MATLAB Central Newsreader - decrease in speed due to appending a row vector

decrease in speed due to appending a row vector

Brian — Thu, 24 May 2007 19:44:44 -0400

I have a row vector and I am adding new data elements in between at
different locations. The approach i take is copy the vector in a
temporary variable, add the new element at the desired location and
then append the remaining elements from the temporary variable. I
believed it should be very fast in Matlab but if i repeat it 15000
times it takes me 10 seconds which is very slow for my application.
Is there any other way to get around this problem.

Re: decrease in speed due to appending a row vecto

Joseph — Thu, 24 May 2007 20:17:29 -0400

pre-allocate the space for your vector

help zeros

Brian wrote:
>
>
> I have a row vector and I am adding new data elements in between at
> different locations. The approach i take is copy the vector in a
> temporary variable, add the new element at the desired location and
> then append the remaining elements from the temporary variable. I
> believed it should be very fast in Matlab but if i repeat it 15000
> times it takes me 10 seconds which is very slow for my application.
> Is there any other way to get around this problem.

Re: decrease in speed due to appending a row vector

Brian — Thu, 24 May 2007 20:20:24 -0400

I don't know how big will the vector grow each time. The space
depends on the input data which will change most of the time and I
have no way of know the final size.

Joseph wrote:
>
>
> pre-allocate the space for your vector
>
> help zeros
>
> Brian wrote:
>>
>>
>> I have a row vector and I am adding new data elements in
between
> at
>> different locations. The approach i take is copy the vector in
a
>> temporary variable, add the new element at the desired location
> and
>> then append the remaining elements from the temporary variable.
I
>> believed it should be very fast in Matlab but if i repeat it
> 15000
>> times it takes me 10 seconds which is very slow for my
> application.
>> Is there any other way to get around this problem.

Re: decrease in speed due to appending a row vector

John D'Errico — Thu, 24 May 2007 20:35:36 -0400

Brian wrote:
>
>
> I don't know how big will the vector grow each time. The space
> depends on the input data which will change most of the time and I
> have no way of know the final size.

Take a look at my growdata or growdata2
codes. They came from discussions here
on c.s-s.m, and were written to allow
arrays to grow in size with a minimum
penalty in time. My goal was to allow
arrays to grow in size to potentially
millions of elements, without advance
knowledge of the final count. These codes
use cell arrays, each cell of which is
preallocated to use moderately large
but still manageable chunks of memory.
Arrays with huge numbers of cells can
have problems too.

<http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=8334&objectType=file>

Of course, its always better to just
preallocate your large arrays whenever
you can.

HTh,
John

Re: decrease in speed due to appending a row vector

NZTideMan — Thu, 24 May 2007 18:18:56 -0400

On May 25, 12:35 pm, "John D'Errico" <woodch...@rochester.rr.com>
wrote:
> Brian wrote:
>
> > I don't know how big will the vector grow each time. The space
> > depends on the input data which will change most of the time and I
> > have no way of know the final size.
>
> Take a look at my growdata or growdata2
> codes. They came from discussions here
> on c.s-s.m, and were written to allow
> arrays to grow in size with a minimum
> penalty in time. My goal was to allow
> arrays to grow in size to potentially
> millions of elements, without advance
> knowledge of the final count. These codes
> use cell arrays, each cell of which is
> preallocated to use moderately large
> but still manageable chunks of memory.
> Arrays with huge numbers of cells can
> have problems too.
>
> <http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objec...>
>
> Of course, its always better to just
> preallocate your large arrays whenever
> you can.
>
> HTh,
> John

I didn't know about your growdata routines, John, so I have devised
the following way for accumulating results of a Monte Carlo
simulation:
I generate a structure:
s(irun).ymx=a;
where irun is the Monte Carlo run number, and the result of the
simulation a(1,n) is a row vector where n is the number of events
(drawn from a Poisson distribution). n varies from run to run.
I understand that when you do it this way, each array in s uses
contiguous memory, but they are not contiguous within the structure.
Thus, the memory allocation problem does not arise.
I'd appreciate your comments on this.

