF mailing list
COFF(a)minnie.tuhs.org
https://minnie.tuhs.org/cgi-bin/mailman/listinfo/coff
_______________________________________________
COFF mailing list
COFF(a)minnie.tuhs.org
https://minnie.tuhs.org/cgi-bin/mailman/listinfo/coff
--
---
Larry McVoy lm at mcvoy.com http://www.mcvoy.com/lm
6:12 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On Tue, Feb 11, 2020 at 10:01 PM Larry McVoy <lm(a)mcvoy.com> wrote:
What little Fortran background I have suggests that
the difference
might be mind set. Fortran programmers are formally trained (at least I
was, there was a whole semester devoted to this) in accumulated errors.
You did a deep dive into how to code stuff so that the error was reduced
each time instead of increased. It has a lot to do with how floating
point works, it's not exact like integers are.
Just a thought, but it might also be the training. My Dad (a
mathematician and 'computer') passed a few years ago, I'd love to have
asked him. But I suspect when he and his peeps were doing this with a
slide rule or at best an Friden mechanical adding machine, they were
acutely aware of how errors accumulated or not. When they started to
convert their processes/techniques to Fortran in the early 1960s, I agree
with you that I think they were conscious of what they were doing. I'm
not sure modern CS types are taught the same things as what might be taught
in a course being run by a pure scientist who cares in the same way folks
like our mothers and fathers did in the 1950s and 60s.
9:55 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On 12/02/2020, Clem Cole <clemc(a)ccc.com> wrote (in part):
But I suspect when he and his peeps were doing this
with a
slide rule or at best an Friden mechanical adding machine, they were
acutely aware of how errors accumulated or not.
This being COFF, I feel justified in chiming in with a story from my
grad student days.
A fellow student was working on something -- I have forgotten the
actual problem but studying some behaviour near a singularity -- and
he wanted some numerical values. So he obtained Fortran code from a
grad student in physics. The result blew up in his region of
interest. The person who gave him the code had no interest in
investigating because he was happy with the results in his region.
(After months of toil, he found serious approximation errors.)
N.
10:11 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On Wed, Feb 12, 2020, 11:13 AM Clem Cole <clemc(a)ccc.com> wrote:
On Tue, Feb 11, 2020 at 10:01 PM Larry McVoy <lm(a)mcvoy.com> wrote:
Most cs types barely know that 2.234 might not be an exact number when
converted to binary... A few, however can do sophisticated analysis on the
average ULP for complex functions over the expected range..
Warner
_______________________________________________
What little Fortran background I have suggests
that the difference
might be mind set. Fortran programmers are formally trained (at least I
was, there was a whole semester devoted to this) in accumulated errors.
You did a deep dive into how to code stuff so that the error was reduced
each time instead of increased. It has a lot to do with how floating
point works, it's not exact like integers are.
Just a thought, but it might also be the training. My Dad (a
mathematician and 'computer') passed a few years ago, I'd love to have
asked him. But I suspect when he and his peeps were doing this with a
slide rule or at best an Friden mechanical adding machine, they were
acutely aware of how errors accumulated or not. When they started to
convert their processes/techniques to Fortran in the early 1960s, I agree
with you that I think they were conscious of what they were doing. I'm
not sure modern CS types are taught the same things as what might be taught
in a course being run by a pure scientist who cares in the same way folks
like our mothers and fathers did in the 1950s and 60s.
COFF mailing list
COFF(a)minnie.tuhs.org
https://minnie.tuhs.org/cgi-bin/mailman/listinfo/coff
10:45 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On 2/12/2020 4:11 PM, Warner Losh wrote:
On Wed, Feb 12, 2020, 11:13 AM Clem Cole <clemc(a)ccc.com
<mailto:clemc@ccc.com>> wrote:
On Tue, Feb 11, 2020 at 10:01 PM Larry McVoy <lm(a)mcvoy.com
<mailto:lm@mcvoy.com>> wrote:
What little Fortran background I have suggests that the difference
might be mind set. Fortran programmers are formally trained (at
least I
was, there was a whole semester devoted to this) in accumulated
errors.
You did a deep dive into how to code stuff so that the error was
reduced
each time instead of increased. It has a lot to do with how
floating
point works, it's not exact like integers are.
Just a thought, but it might also be the training. My Dad (a
mathematician and 'computer') passed a few years ago, I'd love to
have asked him. But I suspect when he and his peeps were doing
this with a slide rule or at best an Friden mechanical adding
machine, they were acutely aware of how errors accumulated or not.
