Memory usage per top 10x usage per heapy

[MrsEntity]

Hi all,

I'm working on some code that parses a 500kb, 2M line file line by line andsaves, per line, some derived strings into various data structures. I thusexpect that memory use should monotonically increase. Currently, the program is taking up so much memory - even on 1/2 sized files - that on 2GB machine I'm thrashing swap. What's strange is that heapy (http://guppy-pe.sourceforge.net/) is showing that the code uses about 10x less memory than reported by top, and the heapy data seems consistent with what I was expecting based on the objects the code stores. I tried using memory_profiler (http://pypi.python.org/pypi/memory_profiler) but it didn't really provide any illuminating information. The code does create and discard a number of objects per line of the file, but they should not be stored anywhere, and heapy seems to confirm that. So, my questions are:

1) For those of you kind enough to help me figure out what's going on, whatadditional data would you like? I didn't want swamp everyone with the codeand heapy/memory_profiler output but I can do so if it's valuable.
2) How can I diagnose (and hopefully fix) what's causing the massive memoryusage when it appears, from heapy, that the code is performing reasonably?

Specs: Ubuntu 12.04 in Virtualbox on Win7/64, Python 2.7/64

Thanks very much.

 

[Tim Chase]

I'm working on some code that parses a 500kb, 2M line file line
by line and saves, per line, some derived strings into various
data structures. I thus expect that memory use should
monotonically increase. Currently, the program is taking up so
much memory - even on 1/2 sized files - that on 2GB machine I'm
thrashing swap.

It might help to know what comprises the "into various data
structures". I do a lot of ETL work on far larger files,
with similar machine specs, and rarely touch swap.

2) How can I diagnose (and hopefully fix) what's causing the
massive memory usage when it appears, from heapy, that the code
is performing reasonably?

I seem to recall that Python holds on to memory that the VM
releases, but that it *should* reuse it later. So you'd get
the symptom of the memory-usage always increasing, never
decreasing.

Things that occur to me:

- check how you're reading the data: are you iterating over
the lines a row at a time, or are you using
.read()/.readlines() to pull in the whole file and then
operate on that?

- check how you're storing them: are you holding onto more
than you think you are? Would it hurt to switch from a
dict to store your data (I'm assuming here) to using the
anydbm module to temporarily persist the large quantity of
data out to disk in order to keep memory usage lower?

Without actual code, it's hard to do a more detailed
analysis.

-tkc

 

[Junkshops]

Hi Tim, thanks for the response.

- check how you're reading the data: are you iterating over
the lines a row at a time, or are you using
.read()/.readlines() to pull in the whole file and then
operate on that?

I'm using enumerate() on an iterable input (which in this case is the
filehandle).

- check how you're storing them: are you holding onto more
than you think you are?

I've used ipython to look through my data structures (without going into
ungainly detail, 2 dicts with X numbers of key/value pairs, where X =
number of lines in the file), and everything seems to be working
correctly. Like I say, heapy output looks reasonable - I don't see
anything surprising there. In one dict I'm storing a id string (the
first token in each line of the file) with values as (again, without
going into massive detail) the md5 of the contents of the line. The
second dict has the md5 as the key and an object with __slots__ set that
stores the line number of the file and the type of object that line
represents.

Would it hurt to switch from a
dict to store your data (I'm assuming here) to using the
anydbm module to temporarily persist the large quantity of
data out to disk in order to keep memory usage lower?

That's the thing though - according to heapy, the memory usage *is* low
and is more or less what I expect. What I don't understand is why top is
reporting such vastly different memory usage. If a memory profiler is
saying everything's ok, it makes it very difficult to figure out what's
causing the problem. Based on heapy, a db based solution would be
serious overkill.

-MrsE

 

[Dave Angel]

Hi all,

I'm working on some code that parses a 500kb, 2M line file

Just curious; which is it, two million lines, or half a million bytes?

line by line and saves, per line, some derived strings into various data structures. I thus expect that memory use should monotonically increase. Currently, the program is taking up so much memory - even on 1/2 sized files - that on 2GB machine

which machine is 2gb, the Windows machine, or the VM? You could get
thrashing at either level.

