Computing elements of a 1000 x 60 matrix exhausts RAMWhat is a Mathematica packed array?Paging RAM in case of memory shortage issueHeavy duty operations, RAM, and ReadyboostEfficient calculation of diagonal matrix elementsHow to know each variable used how much RAMReplacing elements of a matrixHow to clear RAM memory in a running code?How does this code behave with more than 32GiB RAM?Replace diagonal elements in sparse matrixSolution list from “Solve” too large for my RAM spaceLimit the amount of RAM Mathematica may access?
Their answer is discrete, mine is continuous. They baited me into the wrong answer. I have a P Exam question
C SIGINT signal in Linux
How do I write "Show, Don't Tell" as an Asperger?
Can't login after removing Flatpak
How is TD(0) method helpful? What good does it do?
Are there cubesats in GEO?
Building a road to escape Earth's gravity by making a pyramid on Antartica
In this example, which path would a monster affected by the Dissonant Whispers spell take?
Can you `= delete` a templated function on a second declaration?
Did Darth Vader wear the same suit for 20+ years?
Why don’t airliners have temporary liveries?
How to make thick Asian sauces?
Does an ice chest packed full of frozen food need ice? 18 day Grand Canyon trip
Incremental Ranges!
Secure offsite backup, even in the case of hacker root access
How can drunken, homicidal elves successfully conduct a wild hunt?
What are the words for people who cause trouble believing they know better?
How could a government be implemented in a virtual reality?
PhD student with mental health issues and bad performance
What can plausibly explain many of my very long and low-tech bridges?
Sharing one invocation list between multiple events on the same object in C#
PC video game involving floating islands doing aerial combat
How is it possible that Gollum speaks Westron?
Why is c4 bad when playing the London against a King's Indian?
Computing elements of a 1000 x 60 matrix exhausts RAM
What is a Mathematica packed array?Paging RAM in case of memory shortage issueHeavy duty operations, RAM, and ReadyboostEfficient calculation of diagonal matrix elementsHow to know each variable used how much RAMReplacing elements of a matrixHow to clear RAM memory in a running code?How does this code behave with more than 32GiB RAM?Replace diagonal elements in sparse matrixSolution list from “Solve” too large for my RAM spaceLimit the amount of RAM Mathematica may access?
$begingroup$
I am trying to compute a 1000 x 60 matrix or list of lists (and ideally this should go up to 1000 x 500 or 1000 x 1000).
Each element is the result of a FindRoot
operation, so I make my list by doing
Table[Flatten[h /. FindRoot[h == F[h, b, g], b, 1, 1000, g, 1, 60)
but 16GB of RAM are filled up. I think I should be able to hold list of lists much bigger than that, so probably using Table
with FindRoot
is causing Mathematica to store a lot of undeeded stuff in memory.
Here is the code:
ι[m_, n_] := Binomial[n, n*(1 - m)/2]*2^(-n);
f[m_, h_, b_, g_, n_] := (h*m + g/2*m^2) +
1/(n*b)*Log[ι[m, n]];
μ[m_, h_, b_, g_, n_] :=
Exp[b*n*f[m, h, b, g, n] + b*n*(-h + g/2)]/
Sum[Exp[b*n*f[x, h, b, g, n] + b*n*(-h + g/2)], x, -1 + 2/n,
1 - 2/n, 2/n];
moment[h_, x_, b_, g_, n_] := Sum[m^x*μ[m, h, b, g, n], m, -1 + 2/n, 1 - 2/n, 2/n];
var[h_, b_, g_, n_] := moment[h, 2, b, g, n] - moment[h, 1, b, g, n]^2;
cov[h_, b_, g_, n_] := moment[h, 3, b, g, n] - moment[h, 1, b, g,n]*moment[h, 2, b, g, n];
F[h_,b_,g_,n_]:= -d*b*(cov[h, b, gg, n] +
2 var[h, b, gg, n]);
n = 100;
d = 0.9;
glist = Table[g, g, 0.4, 1, 0.01];
blist = Table[b, b, 1.1, 10.1, 0.01];
heatdata = Table[
Flatten[h /.
FindRoot[
h == F[h,b,g,n], h, -0.01]][[1]]
, b, blist, g, glist];
performance-tuning memory
$endgroup$
add a comment |
$begingroup$
I am trying to compute a 1000 x 60 matrix or list of lists (and ideally this should go up to 1000 x 500 or 1000 x 1000).
