[erlang-questions] Strings and Text Processing

J K jmakarlsson@REDACTED
Wed Jan 9 23:44:34 CET 2013 




________________________________
 From: Joe Armstrong <erlang@REDACTED>
To: J K <jmakarlsson@REDACTED> 
Cc: Dmitry Kolesnikov <dmkolesnikov@REDACTED>; Steve Davis <steven.charles.davis@REDACTED>; Erlang Questions <erlang-questions@REDACTED> 
Sent: Thursday, January 3, 2013 10:34 AM
Subject: Re: [erlang-questions] Strings and Text Processing
 




On Thu, Jan 3, 2013 at 12:25 AM, J K <jmakarlsson@REDACTED>wrote:

Hi,
>
>
>I'm not sure I understand your 20.000 files example.

This was just to give an estimate of the total memory size
 
Are you suggesting that the user should limit the number of erlang processes to the number of cores or are you suggesting that the VM compresses the erlang process data when not running? 

You could experiment with limiting the number of active processes and
compressing data when it's not being used (you'd have to do this yourself as
part of the application).

To first approximation one process per parallel activity is a good rule of thumb (and you let the Erlang scheduler figure out who is to run where) - the alternative is
that you limit the number of parallel processes and decide when and where they execute - you are basically saying "because I know a lot about the specific details of my application I can do a better job of  process management than the Erlang VM" - this is pretty tricky.

---
Yes, it would probably take some time to do that. I would rather let the VM limit the number of processes to a number set by the user. 





That would be really nice if it can be done with only a small performance penalty, say 10-20%.

>
>In my case I start 50.000 to 100.000 processes , one per file (it's an (map reduce like) application to do feature extraction for some machine learning algorithms) . 

How big is each file? how much processing is needed per-file to do feature extraction?

---
I would guess most of them are about 200-300KB, but probably vary between 0.1-1MB.

This is the parsing part of one of the feature extraction algorithms:

Definitions.

T = [a-zA-Z0-9!"#¤&%/()=?¡@£$€¥{\[\]}\\,.\-;:_<>\|^~\*'+´`\s\r\n\t\f\e\b\v\d©�]

Rules.

{T}{T}: {token,TokenChars,pushback(TokenChars)}.
{T}: skip_token.

Erlang code.

pushback([_,X]) -> [X].


It's the character bigram algorithm:
http://en.wikipedia.org/wiki/Bigram


A special case of n-gram:
http://en.wikipedia.org/wiki/N-gram


However I just realized I can do this:
{T}{T}: {token,list_to_binary(TokenChars),pushback(TokenChars)}.


 
One erlang node uses about 7GB of memory. I can probably tune it a bit (a lot?) more by using binaries but it would be nice to have an option to compress process data when not running, for people that are lazy/not an erlang expert/does not have that much time (my case)/or just as an indication of how much memory that could be saved by using binaries.
>
>

look up the manual entry for hibernate/3

http://erlang.org/doc/man/erlang.html#hibernate-3 

The hibernate BIF minimise the size of a process before putting it to sleep. It won't
compress any process data - but it does trim the stack and heap before suspending a process. If you have a large number of processes that sleep for
long times this might be a good idea. If your processes sleep for short times and wake up at random then it probably won't help.

Performance tuning is a black art. Unfortunately it's *very* system dependent. If you change the number of cores, or operating system, or amount of memory you have to start again.

If in doubt measure !

---
I actually do explicit garbage collection already. I made some measurement of peak memory usage using different combinations of GC, binary, list and fullsweep_after. I might try hibernate later but I'm not sure if it would make any big difference, after all I have 7GB available on each host :) And I can easily add another host if needed.

I don't have the files available right now so I copied some include files totaling about 27MB (as reported by du -m).


