Re: Holy COW! It worked (I think)
hehe, Mike if you want to go at it take a look at the performance branch
before we reverted it where I ditched all NumericRecords, LabelSST and
String based records for parallel object arrays. I just never could
finish it because at the time the other features were too much of a
moving target. At
that time I decresed memory requirements of a mostly numeric file by
about 1/3.
-Andy
Glen Stampoultzis wrote:
Nice one. You clearly know the Excel file format very well. There
are lots of places where we create too many objects. Looks like
that's one of them.
There are some tests around string handling and SST but it would be
nice to have some for the new code you've written.
Some of our existing tests against head are unfortunately still broken.
Regards,
Glen
Michael Zalewski wrote:
When I wrote this comment, I thought after I posted -- This is
ridiculous.
No sane person is going to try to rewrite BinaryTree like I
suggested. Such
a change must surely destroy every other dependant structure, which
includes
the file structure itself. Only a raving lunatic would actually spend
time
on this.
So I tried :)
Result was memory requirements decreased by about 1/3. (My file
consists of
only Strings, and the strings are short. I suspect most use cases
would not
see such an improvement in memory requirements).
However, time to load my 65,000 unique string workbook decreased by a
factor
of of almost 10 (from over 5 minutes to about 30 sec). The strange
phenomenon with the CPU going idle happened briefly for less than 3
sec, and
only one time.
Here is part of the code (to make my idea more clear)
private static final class Node
implements Map.Entry
{
private Comparable _dataKey; // instead of Comparable[] _data
private Comparable _dataData;
private Node _leftKey; // instead of Node[] _left
private Node _leftData;
private Node _rightKey; // instead of Node[] _right
private Node _rightData;
private Node _parentKey; // instead of Node[] _parent
private Node _parentData;
private boolean _blackKey; // instead of Boolean[] _black
private boolean _blackData;
private int _hashcode;
private boolean _calculated_hashcode;
/**
* Make a new cell with given key and value, and with null
* links, and black (true) colors.
*
* @param key
* @param value
*/
Node(final Comparable key, final Comparable value)
{
_dataKey = key; // much shorter ctor
_dataData = value; // does not create any arrays
}
I'll put this into Bugzilla soon to start the discussion. But I
should point
out that I have run practically no tests. Gotta find the tests first.
Are there any tests?
-Original Message-
From: Michael Zalewski [mailto:[EMAIL PROTECTED] Sent: Friday,
March 25, 2005 12:45 PM
To: 'POI Users List'
Subject: RE: HSSF cannot open files that contain many strings
My own thought is that there are just too gosh darn many objects.
(Gosh darn
many objects = gosh darn long time to process).
The SST table gets deserialized into a humongous double binary tree
structure, (org.apache.poi.util.BinaryTree) which is actually indexed by
both the index of the string and the value of the string. So this
means that
there are at least 10 objects created per String
1) The String structure (type org.apache.poi.hssf.record.UnicodeString)
2) The String value itself (contained as a field in type UnicodeString)
3) The Integer value (which indexes the String). It's an Integer object
instead of a primitive, so it can implement Comparable and be one of the
keys in the double indexed tree structure
4) The Node object (of the tree, which has a reference to both the
String
value and the Integer value)
5) One or more LabelSST records which contain an index into the tree.
If you look inside org.apache.poi.util.BinaryTree, you can see that each
node of the binary tree (there is one node for each string) contains
five
array objects in addition to the ones I listed above.
This means that my file of 65,000 unique strings will end up creating
650,000 objects to represent those strings when deserialized. I'm
probably
missing some objects in this analysis, so my guess is that my 65,000
string
spreadsheet required over a million java objects.
You can get rid of 5 of these objects with a simple refactoring of
BinaryTree -- replace each of the 5 arrays with 2 fields (replace the 5
arrays with 10 primitive fields).
-Original Message-
From: Danny Mui [mailto:[EMAIL PROTECTED] Sent: Friday, March 25,
2005 11:50 AM
To: POI Users List
Subject: Re: HSSF cannot open files that contain many strings
I'm curious about the CPU utilization issues and why it takes so gosh
darn long! Wonder what a profiler will say about loading a file as
you've described.