Re: decrease in speed due to appending a row vector

NZTideMan — Thu, 24 May 2007 18:37:29 -0400

Re: decrease in speed due to appending a row vector

John D'Errico — Thu, 24 May 2007 22:50:37 -0400

NZTideMan wrote:
>
>
> On May 25, 12:35 pm, "John D'Errico"
<woodch...@rochester.rr.com>
> wrote:
>> Brian wrote:
>>
>> > I don't know how big will the vector grow each time. The
space
>> > depends on the input data which will change most of the
time
> and I
>> > have no way of know the final size.
>>
>> Take a look at my growdata or growdata2
>> codes. They came from discussions here
>> on c.s-s.m, and were written to allow
>> arrays to grow in size with a minimum
>> penalty in time. My goal was to allow
>> arrays to grow in size to potentially
>> millions of elements, without advance
>> knowledge of the final count. These codes
>> use cell arrays, each cell of which is
>> preallocated to use moderately large
>> but still manageable chunks of memory.
>> Arrays with huge numbers of cells can
>> have problems too.
>>
>> <http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objec...>
>>
>> Of course, its always better to just
>> preallocate your large arrays whenever
>> you can.
>>
>> HTh,
>> John
>
> I didn't know about your growdata routines, John, so I have devised
> the following way for accumulating results of a Monte Carlo
> simulation:
> I generate a structure:
> s(irun).ymx=a;
> where irun is the Monte Carlo run number, and the result of the
> simulation a(1,n) is a row vector where n is the number of events
> (drawn from a Poisson distribution). n varies from run to run.
> I understand that when you do it this way, each array in s uses
> contiguous memory, but they are not contiguous within the
> structure.
> Thus, the memory allocation problem does not arise.
> I'd appreciate your comments on this.

This is also a reassonable way to solve the
problem.

Use of a single cell for each new appended
element is another way that is very similar
to the structure approach that you used.
Use of pointers here is more efficient
than growing and re-allocating the entire
array. I'm not sure about use of structures,
as I thought that structures were not
efficient users of memory as compared to
cell arrays. This is probably changing
with each matlab release of course. (See
below for a comparison.)

An issue is how many total elements that
you may expect to see. In some earlier
testing, I found that cell arrays with
millions of cells caused some serious
problems, however a few thousand, or even
ten thousand or so cells was not a problem.

I am starting to wonder if its time to
reassess those growdata codes as Matlab
evolves. In some recent tests it seemed
that the problem I had seen with creating
large cell arrays with millions of elements
may no longer be as much of an issue in
recent releases. The underlying matlab
codes must have been improved. I'll want
to compare how use of a structure compares
to a cell array for memory and speed.

Some quick testing shows a structure and
a cell array to be similar in time and
memory here.

tic
for i = 1:1000
  a(i).X =rand(1,3);
end,
whos a
toc

  Name Size Bytes Class Attributes
  a 1x1000 84064 struct

Elapsed time is 0.023403 seconds.

tic
for i = 1:1000
  b{i} =rand(1,3);
end
whos b
toc

  Name Size Bytes Class Attributes
  b 1x1000 84000 cell

Elapsed time is 0.022017 seconds.

The question is, do these scale well with
larger numbers of append operations?

tic
for i = 1:100000
  a(i).X =rand(1,3);
end
whos a
toc

  Name Size Bytes Class Attributes
  a 1x100000 8400064 struct

Elapsed time is 118.482410 seconds.

This has grown by more than 100x. In fact,
it grew in time by a factor of roughly
10000, or 100^2. It is this quadratic
growth in time that I was trying to avoid
with the growdata codes.

How about growdata2 here?

tic
g=growdata2;
for i = 1:100000
  g(rand(1,3))
end
g=g();
whos g
toc

  Name Size Bytes Class Attributes
  g 100000x3 2400000 double

Elapsed time is 12.672389 seconds.

So if you are appending only a few thousand
elements, then either the structure or the
cell array will work quite well. If you
expect to append hundreds of thousands or
more chunks, then the hybrid approach in
the growdata tools is still more reasonable.

Your mileage may vary with a newer release
yet than R2006b. I've not gotten the latest
release installed until I install a new OS
version here.