When they started to convert their processes/techniques to Fortran
in the early 1960s, I agree with you that I think they were
conscious of what they were doing. I'm not sure modern CS types
are taught the same things as what might be taught in a course being
run by a pure scientist who cares in the same way folks like our
mothers and fathers did in the 1950s and 60s.
Most cs types barely know that 2.234 might not be an exact number when
converted to binary... A few, however can do sophisticated analysis on
the average ULP for complex functions over the expected range..
If that is true of some today, that is sad and disappointing. I think I
was taught otherwise in my beginning C.S. course at UT-Austin in 1971.
If I recall correctly:
- all doctoral candidates ended up taking two semesters of numerical
analysis. I still have two volume n.a. text in the attic (orange, but
not "burnt orange", IIRC).
- numerical analysis was covered on the doctoral qualifying exam.
--
voice: +1.512.784.7526 e-mail: sauer(a)technologists.com
fax: +1.512.346.5240 Web: https://technologists.com/sauer/
Facebook/Google/Skype/Twitter: CharlesHSauer
11:05 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On Wed, Feb 12, 2020 at 04:45:54PM -0600, Charles H Sauer wrote:
If I recall correctly:
- all doctoral candidates ended up taking two semesters of numerical
analysis. I still have two volume n.a. text in the attic (orange, but not
"burnt orange", IIRC).
- numerical analysis was covered on the doctoral qualifying exam.
Pretty sure Madison required that stuff for Masters degrees. Or maybe
undergrad, I feel like I took that stuff pretty early on.
I'm very systems oriented so I can't imagine I would have taking that
willingly. I hate the whole idea of floating point, just seems so
error prone.
11:54 p.m.
New subject: floating point (Re: Old and Tradition was [TUHS] V9 shell
On Feb 12, 2020, at 3:05 PM, Larry McVoy <lm(a)mcvoy.com> wrote:
On Wed, Feb 12, 2020 at 04:45:54PM -0600, Charles H Sauer wrote:
David Goldberg's article "What every computer scientist should know
about floating-point arithmetic" is a good one to read:
https://www.itu.dk/~sestoft/bachelor/IEEE754_article.pdf
I still have Bill Press's Numerical Recipes book though not opened
recently (as in not since '80s)!
It is interesting that older languages such as Lisp & APL have a
builtin concept of tolerance. Here 0.3 < 0.1 + 0.2 is false. But
in most modern languages it is true! This is so since 0.1 + 0.2 is
0.30000000000000004. In Fortran you'd write something like
abs(0.3 - (0.1 + 0.2)) > tolerance. You can do the same in C
etc.but for some reason it seems to be uncommon :-)
If I recall correctly:
- all doctoral candidates ended up taking two semesters of numerical
analysis. I still have two volume n.a. text in the attic (orange, but not
"burnt orange", IIRC).
- numerical analysis was covered on the doctoral qualifying exam.
Pretty sure Madison required that stuff for Masters degrees. Or maybe
undergrad, I feel like I took that stuff pretty early on.
I'm very systems oriented so I can't imagine I would have taking that
willingly. I hate the whole idea of floating point, just seems so
error prone.
11:56 p.m.
New subject: floating point (Re: Old and Tradition was [TUHS] V9 shell
On Feb 12, 2020, at 3:54 PM, Bakul Shah
<bakul(a)bitblocks.com> wrote:
On Feb 12, 2020, at 3:05 PM, Larry McVoy <lm(a)mcvoy.com> wrote:
Er... not right. Fl. pt. arithmetic is hard!
On Wed, Feb 12, 2020 at 04:45:54PM -0600, Charles H Sauer wrote:
David Goldberg's article "What every computer scientist should know
about floating-point arithmetic" is a good one to read:
https://www.itu.dk/~sestoft/bachelor/IEEE754_article.pdf
I still have Bill Press's Numerical Recipes book though not opened
recently (as in not since '80s)!
It is interesting that older languages such as Lisp & APL have a
builtin concept of tolerance. Here 0.3 < 0.1 + 0.2 is false. But
in most modern languages it is true! This is so since 0.1 + 0.2 is
0.30000000000000004. In Fortran you'd write something like
abs(0.3 - (0.1 + 0.2)) > tolerance. You can do the same in C
etc.but for some reason it seems to be uncommon :-) If I recall correctly:
- all doctoral candidates ended up taking two semesters of numerical
analysis. I still have two volume n.a. text in the attic (orange, but not
"burnt orange", IIRC).