Each element is the result of a FindRoot
operation, so I make my list by doing
Table[Flatten[h /. FindRoot[h == F[h, b, g], b, 1, 1000, g, 1, 60)
but 16GB of RAM are filled up. I think I should be able to hold list of lists much bigger than that, so probably using Table
with FindRoot
is causing Mathematica to store a lot of undeeded stuff in memory.
Here is the code:
ι[m_, n_] := Binomial[n, n*(1 - m)/2]*2^(-n);
f[m_, h_, b_, g_, n_] := (h*m + g/2*m^2) +
1/(n*b)*Log[ι[m, n]];
μ[m_, h_, b_, g_, n_] :=
Exp[b*n*f[m, h, b, g, n] + b*n*(-h + g/2)]/
Sum[Exp[b*n*f[x, h, b, g, n] + b*n*(-h + g/2)], x, -1 + 2/n,
1 - 2/n, 2/n];
moment[h_, x_, b_, g_, n_] := Sum[m^x*μ[m, h, b, g, n], m, -1 + 2/n, 1 - 2/n, 2/n];
var[h_, b_, g_, n_] := moment[h, 2, b, g, n] - moment[h, 1, b, g, n]^2;
cov[h_, b_, g_, n_] := moment[h, 3, b, g, n] - moment[h, 1, b, g,n]*moment[h, 2, b, g, n];
F[h_,b_,g_,n_]:= -d*b*(cov[h, b, gg, n] +
2 var[h, b, gg, n]);
n = 100;
d = 0.9;
glist = Table[g, g, 0.4, 1, 0.01];
blist = Table[b, b, 1.1, 10.1, 0.01];
heatdata = Table[
Flatten[h /.
FindRoot[
h == F[h,b,g,n], h, -0.01]][[1]]
, b, blist, g, glist];
performance-tuning memory
$endgroup$
$begingroup$
Please show a complete minimal example that reproduces the problem.
$endgroup$
– Szabolcs
May 19 at 8:31
1
$begingroup$
If it's a lot of code, that would be your actual code, not a minimal example. Please make an effort to track down the cause of the problem, and construct a small example that illustrates the problem. See here for guidance: mathematica.meta.stackexchange.com/q/2126/12
$endgroup$
– Szabolcs
May 19 at 8:34
1
$begingroup$
One possible issue is the memoization. Did you check how many values are actually saved? If you are working with floating point numbers, it may be the thing that eats up the memory.
$endgroup$
– Szabolcs
May 19 at 8:35
$begingroup$
Thanks, I've added the code as it isn't really that long. I've removed the memoization and I'm looking to see if this works now (the computation does take a while to run).
$endgroup$
– Three Diag
May 19 at 8:40
add a comment |
$begingroup$
I am trying to compute a 1000 x 60 matrix or list of lists (and ideally this should go up to 1000 x 500 or 1000 x 1000).
Each element is the result of a FindRoot
operation, so I make my list by doing
Table[Flatten[h /. FindRoot[h == F[h, b, g], b, 1, 1000, g, 1, 60)
but 16GB of RAM are filled up. I think I should be able to hold list of lists much bigger than that, so probably using Table
with FindRoot
is causing Mathematica to store a lot of undeeded stuff in memory.
Here is the code:
ι[m_, n_] := Binomial[n, n*(1 - m)/2]*2^(-n);
f[m_, h_, b_, g_, n_] := (h*m + g/2*m^2) +
1/(n*b)*Log[ι[m, n]];
μ[m_, h_, b_, g_, n_] :=
Exp[b*n*f[m, h, b, g, n] + b*n*(-h + g/2)]/
Sum[Exp[b*n*f[x, h, b, g, n] + b*n*(-h + g/2)], x, -1 + 2/n,
1 - 2/n, 2/n];
moment[h_, x_, b_, g_, n_] := Sum[m^x*μ[m, h, b, g, n], m, -1 + 2/n, 1 - 2/n, 2/n];
var[h_, b_, g_, n_] := moment[h, 2, b, g, n] - moment[h, 1, b, g, n]^2;
cov[h_, b_, g_, n_] := moment[h, 3, b, g, n] - moment[h, 1, b, g,n]*moment[h, 2, b, g, n];
F[h_,b_,g_,n_]:= -d*b*(cov[h, b, gg, n] +
2 var[h, b, gg, n]);
n = 100;
d = 0.9;
glist = Table[g, g, 0.4, 1, 0.01];
blist = Table[b, b, 1.1, 10.1, 0.01];
heatdata = Table[
Flatten[h /.