It shouldn't be too
Re: Holy COW! It worked (I think)
I'd like to see a patch version of this that ALSO handles rich string
format! We still don't do that properly.
The BinaryTree was actually a stop gap to de-duplicate strings both when
they weren't really duplicates (thus loosing
rich string formatting) and when they were added (and thus quite
possibly were dupes).
This looks like good work.
Michael Zalewski wrote:
When I wrote this comment, I thought after I posted -- This is ridiculous.
No sane person is going to try to rewrite BinaryTree like I suggested. Such
a change must surely destroy every other dependant structure, which includes
the file structure itself. Only a raving lunatic would actually spend time
on this.
So I tried :)
Result was memory requirements decreased by about 1/3. (My file consists of
only Strings, and the strings are short. I suspect most use cases would not
see such an improvement in memory requirements).
However, time to load my 65,000 unique string workbook decreased by a factor
of of almost 10 (from over 5 minutes to about 30 sec). The strange
phenomenon with the CPU going idle happened briefly for less than 3 sec, and
only one time.
Here is part of the code (to make my idea more clear)
private static final class Node
implements Map.Entry
{
private Comparable _dataKey; // instead of Comparable[] _data
private Comparable _dataData;
private Node _leftKey; // instead of Node[] _left
private Node _leftData;
private Node _rightKey; // instead of Node[] _right
private Node _rightData;
private Node _parentKey; // instead of Node[] _parent
private Node _parentData;
private boolean _blackKey; // instead of Boolean[] _black
private boolean _blackData;
private int _hashcode;
private boolean _calculated_hashcode;
/**
* Make a new cell with given key and value, and with null
* links, and black (true) colors.
*
* @param key
* @param value
*/
Node(final Comparable key, final Comparable value)
{
_dataKey = key; // much shorter ctor
_dataData = value; // does not create any arrays
}
I'll put this into Bugzilla soon to start the discussion. But I should point
out that I have run practically no tests. Gotta find the tests first.
Are there any tests?
-Original Message-
From: Michael Zalewski [mailto:[EMAIL PROTECTED]
Sent: Friday, March 25, 2005 12:45 PM
To: 'POI Users List'
Subject: RE: HSSF cannot open files that contain many strings
My own thought is that there are just too gosh darn many objects. (Gosh darn
many objects = gosh darn long time to process).
The SST table gets deserialized into a humongous double binary tree
structure, (org.apache.poi.util.BinaryTree) which is actually indexed by
both the index of the string and the value of the string. So this means that
there are at least 10 objects created per String
1) The String structure (type org.apache.poi.hssf.record.UnicodeString)
2) The String value itself (contained as a field in type UnicodeString)
3) The Integer value (which indexes the String). It's an Integer object
instead of a primitive, so it can implement Comparable and be one of the
keys in the double indexed tree structure
4) The Node object (of the tree, which has a reference to both the String
value and the Integer value)
5) One or more LabelSST records which contain an index into the tree.
If you look inside org.apache.poi.util.BinaryTree, you can see that each
node of the binary tree (there is one node for each string) contains five
array objects in addition to the ones I listed above.
This means that my file of 65,000 unique strings will end up creating
650,000 objects to represent those strings when deserialized. I'm probably
missing some objects in this analysis, so my guess is that my 65,000 string
spreadsheet required over a million java objects.
You can get rid of 5 of these objects with a simple refactoring of
BinaryTree -- replace each of the 5 arrays with 2 fields (replace the 5
arrays with 10 primitive fields).
-Original Message-
From: Danny Mui [mailto:[EMAIL PROTECTED]
Sent: Friday, March 25, 2005 11:50 AM
To: POI Users List
Subject: Re: HSSF cannot open files that contain many strings
I'm curious about the CPU utilization issues and why it takes so gosh
darn long! Wonder what a profiler will say about loading a file as
you've described.
It shouldn't be too difficult to adjust the way the SST's are
written/loaded to validate/invalidate this problem/fix.
Michael Zalewski wrote:
Ummm...
Yes I think I might have identified an issue with POI and a large number
of
strings. And I was looking at it partly in response to Mike's problem.
But I don't think the issue I found is the root problem. It might explain
why large files generated in POI HSSF would not open correctly in Excel.