John

Re: decrease in speed due to appending a row vector

NZTideMan — Thu, 24 May 2007 20:35:02 -0400

On May 25, 2:50 pm, "John D'Errico" <woodch...@rochester.rr.com>
wrote:
> NZTideMan wrote:
>
> > On May 25, 12:35 pm, "John D'Errico"
>
> <woodch...@rochester.rr.com>
>
>
>
>
>
> > wrote:
> >> Brian wrote:
>
> >> > I don't know how big will the vector grow each time. The
> space
> >> > depends on the input data which will change most of the
> time
> > and I
> >> > have no way of know the final size.
>
> >> Take a look at my growdata or growdata2
> >> codes. They came from discussions here
> >> on c.s-s.m, and were written to allow
> >> arrays to grow in size with a minimum
> >> penalty in time. My goal was to allow
> >> arrays to grow in size to potentially
> >> millions of elements, without advance
> >> knowledge of the final count. These codes
> >> use cell arrays, each cell of which is
> >> preallocated to use moderately large
> >> but still manageable chunks of memory.
> >> Arrays with huge numbers of cells can
> >> have problems too.
>
> >> <http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objec...>
>
> >> Of course, its always better to just
> >> preallocate your large arrays whenever
> >> you can.
>
> >> HTh,
> >> John
>
> > I didn't know about your growdata routines, John, so I have devised
> > the following way for accumulating results of a Monte Carlo
> > simulation:
> > I generate a structure:
> > s(irun).ymx=a;
> > where irun is the Monte Carlo run number, and the result of the
> > simulation a(1,n) is a row vector where n is the number of events
> > (drawn from a Poisson distribution). n varies from run to run.
> > I understand that when you do it this way, each array in s uses
> > contiguous memory, but they are not contiguous within the
> > structure.
> > Thus, the memory allocation problem does not arise.
> > I'd appreciate your comments on this.
>
> This is also a reassonable way to solve the
> problem.
>
> Use of a single cell for each new appended
> element is another way that is very similar
> to the structure approach that you used.
> Use of pointers here is more efficient
> than growing and re-allocating the entire
> array. I'm not sure about use of structures,
> as I thought that structures were not
> efficient users of memory as compared to
> cell arrays. This is probably changing
> with each matlab release of course. (See
> below for a comparison.)
>
> An issue is how many total elements that
> you may expect to see. In some earlier
> testing, I found that cell arrays with
> millions of cells caused some serious
> problems, however a few thousand, or even
> ten thousand or so cells was not a problem.
>
> I am starting to wonder if its time to
> reassess those growdata codes as Matlab
> evolves. In some recent tests it seemed
> that the problem I had seen with creating
> large cell arrays with millions of elements
> may no longer be as much of an issue in
> recent releases. The underlying matlab
> codes must have been improved. I'll want
> to compare how use of a structure compares
> to a cell array for memory and speed.
>
> Some quick testing shows a structure and
> a cell array to be similar in time and
> memory here.
>
> tic
> for i = 1:1000
> a(i).X =rand(1,3);
> end,
> whos a
> toc
>
> Name Size Bytes Class Attributes
> a 1x1000 84064 struct
>
> Elapsed time is 0.023403 seconds.
>
> tic
> for i = 1:1000
> b{i} =rand(1,3);
> end
> whos b
> toc
>
> Name Size Bytes Class Attributes
> b 1x1000 84000 cell
>
> Elapsed time is 0.022017 seconds.
>
> The question is, do these scale well with
> larger numbers of append operations?
>
> tic
> for i = 1:100000
> a(i).X =rand(1,3);
> end
> whos a
> toc
>
> Name Size Bytes Class Attributes
> a 1x100000 8400064 struct
>
> Elapsed time is 118.482410 seconds.
>
> This has grown by more than 100x. In fact,
> it grew in time by a factor of roughly
> 10000, or 100^2. It is this quadratic
> growth in time that I was trying to avoid
> with the growdata codes.
>
> How about growdata2 here?
>
> tic
> g=growdata2;
> for i = 1:100000
> g(rand(1,3))
> end
> g=g();
> whos g
> toc
>
> Name Size Bytes Class Attributes
> g 100000x3 2400000 double
>
> Elapsed time is 12.672389 seconds.
>
> So if you are appending only a few thousand
> elements, then either the structure or the
> cell array will work quite well. If you
> expect to append hundreds of thousands or
> more chunks, then the hybrid approach in
> the growdata tools is still more reasonable.
>
> Your mileage may vary with a newer release
> yet than R2006b. I've not gotten the latest
> release installed until I install a new OS
> version here.
>
> John- Hide quoted text -
>
> - Show quoted text -

That's pretty convincing!
I guess the lesson is that if you have toy problems, using a structure
works fine, but for real-world problems growdata2 is the way to go.
But an advantage of the structure approach is that at the end I have
all the runs separated.
Of course, I can concatenate them if I choose using square brackets.