FindRoot[
h == F[h,b,g,n], h, -0.01]][[1]]
, b, blist, g, glist];
performance-tuning memory
$endgroup$
I am trying to compute a 1000 x 60 matrix or list of lists (and ideally this should go up to 1000 x 500 or 1000 x 1000).
Each element is the result of a FindRoot
operation, so I make my list by doing
Table[Flatten[h /. FindRoot[h == F[h, b, g], b, 1, 1000, g, 1, 60)
but 16GB of RAM are filled up. I think I should be able to hold list of lists much bigger than that, so probably using Table
with FindRoot
is causing Mathematica to store a lot of undeeded stuff in memory.
Here is the code:
ι[m_, n_] := Binomial[n, n*(1 - m)/2]*2^(-n);
f[m_, h_, b_, g_, n_] := (h*m + g/2*m^2) +
1/(n*b)*Log[ι[m, n]];
μ[m_, h_, b_, g_, n_] :=
Exp[b*n*f[m, h, b, g, n] + b*n*(-h + g/2)]/
Sum[Exp[b*n*f[x, h, b, g, n] + b*n*(-h + g/2)], x, -1 + 2/n,
1 - 2/n, 2/n];
moment[h_, x_, b_, g_, n_] := Sum[m^x*μ[m, h, b, g, n], m, -1 + 2/n, 1 - 2/n, 2/n];
var[h_, b_, g_, n_] := moment[h, 2, b, g, n] - moment[h, 1, b, g, n]^2;
cov[h_, b_, g_, n_] := moment[h, 3, b, g, n] - moment[h, 1, b, g,n]*moment[h, 2, b, g, n];
F[h_,b_,g_,n_]:= -d*b*(cov[h, b, gg, n] +
2 var[h, b, gg, n]);
n = 100;
d = 0.9;
glist = Table[g, g, 0.4, 1, 0.01];
blist = Table[b, b, 1.1, 10.1, 0.01];
heatdata = Table[
Flatten[h /.
FindRoot[
h == F[h,b,g,n], h, -0.01]][[1]]
, b, blist, g, glist];
performance-tuning memory
performance-tuning memory
edited May 20 at 0:11
m_goldberg
90.4k873203
90.4k873203
asked May 19 at 8:24
Three DiagThree Diag
346111
346111
$begingroup$
Please show a complete minimal example that reproduces the problem.
$endgroup$
– Szabolcs
May 19 at 8:31
1
$begingroup$
If it's a lot of code, that would be your actual code, not a minimal example. Please make an effort to track down the cause of the problem, and construct a small example that illustrates the problem. See here for guidance: mathematica.meta.stackexchange.com/q/2126/12
$endgroup$
– Szabolcs
May 19 at 8:34
1
$begingroup$
One possible issue is the memoization. Did you check how many values are actually saved? If you are working with floating point numbers, it may be the thing that eats up the memory.
$endgroup$
– Szabolcs
May 19 at 8:35
$begingroup$
Thanks, I've added the code as it isn't really that long. I've removed the memoization and I'm looking to see if this works now (the computation does take a while to run).
$endgroup$
– Three Diag
May 19 at 8:40
add a comment |
$begingroup$
Please show a complete minimal example that reproduces the problem.
$endgroup$
– Szabolcs
May 19 at 8:31
1
$begingroup$
If it's a lot of code, that would be your actual code, not a minimal example. Please make an effort to track down the cause of the problem, and construct a small example that illustrates the problem. See here for guidance: mathematica.meta.stackexchange.com/q/2126/12
$endgroup$
– Szabolcs
May 19 at 8:34
1
$begingroup$
One possible issue is the memoization. Did you check how many values are actually saved? If you are working with floating point numbers, it may be the thing that eats up the memory.
$endgroup$
– Szabolcs
May 19 at 8:35
$begingroup$
Thanks, I've added the code as it isn't really that long. I've removed the memoization and I'm looking to see if this works now (the computation does take a while to run).