In
fact, I
Re: Holy COW! It worked (I think)
Nice one. You clearly know the Excel file format very well. There are
lots of places where we create too many objects. Looks like that's one
of them.
There are some tests around string handling and SST but it would be nice
to have some for the new code you've written.
Some of our existing tests against head are unfortunately still broken.
Regards,
Glen
Michael Zalewski wrote:
When I wrote this comment, I thought after I posted -- This is ridiculous.
No sane person is going to try to rewrite BinaryTree like I suggested. Such
a change must surely destroy every other dependant structure, which includes
the file structure itself. Only a raving lunatic would actually spend time
on this.
So I tried :)
Result was memory requirements decreased by about 1/3. (My file consists of
only Strings, and the strings are short. I suspect most use cases would not
see such an improvement in memory requirements).
However, time to load my 65,000 unique string workbook decreased by a factor
of of almost 10 (from over 5 minutes to about 30 sec). The strange
phenomenon with the CPU going idle happened briefly for less than 3 sec, and
only one time.
Here is part of the code (to make my idea more clear)
private static final class Node
implements Map.Entry
{
private Comparable _dataKey; // instead of Comparable[] _data
private Comparable _dataData;
private Node _leftKey; // instead of Node[] _left
private Node _leftData;
private Node _rightKey; // instead of Node[] _right
private Node _rightData;
private Node _parentKey; // instead of Node[] _parent
private Node _parentData;
private boolean _blackKey; // instead of Boolean[] _black
private boolean _blackData;
private int _hashcode;
private boolean _calculated_hashcode;
/**
* Make a new cell with given key and value, and with null
* links, and black (true) colors.
*
* @param key
* @param value
*/
Node(final Comparable key, final Comparable value)
{
_dataKey = key; // much shorter ctor
_dataData = value; // does not create any arrays
}
I'll put this into Bugzilla soon to start the discussion. But I should point
out that I have run practically no tests. Gotta find the tests first.
Are there any tests?
-Original Message-
From: Michael Zalewski [mailto:[EMAIL PROTECTED]
Sent: Friday, March 25, 2005 12:45 PM
To: 'POI Users List'
Subject: RE: HSSF cannot open files that contain many strings
My own thought is that there are just too gosh darn many objects. (Gosh darn
many objects = gosh darn long time to process).
The SST table gets deserialized into a humongous double binary tree
structure, (org.apache.poi.util.BinaryTree) which is actually indexed by
both the index of the string and the value of the string. So this means that
there are at least 10 objects created per String
1) The String structure (type org.apache.poi.hssf.record.UnicodeString)
2) The String value itself (contained as a field in type UnicodeString)
3) The Integer value (which indexes the String). It's an Integer object
instead of a primitive, so it can implement Comparable and be one of the
keys in the double indexed tree structure
4) The Node object (of the tree, which has a reference to both the String
value and the Integer value)
5) One or more LabelSST records which contain an index into the tree.
If you look inside org.apache.poi.util.BinaryTree, you can see that each
node of the binary tree (there is one node for each string) contains five
array objects in addition to the ones I listed above.
This means that my file of 65,000 unique strings will end up creating
650,000 objects to represent those strings when deserialized. I'm probably
missing some objects in this analysis, so my guess is that my 65,000 string
spreadsheet required over a million java objects.
You can get rid of 5 of these objects with a simple refactoring of
BinaryTree -- replace each of the 5 arrays with 2 fields (replace the 5
arrays with 10 primitive fields).
-Original Message-
From: Danny Mui [mailto:[EMAIL PROTECTED]
Sent: Friday, March 25, 2005 11:50 AM
To: POI Users List
Subject: Re: HSSF cannot open files that contain many strings
I'm curious about the CPU utilization issues and why it takes so gosh
darn long! Wonder what a profiler will say about loading a file as
you've described.
It shouldn't be too difficult to adjust the way the SST's are
written/loaded to validate/invalidate this problem/fix.
Michael Zalewski wrote:
Ummm...
Yes I think I might have identified an issue with POI and a large number
of
strings. And I was looking at it partly in response to Mike's problem.
But I don't think the issue I found is the root problem. It might explain
why large files generated in POI HSSF would not open correctly in Excel.