I'm thrashing swap. What's strange is that heapy (http://guppy-pe.sourceforge.net/) is showing that the code uses about 10x less memory than reported by top, and the heapy data seems consistent with what I was expecting based on the objects the code stores. I tried using memory_profiler (http://pypi.python.org/pypi/memory_profiler) but it didn't really provide any illuminating information. The code does create and discard a number of objects per line of the file, but they should not be stored anywhere, and heapy seems to confirm that. So, my questions are:

1) For those of you kind enough to help me figure out what's going on, what additional data would you like? I didn't want swamp everyone with the code and heapy/memory_profiler output but I can do so if it's valuable.
2) How can I diagnose (and hopefully fix) what's causing the massive memory usage when it appears, from heapy, that the code is performing reasonably?

Specs: Ubuntu 12.04 in Virtualbox on Win7/64, Python 2.7/64

Thanks very much.

Tim raised most of my concerns, but I would point out that just because
you free up the memory from the Python doesn't mean it gets released
back to the system. The C runtime manages its own heap, and is pretty
persistent about hanging onto memory once obtained. It's not normally a
problem, since most small blocks are reused. But it can get
fragmented. And i have no idea how well Virtual Box maps the Linux
memory map into the Windows one.

 

[Junkshops]

Just curious; which is it, two million lines, or half a million bytes?
I have, in fact, this very afternoon, invented a means of writing a
carriage return character using only 2 bits of information. I am
prepared to sell licenses to this revolutionary technology for the low
price of $29.95 plus tax.

Sorry, that should've been a 500Mb, 2M line file.

which machine is 2gb, the Windows machine, or the VM?

VM. Winders is 4gb.

...but I would point out that just because
you free up the memory from the Python doesn't mean it gets released
back to the system. The C runtime manages its own heap, and is pretty
persistent about hanging onto memory once obtained. It's not normally a
problem, since most small blocks are reused. But it can get
fragmented. And i have no idea how well Virtual Box maps the Linux
memory map into the Windows one.

Right, I understand that - but what's confusing me is that, given the
memory use is (I assume) monotonically increasing, the code should never
use more than what's reported by heapy once all the data is loaded into
memory, given that memory released by the code to the Python runtime is
reused. To the best of my ability to tell I'm not storing anything I
shouldn't, so the only thing I can think of is that all the object
creation and destruction, for some reason, it preventing reuse of
memory. I'm at a bit of a loss regarding what to try next.

Cheers, MrsE

 

[Dennis Lee Bieber]

Hi all,

I'm working on some code that parses a 500kb, 2M line file line by line and saves, per line, some derived strings

Pardon? A 2million line file will contain, at the minimum 2million
line-end characters. That four times 500kB just in the line-ends,
ignoring any data.

 

[Tim Chase]

Pardon? A 2million line file will contain, at the minimum 2million
line-end characters. That four times 500kB just in the line-ends,
ignoring any data.

As corrected later in the thread, MrsEntity writes

"""
I have, in fact, this very afternoon, invented a means of writing a
carriage return character using only 2 bits of information. I am
prepared to sell licenses to this revolutionary technology for the
low price of $29.95 plus tax.

Sorry, that should've been a 500Mb, 2M line file.
"""

If only other unnamed persons on the list were so gracious rather
than turning the flame-dial to 11.

I hope that when people come to the list, *this* is what they see,
laugh, and want to participate.

Although, MrsEntity could be zombie David A. Huffman, whose encoding
scheme actually *can* store 2M lines in 500kb

-tkc

 

[Dave Angel]

Sorry, that should've been a 500Mb, 2M line file.

VM. Winders is 4gb.

Right, I understand that - but what's confusing me is that, given the
memory use is (I assume) monotonically increasing, the code should never
use more than what's reported by heapy once all the data is loaded into
memory, given that memory released by the code to the Python runtime is
reused. To the best of my ability to tell I'm not storing anything I
shouldn't, so the only thing I can think of is that all the object
creation and destruction, for some reason, it preventing reuse of
memory. I'm at a bit of a loss regarding what to try next.

I'm not familiar with heapy, but perhaps it's missing something there.
I'm a bit surprised you aren't beyond the 2gb limit, just with the
structures you describe for the file. You do realize that each object
has quite a few bytes of overhead, so it's not surprising to use several
times the size of a file, to store the file in an organized way. I also
wonder if heapy has been written to take into account the larger size of
pointers in a 64bit build.

Perhaps one way to save space would be to use a long to store those md5
values. You'd have to measure it, but I suspect it'd help (at the cost
of lots of extra hexlify-type calls). Another thing is to make sure
that the md5 object used in your two maps is the same object, and not
just one with the same value.

 

[Mark Lawrence]

On 25/09/2012 11:51, Tim Chase wrote:
[snip]

If only other unnamed persons on the list were so gracious rather
than turning the flame-dial to 11.