$endgroup$
– Three Diag
May 19 at 8:40
$begingroup$
Please show a complete minimal example that reproduces the problem.
$endgroup$
– Szabolcs
May 19 at 8:31
$begingroup$
Please show a complete minimal example that reproduces the problem.
$endgroup$
– Szabolcs
May 19 at 8:31
1
1
$begingroup$
If it's a lot of code, that would be your actual code, not a minimal example. Please make an effort to track down the cause of the problem, and construct a small example that illustrates the problem. See here for guidance: mathematica.meta.stackexchange.com/q/2126/12
$endgroup$
– Szabolcs
May 19 at 8:34
$begingroup$
If it's a lot of code, that would be your actual code, not a minimal example. Please make an effort to track down the cause of the problem, and construct a small example that illustrates the problem. See here for guidance: mathematica.meta.stackexchange.com/q/2126/12
$endgroup$
– Szabolcs
May 19 at 8:34
1
1
$begingroup$
One possible issue is the memoization. Did you check how many values are actually saved? If you are working with floating point numbers, it may be the thing that eats up the memory.
$endgroup$
– Szabolcs
May 19 at 8:35
$begingroup$
One possible issue is the memoization. Did you check how many values are actually saved? If you are working with floating point numbers, it may be the thing that eats up the memory.
$endgroup$
– Szabolcs
May 19 at 8:35
$begingroup$
Thanks, I've added the code as it isn't really that long. I've removed the memoization and I'm looking to see if this works now (the computation does take a while to run).
$endgroup$
– Three Diag
May 19 at 8:40
$begingroup$
Thanks, I've added the code as it isn't really that long. I've removed the memoization and I'm looking to see if this works now (the computation does take a while to run).
$endgroup$
– Three Diag
May 19 at 8:40
add a comment |
1 Answer
1
active
oldest
votes
$begingroup$
Your function F
is implemented really, really inefficiently. By quite simple means and in the proposed situation, it can be sped up by a factor of 20000. The key is to start with calculations in machine precision as early as possible and to store frequently used data in packed arrays.
n = 100;
mlist = Range[-1. + 2/n, 1. - 2/n, 2./n];
m2list = mlist^2;
m3list = mlist^3;
logiotalist = Log[Binomial[n, n*(1 - mlist)/2]*2^(-n)];
d = 0.9;
glist = Range[0.4, 1, 0.01];
blist = Range[1.1, 10.1, 0.01];
ClearAll[F];
F[h_?NumericQ, b_, g_] :=
Module[var, cov, explist, μlist, mom1, mom2, mom3,
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
μlist = explist/Total[explist];
mom1 = μlist.mlist;
mom2 = μlist.m2list;
mom3 = μlist.m3list;
var = Subtract[mom2, mom1 mom1];
cov = Subtract[mom3, mom1 mom2];
(-d b) (cov + 2. var)
];
Just a quick test for precision and speed:
t1, r1 = F[0.1, blist[[1]], glist[[1]], n] // RepeatedTiming;
t2, r2 = Fnew[0.1, blist[[1]], glist[[1]]] // RepeatedTiming;
Abs[r1 - r2]/r1
t1/t2
-1.32375*10^-14
2.1*10^4
Now the parallelized solve loop requires about 10 seconds on my Quad Core Haswell CPU:
ParallelEvaluate[Off[General::munfl]];
heatdata = Developer`ToPackedArray[
ParallelTable[
Block[h0, h,
h0 = -0.01;
Developer`ToPackedArray[
Table[
h0 = h /. FindRoot[h == Fnew[h, b, g], h, h0],
b, blist]
]
],
g, glist]
]; // AbsoluteTiming // First
10.072
Memory considerations
You also see: Limited amount of RAM is not an issue here. That must have been caused by excessive memoziation. For the timing, it is crucial how information is stored and retrieved.
Storing computed values in a packed array for retrieving them later is significantly more efficient than memoization. Memoization into DownValues
uses a complex data structure such as a hash table at its backend, and this data structure has certain overhead. In contrast, a packed array represents basically a connected block of physical memory, accompanied by some bytes of meta information (array dimensions and maybe some row pointers). Moreover, computation with data stored in packed arrays can take advantage of vectorization, which is most crucially employed in the following line:
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
Remark on precision
Finally, I have to note that there is numerical underflow occurring in the course of the computation. This is probably caused by calling Exp
with negative numbers of oversized absolute value. I decided to turn off the warning message, but this may lead to a significant loss of precision. So use with care. If one wants to do it correctly, one should investigate this further and apply, e.g. Clip
or Threshold
.