Re: decrease in speed due to appending a row vecto

Eric Sampson — Fri, 25 May 2007 10:05:05 -0400

John D'Errico wrote:
>
>
(snip)
> Some quick testing shows a structure and
> a cell array to be similar in time and
> memory here.
>
> tic
> for i = 1:1000
> a(i).X =rand(1,3);
> end,
> whos a
> toc
>
> Name Size Bytes Class Attributes
> a 1x1000 84064 struct
>
> Elapsed time is 0.023403 seconds.
>
> tic
> for i = 1:1000
> b{i} =rand(1,3);
> end
> whos b
> toc
>
> Name Size Bytes Class Attributes
> b 1x1000 84000 cell
>
> Elapsed time is 0.022017 seconds.
>
> The question is, do these scale well with
> larger numbers of append operations?
>
> tic
> for i = 1:100000
> a(i).X =rand(1,3);
> end
> whos a
> toc
>
> Name Size Bytes Class Attributes
> a 1x100000 8400064 struct
>
> Elapsed time is 118.482410 seconds.
>
> This has grown by more than 100x. In fact,
> it grew in time by a factor of roughly
> 10000, or 100^2. It is this quadratic
> growth in time that I was trying to avoid
> with the growdata codes.
>
> How about growdata2 here?
>
> tic
> g=growdata2;
> for i = 1:100000
> g(rand(1,3))
> end
> g=g();
> whos g
> toc
>
> Name Size Bytes Class Attributes
> g 100000x3 2400000 double
>
> Elapsed time is 12.672389 seconds.
>
> So if you are appending only a few thousand
> elements, then either the structure or the
> cell array will work quite well. If you
> expect to append hundreds of thousands or
> more chunks, then the hybrid approach in
> the growdata tools is still more reasonable.
>
> Your mileage may vary with a newer release
> yet than R2006b. I've not gotten the latest
> release installed until I install a new OS
> version here.
>
> John

John, I can't reproduce your results on my computer (winxp32), on
R2006b or R2007a. The cell and struct both take virtually the same
amount of time, about 20sec for the 100000 case. Growdata is still
faster, of course, coming in at about 10sec.

Regards,
Eric Sampson
The MathWorks, Inc.

Re: decrease in speed due to appending a row vecto

John D'Errico — Fri, 25 May 2007 10:35:59 -0400

Eric Sampson wrote:
>
>
> John D'Errico wrote:
>>
>>
> (snip)
>> Some quick testing shows a structure and
>> a cell array to be similar in time and
>> memory here.
>>
>> tic
>> for i = 1:1000
>> a(i).X =rand(1,3);
>> end,
>> whos a
>> toc
>>
>> Name Size Bytes Class Attributes
>> a 1x1000 84064 struct
>>
>> Elapsed time is 0.023403 seconds.
>>
>> tic
>> for i = 1:1000
>> b{i} =rand(1,3);
>> end
>> whos b
>> toc
>>
>> Name Size Bytes Class Attributes
>> b 1x1000 84000 cell
>>
>> Elapsed time is 0.022017 seconds.
>>
>> The question is, do these scale well with
>> larger numbers of append operations?
>>
>> tic
>> for i = 1:100000
>> a(i).X =rand(1,3);
>> end
>> whos a
>> toc
>>
>> Name Size Bytes Class Attributes
>> a 1x100000 8400064 struct
>>
>> Elapsed time is 118.482410 seconds.
>>
>> This has grown by more than 100x. In fact,
>> it grew in time by a factor of roughly
>> 10000, or 100^2. It is this quadratic
>> growth in time that I was trying to avoid
>> with the growdata codes.
>>
>> How about growdata2 here?
>>
>> tic
>> g=growdata2;
>> for i = 1:100000
>> g(rand(1,3))
>> end
>> g=g();
>> whos g
>> toc
>>
>> Name Size Bytes Class Attributes
>> g 100000x3 2400000 double
>>
>> Elapsed time is 12.672389 seconds.
>>
>> So if you are appending only a few thousand
>> elements, then either the structure or the
>> cell array will work quite well. If you
>> expect to append hundreds of thousands or
>> more chunks, then the hybrid approach in
>> the growdata tools is still more reasonable.
>>
>> Your mileage may vary with a newer release
>> yet than R2006b. I've not gotten the latest
>> release installed until I install a new OS
>> version here.
>>
>> John
>
>
> John, I can't reproduce your results on my computer (winxp32), on
> R2006b or R2007a. The cell and struct both take virtually the same
> amount of time, about 20sec for the 100000 case. Growdata is still
> faster, of course, coming in at about 10sec.
>
> Regards,
> Eric Sampson
> The MathWorks, Inc.