In
Re: Holy COW! It worked (I think)
Michael Zalewski, you are a holy cleric.
(I hope I'm not jumping the gun too much there :)
Michael Zalewski wrote:
When I wrote this comment, I thought after I posted -- This is
ridiculous.
No sane person is going to try to rewrite BinaryTree like I suggested.
Such
a change must surely destroy every other dependant structure, which
includes
the file structure itself. Only a raving lunatic would actually spend
time
on this.
So I tried :)
Result was memory requirements decreased by about 1/3. (My file
consists of
only Strings, and the strings are short. I suspect most use cases would
not
see such an improvement in memory requirements).
However, time to load my 65,000 unique string workbook decreased by a
factor
of of almost 10 (from over 5 minutes to about 30 sec). The strange
phenomenon with the CPU going idle happened briefly for less than 3
sec, and
only one time.
Here is part of the code (to make my idea more clear)
private static final class Node
implements Map.Entry
{
private Comparable _dataKey; // instead of Comparable[] _data
private Comparable _dataData;
private Node _leftKey; // instead of Node[] _left
private Node _leftData;
private Node _rightKey; // instead of Node[] _right
private Node _rightData;
private Node _parentKey; // instead of Node[] _parent
private Node _parentData;
private boolean _blackKey; // instead of Boolean[] _black
private boolean _blackData;
private int _hashcode;
private boolean _calculated_hashcode;
/**
* Make a new cell with given key and value, and with null
* links, and black (true) colors.
*
* @param key
* @param value
*/
Node(final Comparable key, final Comparable value)
{
_dataKey = key; // much shorter ctor
_dataData = value; // does not create any arrays
}
I'll put this into Bugzilla soon to start the discussion. But I should
point
out that I have run practically no tests. Gotta find the tests first.
Are there any tests?
-Original Message-
From: Michael Zalewski [mailto:[EMAIL PROTECTED]
Sent: Friday, March 25, 2005 12:45 PM
To: 'POI Users List'
Subject: RE: HSSF cannot open files that contain many strings
My own thought is that there are just too gosh darn many objects. (Gosh
darn
many objects = gosh darn long time to process).
The SST table gets deserialized into a humongous double binary tree
structure, (org.apache.poi.util.BinaryTree) which is actually indexed by
both the index of the string and the value of the string. So this means
that
there are at least 10 objects created per String
1) The String structure (type org.apache.poi.hssf.record.UnicodeString)
2) The String value itself (contained as a field in type UnicodeString)
3) The Integer value (which indexes the String). It's an Integer object
instead of a primitive, so it can implement Comparable and be one of the
keys in the double indexed tree structure
4) The Node object (of the tree, which has a reference to both the
String
value and the Integer value)
5) One or more LabelSST records which contain an index into the tree.
If you look inside org.apache.poi.util.BinaryTree, you can see that each
node of the binary tree (there is one node for each string) contains
five
array objects in addition to the ones I listed above.
This means that my file of 65,000 unique strings will end up creating
650,000 objects to represent those strings when deserialized. I'm
probably
missing some objects in this analysis, so my guess is that my 65,000
string
spreadsheet required over a million java objects.
You can get rid of 5 of these objects with a simple refactoring of
BinaryTree -- replace each of the 5 arrays with 2 fields (replace the 5
arrays with 10 primitive fields).
-Original Message-
From: Danny Mui [mailto:[EMAIL PROTECTED]
Sent: Friday, March 25, 2005 11:50 AM
To: POI Users List
Subject: Re: HSSF cannot open files that contain many strings
I'm curious about the CPU utilization issues and why it takes so gosh
darn long! Wonder what a profiler will say about loading a file as
you've described.
It shouldn't be too difficult to adjust the way the SST's are
written/loaded to validate/invalidate this problem/fix.
Michael Zalewski wrote:
Ummm...
Yes I think I might have identified an issue with POI and a large
number
of
strings. And I was looking at it partly in response to Mike's problem.
But I don't think the issue I found is the root problem. It might
explain
why large files generated in POI HSSF would not open correctly in
Excel.
In
fact, I couldn't find any problem with the way POI handles things. At
this
point, I would say that what I have identified is just a difference in
the
way Excel writes a file with more than 1024 strings, and