$endgroup$
$begingroup$
Thanks for this very helpful and informative answer!
$endgroup$
– Three Diag
May 19 at 23:07
$begingroup$
You're welcome!
$endgroup$
– Henrik Schumacher
May 20 at 6:40
add a comment |
Your Answer
StackExchange.ready(function()
var channelOptions =
tags: "".split(" "),
id: "387"
;
initTagRenderer("".split(" "), "".split(" "), channelOptions);
StackExchange.using("externalEditor", function()
// Have to fire editor after snippets, if snippets enabled
if (StackExchange.settings.snippets.snippetsEnabled)
StackExchange.using("snippets", function()
createEditor();
);
else
createEditor();
);
function createEditor()
StackExchange.prepareEditor(
heartbeatType: 'answer',
autoActivateHeartbeat: false,
convertImagesToLinks: false,
noModals: true,
showLowRepImageUploadWarning: true,
reputationToPostImages: null,
bindNavPrevention: true,
postfix: "",
imageUploader:
brandingHtml: "Powered by u003ca class="icon-imgur-white" href="https://imgur.com/"u003eu003c/au003e",
contentPolicyHtml: "User contributions licensed under u003ca href="https://creativecommons.org/licenses/by-sa/3.0/"u003ecc by-sa 3.0 with attribution requiredu003c/au003e u003ca href="https://stackoverflow.com/legal/content-policy"u003e(content policy)u003c/au003e",
allowUrls: true
,
onDemand: true,
discardSelector: ".discard-answer"
,immediatelyShowMarkdownHelp:true
);
);
Sign up or log in
StackExchange.ready(function ()
StackExchange.helpers.onClickDraftSave('#login-link');
);
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
StackExchange.ready(
function ()
StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fmathematica.stackexchange.com%2fquestions%2f198651%2fcomputing-elements-of-a-1000-x-60-matrix-exhausts-ram%23new-answer', 'question_page');
);
Post as a guest
Required, but never shown
1 Answer
1
active
oldest
votes
1 Answer
1
active
oldest
votes
active
oldest
votes
active
oldest
votes
$begingroup$
Your function F
is implemented really, really inefficiently. By quite simple means and in the proposed situation, it can be sped up by a factor of 20000. The key is to start with calculations in machine precision as early as possible and to store frequently used data in packed arrays.
n = 100;
mlist = Range[-1. + 2/n, 1. - 2/n, 2./n];
m2list = mlist^2;
m3list = mlist^3;
logiotalist = Log[Binomial[n, n*(1 - mlist)/2]*2^(-n)];
d = 0.9;
glist = Range[0.4, 1, 0.01];
blist = Range[1.1, 10.1, 0.01];
ClearAll[F];
F[h_?NumericQ, b_, g_] :=
Module[var, cov, explist, μlist, mom1, mom2, mom3,
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
μlist = explist/Total[explist];
mom1 = μlist.mlist;
mom2 = μlist.m2list;
mom3 = μlist.m3list;
var = Subtract[mom2, mom1 mom1];
cov = Subtract[mom3, mom1 mom2];
(-d b) (cov + 2. var)
];
Just a quick test for precision and speed:
t1, r1 = F[0.1, blist[[1]], glist[[1]], n] // RepeatedTiming;
t2, r2 = Fnew[0.1, blist[[1]], glist[[1]]] // RepeatedTiming;
Abs[r1 - r2]/r1
t1/t2
-1.32375*10^-14
2.1*10^4
Now the parallelized solve loop requires about 10 seconds on my Quad Core Haswell CPU:
ParallelEvaluate[Off[General::munfl]];
heatdata = Developer`ToPackedArray[
ParallelTable[
Block[h0, h,
h0 = -0.01;
Developer`ToPackedArray[
Table[
h0 = h /. FindRoot[h == Fnew[h, b, g], h, h0],
b, blist]
]
],
g, glist]
]; // AbsoluteTiming // First
10.072
Memory considerations
You also see: Limited amount of RAM is not an issue here. That must have been caused by excessive memoziation. For the timing, it is crucial how information is stored and retrieved.