The cell and the struct were both the same
for me, but it was definitely 2 minutes.

Its possible this is a CPU issue, since
I'm running on a Mac, or a memory limitation
issue (3/4 gig of ram.) I was running a web
browser on the side, so there might have
been a virtual problem.

I'll re-run the test this afternoon with
as little else competing for ram as possible
(after I also reboot.)

John

Re: decrease in speed due to appending a row vecto

John D'Errico — Fri, 25 May 2007 16:24:01 -0400

John D'Errico wrote:
>>
>> John, I can't reproduce your results on my computer (winxp32),
on
>> R2006b or R2007a. The cell and struct both take virtually the
> same
>> amount of time, about 20sec for the 100000 case. Growdata is
> still
>> faster, of course, coming in at about 10sec.
>>
>> Regards,
>> Eric Sampson
>> The MathWorks, Inc.
>
> The cell and the struct were both the same
> for me, but it was definitely 2 minutes.
>
> Its possible this is a CPU issue, since
> I'm running on a Mac, or a memory limitation
> issue (3/4 gig of ram.) I was running a web
> browser on the side, so there might have
> been a virtual problem.
>
> I'll re-run the test this afternoon with
> as little else competing for ram as possible
> (after I also reboot.)
>
> John

I reran my test. Both the struct and the
cell take 100+ seconds on my machine, not
touching the mouse during that time.

tic,for i = 1:100000,b{i} =rand(1,3);end,toc

Elapsed time is 100.453774 seconds.

I did leave a cpu monitor running, and it
showed high cpu utilization for the entire
time. There was no hard disk thrashing to
go to virtual memory.

The growdata2 call was consistently about
12 seconds.

John

Re: decrease in speed due to appending a row vecto

us — Fri, 25 May 2007 17:04:00 -0400

John D'Errico:
<SNIP growing up fast...

here are my values for <john d'errico>'s test suite
- n = 100000
- 1) struct
- 2) cell
- 3) growdata2
- wintel env: ic2d/2*2.4ghz/2gb/winvista/r2007a
- reboot, ML only

Elapsed time is 6.428093 seconds. <- struct
Elapsed time is 6.187052 seconds. <- cell
Elapsed time is 2.578906 seconds. <- growdata2

us

Re: decrease in speed due to appending a row vecto

John D'Errico — Fri, 25 May 2007 21:06:34 -0400

us wrote:
>
>
> John D'Errico:
> <SNIP growing up fast...
>
> here are my values for <john d'errico>'s test suite
> - n = 100000
> - 1) struct
> - 2) cell
> - 3) growdata2
> - wintel env: ic2d/2*2.4ghz/2gb/winvista/r2007a
> - reboot, ML only
>
> Elapsed time is 6.428093 seconds. <- struct
> Elapsed time is 6.187052 seconds. <- cell
> Elapsed time is 2.578906 seconds. <- growdata2
>
> us

Gotta get me one of them souped up computers.

Of more interest is why is there a 10:1
ratio on my system?