Oh heck what have I said this time?

 

[Tim Chase]

Oh heck what have I said this time?

You'd *like* to take credit? ;-)

Nah, not you or any of the regulars here. The comment was regarding
the flame-fest that's been running in some parallel threads over the
last ~12hr or so. Mostly instigated by one person with a
particularly quick trigger, vitriolic tongue, and a disregard for
pythonic code.

-tkc

 

[Mark Lawrence]

You'd *like* to take credit? ;-)

Nah, not you or any of the regulars here. The comment was regarding
the flame-fest that's been running in some parallel threads over the
last ~12hr or so. Mostly instigated by one person with a
particularly quick trigger, vitriolic tongue, and a disregard for
pythonic code.

-tkc

Well thank goodness for that. Of course the person to whom you've
alluded has been defended over on the tutor mailing list, seriously, and
as I've said elsewhere after referring to my family as pigs!!!

 

[alex23]

Mostly instigated by one person with a
particularly quick trigger, vitriolic tongue, and a disregard for
pythonic code.

I'm sorry. I'll get me coat.

 

[Mark Lawrence]

I'm sorry. I'll get me coat.

Oi, back of the queue if you don't mind

 

[Steven D'Aprano]

Well thank goodness for that. Of course the person to whom you've
alluded has been defended over on the tutor mailing list, seriously, and
as I've said elsewhere after referring to my family as pigs!!!

Since pigs are at least as intelligent as dogs, and in their natural
state nowhere near as filthy as the stereotype of the pig in a sty, that
isn't as big an insult as it was intended.

 

[Dave Angel]

On 09/25/2012 01:39 PM, Junkshops wrote:

Procedural point: I know you're trying to conform to the standard that
this mailing list uses, but you're off a little, and it's distracting.
It's also probably more work for you, and certainly for us.

You need an attribution in front of the quoted portions. This next
section is by me, but you don't say so. That's because you copy/pasted
it from elsewhere in the reply, and didn't copy the "... Dave Angel
wrote" part.

Much easier is to take the reply, and remove the parts you're not going
to respond to, putting your own comments in between the parts that are
left (as you're doing). And generally, there's no need for anything
after your last remark, so you just delete up to your signature, if any.

I did some back of the envelope calcs which more or less agreed with
heapy. The code stores 1 string, which is, on average, about 50 chars or
so, and one MD5 hex string per line of code. There's about 40 bytes or
so of overhead per string per sys.getsizeof(). I'm also storing an int
(24b) and a <10 char string in an object with __slots__ set. Each
object, per heapy (this is one area where I might be underestimating
things) takes 64 bytes plus instance variable storage, so per line:

50 + 32 + 10 + 3 * 40 + 24 + 64 = 300 bytes per line * 2M lines = ~600MB
plus some memory for the dicts, which is about what heapy is reporting
(note I'm currently not actually running all 2M lines, I'm just running
subsets for my tests).

Is there something I'm missing? Here's the heapy output after loading
~300k lines:

Partition of a set of 1199849 objects. Total size = 89965376 bytes.
Index Count % Size % Cumulative % Kind
0 599999 50 38399920 43 38399920 43 str
1 5 0 25167224 28 63567144 71 dict
2 299998 25 19199872 21 82767016 92 0xa13330
3 299836 25 7196064 8 89963080 100 int
4 4 0 1152 0 89964232 100
collections.defaultdict

Note that 3 of the dicts are empty. I assumet 0xa13330 is the
address of the object. I'd actually expect to see 900k strings, but the
<10 char string is always the same in this case so perhaps the runtime
is using the same object...?

CPython currently interns short strings that conform to variable name
rules. You can't count on that behavior (and i probably don't have it
quite right anyway), but it's probably what you're seeing.

At this point, top reports python as using
1.1g of virt and 1.0g of res.

That I don't know, but that would only explain, at most, a 2x increase
in memory over the heapy report, wouldn't it? Not the ~10x I'm seeing.

That's certainly the way the code is written, and heapy seems to confirm
that the strings aren't duplicated in memory.

Thanks for sticking with me on this,

You're certainly welcome. I suspect that heapy has some limitation in
its reporting, and that's what the discrepancy. Oscar points out that
you have a bunch of exception objects, which certainly looks suspicious.
If you're somehow storing one of these per line, and heapy isn't
reporting them, that could be a large discrepancy.