Storing computed values in a packed array for retrieving them later is significantly more efficient than memoization. Memoization into DownValues
uses a complex data structure such as a hash table at its backend, and this data structure has certain overhead. In contrast, a packed array represents basically a connected block of physical memory, accompanied by some bytes of meta information (array dimensions and maybe some row pointers). Moreover, computation with data stored in packed arrays can take advantage of vectorization, which is most crucially employed in the following line:
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
Remark on precision
Finally, I have to note that there is numerical underflow occurring in the course of the computation. This is probably caused by calling Exp
with negative numbers of oversized absolute value. I decided to turn off the warning message, but this may lead to a significant loss of precision. So use with care. If one wants to do it correctly, one should investigate this further and apply, e.g. Clip
or Threshold
.
$endgroup$
$begingroup$
Thanks for this very helpful and informative answer!
$endgroup$
– Three Diag
May 19 at 23:07
$begingroup$
You're welcome!
$endgroup$
– Henrik Schumacher
May 20 at 6:40
add a comment |
$begingroup$
Your function F
is implemented really, really inefficiently. By quite simple means and in the proposed situation, it can be sped up by a factor of 20000. The key is to start with calculations in machine precision as early as possible and to store frequently used data in packed arrays.
n = 100;
mlist = Range[-1. + 2/n, 1. - 2/n, 2./n];
m2list = mlist^2;
m3list = mlist^3;
logiotalist = Log[Binomial[n, n*(1 - mlist)/2]*2^(-n)];
d = 0.9;
glist = Range[0.4, 1, 0.01];
blist = Range[1.1, 10.1, 0.01];
ClearAll[F];
F[h_?NumericQ, b_, g_] :=
Module[var, cov, explist, μlist, mom1, mom2, mom3,
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
μlist = explist/Total[explist];
mom1 = μlist.mlist;
mom2 = μlist.m2list;
mom3 = μlist.m3list;
var = Subtract[mom2, mom1 mom1];
cov = Subtract[mom3, mom1 mom2];
(-d b) (cov + 2. var)
];
Just a quick test for precision and speed:
t1, r1 = F[0.1, blist[[1]], glist[[1]], n] // RepeatedTiming;
t2, r2 = Fnew[0.1, blist[[1]], glist[[1]]] // RepeatedTiming;
Abs[r1 - r2]/r1
t1/t2
-1.32375*10^-14
2.1*10^4
Now the parallelized solve loop requires about 10 seconds on my Quad Core Haswell CPU:
ParallelEvaluate[Off[General::munfl]];
heatdata = Developer`ToPackedArray[
ParallelTable[
Block[h0, h,
h0 = -0.01;
Developer`ToPackedArray[
Table[
h0 = h /. FindRoot[h == Fnew[h, b, g], h, h0],
b, blist]
]
],
g, glist]
]; // AbsoluteTiming // First
10.072
Memory considerations
You also see: Limited amount of RAM is not an issue here. That must have been caused by excessive memoziation. For the timing, it is crucial how information is stored and retrieved.
Storing computed values in a packed array for retrieving them later is significantly more efficient than memoization. Memoization into DownValues
uses a complex data structure such as a hash table at its backend, and this data structure has certain overhead. In contrast, a packed array represents basically a connected block of physical memory, accompanied by some bytes of meta information (array dimensions and maybe some row pointers). Moreover, computation with data stored in packed arrays can take advantage of vectorization, which is most crucially employed in the following line:
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
Remark on precision
Finally, I have to note that there is numerical underflow occurring in the course of the computation. This is probably caused by calling Exp
with negative numbers of oversized absolute value. I decided to turn off the warning message, but this may lead to a significant loss of precision. So use with care. If one wants to do it correctly, one should investigate this further and apply, e.g. Clip
or Threshold
.
$endgroup$
$begingroup$
Thanks for this very helpful and informative answer!
$endgroup$
– Three Diag
May 19 at 23:07
$begingroup$
You're welcome!