John

Re: decrease in speed due to appending a row vector

Loren Shure — Mon, 28 May 2007 01:23:09 -0400

In article <ef580c1.5@webcrossing.raydaftYaTP>,
woodchips@rochester.rr.com says...
> NZTideMan wrote:
> >
> >
> > On May 25, 12:35 pm, "John D'Errico"
> <woodch...@rochester.rr.com>
> > wrote:
> >> Brian wrote:
> >>
> >> > I don't know how big will the vector grow each time. The
> space
> >> > depends on the input data which will change most of the
> time
> > and I
> >> > have no way of know the final size.
> >>
> >> Take a look at my growdata or growdata2
> >> codes. They came from discussions here
> >> on c.s-s.m, and were written to allow
> >> arrays to grow in size with a minimum
> >> penalty in time. My goal was to allow
> >> arrays to grow in size to potentially
> >> millions of elements, without advance
> >> knowledge of the final count. These codes
> >> use cell arrays, each cell of which is
> >> preallocated to use moderately large
> >> but still manageable chunks of memory.
> >> Arrays with huge numbers of cells can
> >> have problems too.
> >>
> >> <http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objec...>
> >>
> >> Of course, its always better to just
> >> preallocate your large arrays whenever
> >> you can.
> >>
> >> HTh,
> >> John
> >
> > I didn't know about your growdata routines, John, so I have devised
> > the following way for accumulating results of a Monte Carlo
> > simulation:
> > I generate a structure:
> > s(irun).ymx=a;
> > where irun is the Monte Carlo run number, and the result of the
> > simulation a(1,n) is a row vector where n is the number of events
> > (drawn from a Poisson distribution). n varies from run to run.
> > I understand that when you do it this way, each array in s uses
> > contiguous memory, but they are not contiguous within the
> > structure.
> > Thus, the memory allocation problem does not arise.
> > I'd appreciate your comments on this.
>
> This is also a reassonable way to solve the
> problem.
>
> Use of a single cell for each new appended
> element is another way that is very similar
> to the structure approach that you used.
> Use of pointers here is more efficient
> than growing and re-allocating the entire
> array. I'm not sure about use of structures,
> as I thought that structures were not
> efficient users of memory as compared to
> cell arrays. This is probably changing
> with each matlab release of course. (See
> below for a comparison.)
>
> An issue is how many total elements that
> you may expect to see. In some earlier
> testing, I found that cell arrays with
> millions of cells caused some serious
> problems, however a few thousand, or even
> ten thousand or so cells was not a problem.
>
> I am starting to wonder if its time to
> reassess those growdata codes as Matlab
> evolves. In some recent tests it seemed
> that the problem I had seen with creating
> large cell arrays with millions of elements
> may no longer be as much of an issue in
> recent releases. The underlying matlab
> codes must have been improved. I'll want
> to compare how use of a structure compares
> to a cell array for memory and speed.
>
> Some quick testing shows a structure and
> a cell array to be similar in time and
> memory here.
>
> tic
> for i = 1:1000
> a(i).X =rand(1,3);
> end,
> whos a
> toc
>
> Name Size Bytes Class Attributes
> a 1x1000 84064 struct
>
> Elapsed time is 0.023403 seconds.
>
> tic
> for i = 1:1000
> b{i} =rand(1,3);
> end
> whos b
> toc
>
> Name Size Bytes Class Attributes
> b 1x1000 84000 cell
>
> Elapsed time is 0.022017 seconds.
>
> The question is, do these scale well with
> larger numbers of append operations?
>
> tic
> for i = 1:100000
> a(i).X =rand(1,3);
> end
> whos a
> toc
>
> Name Size Bytes Class Attributes
> a 1x100000 8400064 struct
>
> Elapsed time is 118.482410 seconds.
>
> This has grown by more than 100x. In fact,
> it grew in time by a factor of roughly
> 10000, or 100^2. It is this quadratic
> growth in time that I was trying to avoid
> with the growdata codes.
>
> How about growdata2 here?
>
> tic
> g=growdata2;
> for i = 1:100000
> g(rand(1,3))
> end
> g=g();
> whos g
> toc
>
> Name Size Bytes Class Attributes
> g 100000x3 2400000 double
>
> Elapsed time is 12.672389 seconds.
>
> So if you are appending only a few thousand
> elements, then either the structure or the
> cell array will work quite well. If you
> expect to append hundreds of thousands or
> more chunks, then the hybrid approach in
> the growdata tools is still more reasonable.
>
> Your mileage may vary with a newer release
> yet than R2006b. I've not gotten the latest
> release installed until I install a new OS
> version here.
>
> John
>

structs and cells should generally perform about the same. Depends on
what you are doing (accessing, appending, overwriting), where you use
them (command line, function, script), but structs have a slight to
larger advantage sometimes.

-- Loren
http://blogs.mathworks.com/loren/