- numerical analysis was covered on the doctoral qualifying exam.
Pretty sure Madison required that stuff for Masters degrees. Or maybe
undergrad, I feel like I took that stuff pretty early on.
I'm very systems oriented so I can't imagine I would have taking that
willingly. I hate the whole idea of floating point, just seems so
error prone. 
13 Feb
13 Feb
1:21 a.m.
New subject: floating point (Re: Old and Tradition was [TUHS] V9 shell
On 2020-02-12 6:54 PM, Bakul Shah wrote:
On Feb 12, 2020, at 3:05 PM, Larry McVoy
<lm(a)mcvoy.com> wrote:
Mathematica also does significance tracking. It is probably the single
most egregious omission from IEEE FP and unfortunately not corrected in
the new standard.
Check Steve Richfield's old posts on Usenet for more.
--Toby
On Wed, Feb 12, 2020 at 04:45:54PM -0600, Charles H Sauer wrote:
David Goldberg's article "What every computer scientist should know
about floating-point arithmetic" is a good one to read:
https://www.itu.dk/~sestoft/bachelor/IEEE754_article.pdf
I still have Bill Press's Numerical Recipes book though not opened
recently (as in not since '80s)!
It is interesting that older languages such as Lisp & APL have a
builtin concept of tolerance. Here 0.3 < 0.1 + 0.2 is false. But If I recall correctly:
- all doctoral candidates ended up taking two semesters of numerical
analysis. I still have two volume n.a. text in the attic (orange, but not
"burnt orange", IIRC).
- numerical analysis was covered on the doctoral qualifying exam.
Pretty sure Madison required that stuff for Masters degrees. Or maybe
undergrad, I feel like I took that stuff pretty early on.
I'm very systems oriented so I can't imagine I would have taking that
willingly. I hate the whole idea of floating point, just seems so
error prone. in most modern languages it is true! This is so since
0.1 + 0.2 is
0.30000000000000004. In Fortran you'd write something like
abs(0.3 - (0.1 + 0.2)) > tolerance. You can do the same in C
etc.but for some reason it seems to be uncommon :-)
_______________________________________________
COFF mailing list
COFF(a)minnie.tuhs.org
https://minnie.tuhs.org/cgi-bin/mailman/listinfo/coff
6:57 a.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On 2020-Feb-12 15:11:27 -0700, Warner Losh <imp(a)bsdimp.com> wrote:
Most cs types barely know that 2.234 might not be an
exact number when
converted to binary...
My favourite example (based on a real bug that I was involved in
investigating) is: int(0.29 * 100) = 28 (using IEEE doubles). That result
is very counter-intuitive. (Detailed explanation available on request).
In some ways IEEE-754 has made things worse - before that standard, almost
every computer system implemented something different. Floating point was
well known to be "here be dragons" territory and people who ventured there
were either foolhardy or had enough numerical analysis training to be able
to understand (and cope with) what the computer was doing.
IEEE-754 offered a seeming nirvana - where the same inputs would result in
the same outputs (down to the bit) irrespective of what computer you ran
your code on. The underlying problems were still present but, at least in
simple cases, were masked. Why employ an experienced numerical analyst to
develop the code when a high-school student can translate the mathematical
equations into running code in the language-du-jour? (And some languages,
like Java, make a virtue out of hiding all the information needed to
properly handle exceptional conditions).
--
Peter Jeremy
16 Feb
16 Feb
9:47 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
I think it's implicit with fooling around with slide rules. Everything
is logarithmic, therefore imprecise beyond a certain level (floating
point number or iteration, it's the same problem). You learn to
approximate, within a certain level of confidence (or diffidence :).
(FWVLIW - I bought a Dover book on slide rule in the late 70s while at
high school, and shortly after, a real slide rule, and it's stuck with
me.)
Wesley Parish
On 2/13/20, Clem Cole <clemc(a)ccc.com> wrote:
On Tue, Feb 11, 2020 at 10:01 PM Larry McVoy
<lm(a)mcvoy.com> wrote:
What little Fortran background I have suggests
that the difference
might be mind set. Fortran programmers are formally trained (at least I
was, there was a whole semester devoted to this) in accumulated errors.
You did a deep dive into how to code stuff so that the error was reduced
each time instead of increased. It has a lot to do with how floating
point works, it's not exact like integers are.