He also points out that you have a couple of lambda functions stored in
one of your dictionary. A lambda function can be an expensive
proposition if you are building millions of them. So can nested
functions with non-local variable references, in case you have any of those.

Oscar also reminds you of what I suggested for the md5 fields. Stored
as ints instead of hex strings could save a good bit. Just remember to
use the same one for both dicts, as you've been doing with the strings.


Other than that, I'm stumped.

 

[Junkshops]

I suspect that heapy has some limitation in its reporting, and that's
what the discrepancy.

That would be my first suspicion as well - except that heapy's results
agree so well with what I expect, and I can't think of any reason I'd be
using 10x more memory. If heapy is wrong, then I need to try and figure
out what's using up all that memory some other way... but I don't know
what that way might be.

... can be an expensive proposition if you are building millions of
them. So can nested functions with non-local variable references, in
case you have any of those.

Not as far as I know.

Cheers, MrsEntity

 

[Ian Kelly]

Also I think lambda functions might be able to keep the frame alive. Are
they by any chance being created in a function that is called in a loop?

I'm pretty sure they don't. Closures don't keep a reference to the
calling frame, only to the appropriate cellvars.

Also note that whether a function is a closure has nothing to do with
whether it was defined by a lambda or a def statement. In fact,
there's no difference between functions created by one vs. the other,
except that one has an interesting __name__ and the other does not.

 

[Tim Chase]

I don't know whether it would be better or worse but it might be
worth seeing what happens if you replace the FileContext objects
with tuples.

If tuples provide a savings but you find them opaque, you might also
consider named-tuples for clarity.

-tkc

 

[Tim Chase]

Do they have the same memory usage?

Since tuples don't have a per-instance __dict__, I'd expect them to be a
lot lighter. I'm not sure if I'm interpreting the results below properly
but they seem to suggest that a namedtuple can have a memory consumption
several times larger than an ordinary tuple.

I think the "how much memory is $METHOD using" topic of the thread
is the root of the problem. From my testing of your question:

import collections, sys
A = collections.namedtuple('A', ['x', 'y'])
nt = A(1,3)
t = (1,3)
sys.getsizeof(nt) 72
sys.getsizeof(t) 72
nt_s = set(dir(nt))
t_s = set(dir(t))
t_s ^ nt_s

set(['__module__', '_make', '_asdict', '_replace', '_fields',
'__slots__', 'y', 'x'])set([])

So a named-tuple has 6+n (where "n" is the number of fields) extra
attributes, but it seems that namedtuples & tuples seem to occupy
the same amount of space (72).

Additionally, pulling up a second console and issuing

ps v | grep [p]ython

shows the memory usage of the process as I perform these, and after
them, and they both show the same usage (actual test was

1) pull up a fresh python
2) import sys, collections; A = collections.namedtuple('A',['x','y'])
3) check memory usage in other window
4a) x = (1,2)
4b) x = A(1,2)
5) check memory usage again in other window
6) quit python

performing 4a on one run, and 4b on the second run.

Both showed identical memory usage as well (Debian Linux (Stable),
stock Python 2.6.6) at the system level.

I don't know if that little testing is actually worth anything, but
at least it's another data-point as we muddle towards helping
MrsEntity/junkshops.

-tkc

 

[bryanjugglercryptographer]

Based on heapy, a db based solution would be serious overkill.

I've embraced overkill and my life is better for it. Don't confuse overkillwith cost. Overkill is your friend.

The facts of the case: You need to save some derived strings for each of 2Minput lines. Even half the input runs over the 2GB RAM in your (virtual) machine. You're using Ubuntu 12.04 in Virtualbox on Win7/64, Python 2.7/64.

That screams "sqlite3". It's overkill, in a good way. It's already there for the importing.

Other approaches? You could try to keep everything in RAM, but use less. Tim Chase pointed out the memory-efficiency of named tuples. You could save some more by switching to Win7/32, Python 2.7/32; VirtualBox makes trying such alternatives quick and easy.

Or you could add memory. Compared to good old 32-bit, 64-bit operation consumes significantly more memory and supports vastly more memory. There's a bit of a mis-match in a 64-bit system with just 2GB of RAM. I know, sounds weird, "just" two billion bytes of RAM. I'll rephrase: just ten dollars worth of RAM. Less if you buy it where I do.

I don't know why the memory profiling tools are misleading you. I can thinkof plausible explanations, but they'd just be guesses. There's nothing allthat surprising in running out of RAM, given what you've explained. A couple K per line is easy to burn.

-Bryan