$endgroup$
– Henrik Schumacher
May 20 at 6:40
add a comment |
$begingroup$
Your function F
is implemented really, really inefficiently. By quite simple means and in the proposed situation, it can be sped up by a factor of 20000. The key is to start with calculations in machine precision as early as possible and to store frequently used data in packed arrays.
n = 100;
mlist = Range[-1. + 2/n, 1. - 2/n, 2./n];
m2list = mlist^2;
m3list = mlist^3;
logiotalist = Log[Binomial[n, n*(1 - mlist)/2]*2^(-n)];
d = 0.9;
glist = Range[0.4, 1, 0.01];
blist = Range[1.1, 10.1, 0.01];
ClearAll[F];
F[h_?NumericQ, b_, g_] :=
Module[var, cov, explist, μlist, mom1, mom2, mom3,
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
μlist = explist/Total[explist];
mom1 = μlist.mlist;
mom2 = μlist.m2list;
mom3 = μlist.m3list;
var = Subtract[mom2, mom1 mom1];
cov = Subtract[mom3, mom1 mom2];
(-d b) (cov + 2. var)
];
Just a quick test for precision and speed:
t1, r1 = F[0.1, blist[[1]], glist[[1]], n] // RepeatedTiming;
t2, r2 = Fnew[0.1, blist[[1]], glist[[1]]] // RepeatedTiming;
Abs[r1 - r2]/r1
t1/t2
-1.32375*10^-14
2.1*10^4
Now the parallelized solve loop requires about 10 seconds on my Quad Core Haswell CPU:
ParallelEvaluate[Off[General::munfl]];
heatdata = Developer`ToPackedArray[
ParallelTable[
Block[h0, h,
h0 = -0.01;
Developer`ToPackedArray[
Table[
h0 = h /. FindRoot[h == Fnew[h, b, g], h, h0],
b, blist]
]
],
g, glist]
]; // AbsoluteTiming // First
10.072
Memory considerations
You also see: Limited amount of RAM is not an issue here. That must have been caused by excessive memoziation. For the timing, it is crucial how information is stored and retrieved.
Storing computed values in a packed array for retrieving them later is significantly more efficient than memoization. Memoization into DownValues
uses a complex data structure such as a hash table at its backend, and this data structure has certain overhead. In contrast, a packed array represents basically a connected block of physical memory, accompanied by some bytes of meta information (array dimensions and maybe some row pointers). Moreover, computation with data stored in packed arrays can take advantage of vectorization, which is most crucially employed in the following line:
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
Remark on precision
Finally, I have to note that there is numerical underflow occurring in the course of the computation. This is probably caused by calling Exp
with negative numbers of oversized absolute value. I decided to turn off the warning message, but this may lead to a significant loss of precision. So use with care. If one wants to do it correctly, one should investigate this further and apply, e.g. Clip
or Threshold
.
$endgroup$
Your function F
is implemented really, really inefficiently. By quite simple means and in the proposed situation, it can be sped up by a factor of 20000. The key is to start with calculations in machine precision as early as possible and to store frequently used data in packed arrays.
n = 100;
mlist = Range[-1. + 2/n, 1. - 2/n, 2./n];
m2list = mlist^2;
m3list = mlist^3;
logiotalist = Log[Binomial[n, n*(1 - mlist)/2]*2^(-n)];
d = 0.9;
glist = Range[0.4, 1, 0.01];
blist = Range[1.1, 10.1, 0.01];
ClearAll[F];
F[h_?NumericQ, b_, g_] :=
Module[var, cov, explist, μlist, mom1, mom2, mom3,
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
μlist = explist/Total[explist];
mom1 = μlist.mlist;
mom2 = μlist.m2list;
mom3 = μlist.m3list;
var = Subtract[mom2, mom1 mom1];
cov = Subtract[mom3, mom1 mom2];
(-d b) (cov + 2. var)
];
Just a quick test for precision and speed:
t1, r1 = F[0.1, blist[[1]], glist[[1]], n] // RepeatedTiming;
t2, r2 = Fnew[0.1, blist[[1]], glist[[1]]] // RepeatedTiming;
Abs[r1 - r2]/r1
t1/t2
-1.32375*10^-14
2.1*10^4
Now the parallelized solve loop requires about 10 seconds on my Quad Core Haswell CPU:
ParallelEvaluate[Off[General::munfl]];
heatdata = Developer`ToPackedArray[
ParallelTable[
Block[h0, h,
h0 = -0.01;
Developer`ToPackedArray[
Table[
h0 = h /. FindRoot[h == Fnew[h, b, g], h, h0],
b, blist]
]
],
g, glist]
]; // AbsoluteTiming // First
10.072
Memory considerations
You also see: Limited amount of RAM is not an issue here. That must have been caused by excessive memoziation. For the timing, it is crucial how information is stored and retrieved.