Just a thought, but it might also be the training. My Dad (a
mathematician and 'computer') passed a few years ago, I'd love to have
asked him. But I suspect when he and his peeps were doing this with a
slide rule or at best an Friden mechanical adding machine, they were
acutely aware of how errors accumulated or not. When they started to
convert their processes/techniques to Fortran in the early 1960s, I agree
with you that I think they were conscious of what they were doing. I'm
not sure modern CS types are taught the same things as what might be taught
in a course being run by a pure scientist who cares in the same way folks
like our mothers and fathers did in the 1950s and 60s.
10:10 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On Sun, Feb 16, 2020 at 4:47 PM Wesley Parish <wobblygong(a)gmail.com> wrote:
FWVLIW - I bought a Dover book on slide rule in the
late 70s while at
high school, and shortly after, a real slide rule, and it's stuck with
me.
My dad taught me with a plastic slide rule he put in our stocking in the
early/mid 1960s. This was how I (and many others) learn about
interpolation. I also learned to make log/log paper with it. A few years
later, my grandfather died when I was in engineering school. My grandmother
sent me his slide rule to remember him by (which I still have). Although
she did not know that at the time, I already owned the then hot item, a TI
SR50 scientific calculator - which I paid the $150 in 1972 dollars (about
$900 in today's money). I also got his drafting table, but I no longer
have that. The slide rule is made of ivory on top of metal (I think
bronze but I never had it checked). It was probably made in the 1920s. It
stays in a box in desk ;-)
A slightly, sad part is I don't think either of my kids knows how to use
it, and while both have degrees in science, I don't think either wants it.
10:45 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On 2/16/2020 5:10 PM, Clem Cole wrote:
This was how I (and many others) learn about
interpolation
I learned about interpolation in either middle school or high school
math. And used it many times in video games I had my hands in. I was a
year or two ahead (depending on which year) in math.
Create an array (or a cache at startup) with all the logs I needed, and
then interpolate on the fly. And do the reverse for the alog().
Quite handy.
As for slide-rules, I remember my father had one, and used it regularly.
I never even looked at it. He died when I was 16. A few years later, I
took it out of it's case, looked at it, and thought "ah! logs". I had
never known what a slide-rule was. And I was born in 1965 ;) They
certainly didn't teach us about slide rules in school.
ak
PS: Take the log of a number, divide by two, take the alog() and get the
sqrt(). Divide by 3, cube root, etc. I was impressed, to say the least.
11:50 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On Sun, 16 Feb 2020 17:10:03 -0500 Clem Cole <clemc(a)ccc.com> wrote:
On Sun, Feb 16, 2020 at 4:47 PM Wesley Parish <wobblygong(a)gmail.com> wrote:
What brand was your grandfather's sliderule?
I had a yellow Pickett sliderule in my undergrad days. IIRC it
had some sort of aluminum alloy slide and didn't work as
smoothly as an Aristo or a Faber-Castell. A friend had TI
scinetific calculator with RED LED digits -- may have been the
SR50. Aesthetically it didn't hold a candle to the beautfiul
sliderules!
Some sliderule simulators here: https://www.sliderules.org/
(mine looked exactly like the N600)
Online Museum here: https://sliderulemuseum.com/SRM_Home.htm
FWVLIW - I bought a Dover book on slide rule in
the late 70s while at
high school, and shortly after, a real slide rule, and it's stuck with
me.
My dad taught me with a plastic slide rule he put in our stocking in the
early/mid 1960s. This was how I (and many others) learn about
interpolation. I also learned to make log/log paper with it. A few years
later, my grandfather died when I was in engineering school. My grandmother
sent me his slide rule to remember him by (which I still have). Although
she did not know that at the time, I already owned the then hot item, a TI
SR50 scientific calculator - which I paid the $150 in 1972 dollars (about
$900 in today's money). I also got his drafting table, but I no longer
have that. The slide rule is made of ivory on top of metal (I think
bronze but I never had it checked). It was probably made in the 1920s. It
stays in a box in desk ;-) A slightly, sad part is I don't think either of
my kids knows how to use
it, and while both have degrees in science, I don't think either wants it.
So it goes!
18 Feb
18 Feb
12:17 a.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On Sun, 16 Feb 2020, Bakul Shah wrote:
What brand was your grandfather's sliderule?
Mine was an Arista, and had it right through school; have no idea where it
is now. I also had a circular slide rule, which was fascinating.
Trivia: slide rules were banned for exams (and no calculators in those
days), but log tables were OK; come some important exam, some idiot of a
teacher forgot to specify that log tables were allowed so they were
forbidden. I merely worked the problems through right down to the long
division, and left them there with a note saying that we were supposed to
use log tables; I passed...
-- Dave
12:48 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
My uncle had a cylindrical sliderule. It had about a dozen "rules" that
rotated around the central cylinder, with different gratings on the
opposite side of each rule. And the central cylinder had at least a dozen,
probably 20 or more, gratings. That yielded hundreds of combinations. I
wish he had bequeathed it to me. I have no idea what became of it.
On Mon, Feb 17, 2020 at 7:18 PM Dave Horsfall <dave(a)horsfall.org> wrote:
On Sun, 16 Feb 2020, Bakul Shah wrote:
What brand was your grandfather's sliderule?
Mine was an Arista, and had it right through school; have no idea where it
is now. I also had a circular slide rule, which was fascinating.
Trivia: slide rules were banned for exams (and no calculators in those
days), but log tables were OK; come some important exam, some idiot of a
teacher forgot to specify that log tables were allowed so they were
forbidden. I merely worked the problems through right down to the long
division, and left them there with a note saying that we were supposed to
use log tables; I passed...
-- Dave
_______________________________________________
COFF mailing list
COFF(a)minnie.tuhs.org
https://minnie.tuhs.org/cgi-bin/mailman/listinfo/coff
24 Feb
24 Feb
9:40 a.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
Larry McVoy <lm(a)mcvoy.com> wrote:
Fortran programmers are formally trained (at least I
was, there was a whole semester devoted to this) in accumulated errors.
You did a deep dive into how to code stuff so that the error was reduced
each time instead of increased. It has a lot to do with how floating
point works, it's not exact like integers are.
I was unaware that there's formal training to be had around this but
it's something I'd like to learn more about. Any recommendations on
materials? I don't mind diving into Fortran itself either.
Background: I've been playing around with actuarial calculations
(pensions, life insurance, etc) which involve lots of accumulation over
time and I'm finding that, unsurprisingly, different implementations of
the same rules yield fairly different outcomes due to accumulated
errors.
Sijmen
3:19 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On Mon, Feb 24, 2020 at 10:40:10AM +0100, Sijmen J. Mulder wrote:
Larry McVoy <lm(a)mcvoy.com> wrote:
My training was 35 years ago, I have no idea where to go look for this
stuff now. I googled and didn't find much. I'd go to the local
University that teaches Fortran and ask around.
Fortran programmers are formally trained (at
least I
was, there was a whole semester devoted to this) in accumulated errors.
You did a deep dive into how to code stuff so that the error was reduced
each time instead of increased. It has a lot to do with how floating
point works, it's not exact like integers are.
I was unaware that there's formal training to be had around this but
it's something I'd like to learn more about. Any recommendations on
materials? I don't mind diving into Fortran itself either.
4:27 p.m.
New subject: Fwd: Old and Tradition was [TUHS] V9 shell
On Mon, Feb 24, 2020 at 10:19 AM Larry McVoy <lm(a)mcvoy.com> wrote:
On Mon, Feb 24, 2020 at 10:40:10AM +0100, Sijmen J.
Mulder wrote:
My training was 35 years ago, I have no idea where to go look for this
stuff now. I googled and didn't find much. I'd go to the local
University that teaches Fortran and ask around.
1. Download the Intel Fortran compiler for your Mac, Linux or Windows
box. This will also give you a number of tuning tools. This compiler can
understand syntax back to FORTRAN-66 through Fortran2018 (and often
without any switches - it can usually figure it out and do the right thing).
2. Get a copy of
https://www.amazon.com/Explained-Numerical-Mathematics-Scientific-Computati…
3. Go to the physics and chem depts a place start as Larry said
4. Or head to a supercomputer center in the US or EU
Larry McVoy <lm(a)mcvoy.com> wrote:
> Fortran programmers are formally trained (at least I
> was, there was a whole semester devoted to this) in accumulated errors.
> You did a deep dive into how to code stuff so that the error was
reduced
each time
instead of increased. It has a lot to do with how floating
point works, it's not exact like integers are.
I was unaware that there's formal training to be had around this but
it's something I'd like to learn more about. Any recommendations on
materials? I don't mind diving into Fortran itself either.
1516
days inactive
1528
days old
20 comments
14 participants
participants (14)
-
Adam Thornton -
Arthur Krewat -
Bakul Shah -
Charles H Sauer -
Clem Cole -
Dave Horsfall -
John P. Linderman -
Larry McVoy -
Nemo -
Peter Jeremy -
Sijmen J. Mulder -
Toby Thain -
Warner Losh -
Wesley Parish