Storing computed values in a packed array for retrieving them later is significantly more efficient than memoization. Memoization into DownValues
uses a complex data structure such as a hash table at its backend, and this data structure has certain overhead. In contrast, a packed array represents basically a connected block of physical memory, accompanied by some bytes of meta information (array dimensions and maybe some row pointers). Moreover, computation with data stored in packed arrays can take advantage of vectorization, which is most crucially employed in the following line:
explist = Exp[(b n h) mlist + (b n g/2) m2list + logiotalist + b n (-h + g/2)];
Remark on precision
Finally, I have to note that there is numerical underflow occurring in the course of the computation. This is probably caused by calling Exp
with negative numbers of oversized absolute value. I decided to turn off the warning message, but this may lead to a significant loss of precision. So use with care. If one wants to do it correctly, one should investigate this further and apply, e.g. Clip
or Threshold
.
edited May 19 at 17:44
answered May 19 at 10:03
Henrik SchumacherHenrik Schumacher
63.5k589177
63.5k589177
$begingroup$
Thanks for this very helpful and informative answer!
$endgroup$
– Three Diag
May 19 at 23:07
$begingroup$
You're welcome!
$endgroup$
– Henrik Schumacher
May 20 at 6:40
add a comment |
$begingroup$
Thanks for this very helpful and informative answer!
$endgroup$
– Three Diag
May 19 at 23:07
$begingroup$
You're welcome!
$endgroup$
– Henrik Schumacher
May 20 at 6:40
$begingroup$
Thanks for this very helpful and informative answer!
$endgroup$
– Three Diag
May 19 at 23:07
$begingroup$
Thanks for this very helpful and informative answer!
$endgroup$
– Three Diag
May 19 at 23:07
$begingroup$
You're welcome!
$endgroup$
– Henrik Schumacher
May 20 at 6:40
$begingroup$
You're welcome!
$endgroup$
– Henrik Schumacher
May 20 at 6:40
add a comment |
Thanks for contributing an answer to Mathematica Stack Exchange!
- Please be sure to answer the question. Provide details and share your research!
But avoid …
- Asking for help, clarification, or responding to other answers.
- Making statements based on opinion; back them up with references or personal experience.
Use MathJax to format equations. MathJax reference.
To learn more, see our tips on writing great answers.
Sign up or log in
StackExchange.ready(function ()
StackExchange.helpers.onClickDraftSave('#login-link');
);
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
StackExchange.ready(
function ()
StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fmathematica.stackexchange.com%2fquestions%2f198651%2fcomputing-elements-of-a-1000-x-60-matrix-exhausts-ram%23new-answer', 'question_page');
);
Post as a guest
Required, but never shown
Sign up or log in
StackExchange.ready(function ()
StackExchange.helpers.onClickDraftSave('#login-link');
);
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
Sign up or log in
StackExchange.ready(function ()
StackExchange.helpers.onClickDraftSave('#login-link');
);
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
Sign up or log in
StackExchange.ready(function ()
StackExchange.helpers.onClickDraftSave('#login-link');
);
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
$begingroup$
Please show a complete minimal example that reproduces the problem.
$endgroup$
– Szabolcs
May 19 at 8:31
1
$begingroup$
If it's a lot of code, that would be your actual code, not a minimal example. Please make an effort to track down the cause of the problem, and construct a small example that illustrates the problem. See here for guidance: mathematica.meta.stackexchange.com/q/2126/12
$endgroup$
– Szabolcs
May 19 at 8:34
1
$begingroup$
One possible issue is the memoization. Did you check how many values are actually saved? If you are working with floating point numbers, it may be the thing that eats up the memory.
$endgroup$
– Szabolcs
May 19 at 8:35
$begingroup$
Thanks, I've added the code as it isn't really that long. I've removed the memoization and I'm looking to see if this works now (the computation does take a while to run).
$endgroup$
– Three Diag
May 19 at 8:40