Peak mem usage (using this: http://www.linuxquestions.org/questions/linux-software-2/how-to-measure-a-programs-peak-memory-consumption-336376/ ):

Using: {T}{T}: {token,TokenChars,pushback(TokenChars)}.
- No explicit GC, no fullsweep_after: 
    Min 1318, Max1578, Individual Measurements: 1346, 1318, 1342, 1321, 1587, 1343 MB
- No explicit GC, fullsweep_after 0:

    Min 1359, Max 1414, Meas: 1361, 1363, 1359, 1414, 1365 MB
- Explicit GC, no fullsweep_after:
    Min 1009, Max 1093, Meas: 1010, 1015, 1009, 1093, 1015, 1022 MB
- Explicit GC, fullsweep_after 0
    Min 840, Max 1193, Meas: 864, 844, 847, 840, 868, 1193, 860, 859 MB, 

Using: {T}{T}: {token,list_to_binary(TokenChars),pushback(TokenChars)}.
- No explicit GC, no fullsweep_after:
    1215-1232. 1232, 1215, 1225, 1222 MB
- No explicit GC, fullsweep_after 0
    1191-1304. 1205, 1239, 1304, 1267, 1238, 1191, 1220 MB
- Explicit GC, no fullsweep_after:
    750-955. 890, 813, 885, 822, 874, 955, 846, 890, 750 MB
- Explicit GC, fullsweep_after 0
    813-822, 822, 820, 813, 815, 821, 817 MB

In summary (for my case): using binaries is a good idea but is not critical if using explicit GC and fullsweep_after 0. Also worth mentioning is that I create an ets table for each file to save the parsed result  (working with lists is too slow). This is when the memory usage peaks. 

It would also be good if the fullsweep_after option was mentioned in the GC chapter:
http://erlang.org/doc/man/erlang.html#garbage_collect-0


Johan




/Joe
 
JK
>
>
>
>
>
>________________________________
> From: Joe Armstrong <erlang@REDACTED>
>To: Dmitry Kolesnikov <dmkolesnikov@REDACTED> 
>Cc: Steve Davis <steven.charles.davis@REDACTED>; Erlang Questions <erlang-questions@REDACTED> 
>Sent: Saturday, December 29, 2012 4:30 PM
>
>Subject: Re: [erlang-questions] Strings and Text Processing
> 
>
>
>
>
>
>On Sat, Dec 29, 2012 at 3:20 PM, Dmitry Kolesnikov <dmkolesnikov@REDACTED> wrote:
>
>Hello Steve,
>>
>>You have raised a good point here.
>>One more reason for binary is memory consumption and IPC overhead.
>>
>
>
>The point about memory consumption is raised *many* times - on a modern
>machine this is not a problem.
>
>
>Example: I am working on a text file of 84KB - in a 32 bit Erlang we use
>8 bytes/character - so I use 0.6 MB - I have 4GB memory - so I use 0.015% of
>memory - ie no problem.
>
>
>My strategy is to keep large strings as binaries when I'm not working on them,
>turn them into lists in order to work on them, and turn them back into binaries
>when I'm done. Just because a string starts off in a binary does not mean
>that it has to stay as a binary as you work on it.
>
>
>
>
>Imagine I have a lot of text files, say each of 50KB, I can store 20 per/MB or
>20,000 files per GB. Assume I have a quad core. I can only work on four things
>at the same time - so having (say) 20,000 files (at 50K) and work on four of them
>(unpacked) at a time is another 1.6 Meg.
>
>
>Gigabyte memories mean (among other things) what saving the odd byte here are there is hardly relevant.
>
>
> 
>On another hand list allows to represent a code point per element.
>>
>
>
>yes - the convenience of having one character per list element far outweighs
>the space saving of storing strings in binaries
>
>
> 
>iolists are also very handy to dynamically compose a complex strings.
>>
>>I am afraid that this is an application specific questions… However, I tend to use binary for strings...
>>
>>
>
>
>My strings change form depending on what I'm doing. Sometimes they are
>binaries, sometimes lists, sometimes trees, ...
>
>
>Cheers
>
>
>/Joe
>
>
> 
>- Dmitry
>>
>>
>>
>>On Dec 29, 2012, at 4:08 PM, Steve Davis <steven.charles.davis@REDACTED> wrote:
>>
>>> Disclaimer :-) All the below is prefixed by a big IMHO
>>>
>>> Erlang has been correctly criticized for the difficulty of handling "strings".
>>>
>>> There are two reasons for this (fundamental decisions that were taken way-back-when):
>>> 1) "strings" are "just lists of integers"
>>> 2) "strings" are by default latin-1 representations
>>>
>>> This introduces major inconveniences, some of which are not resolvable:
>>> When faced with any list during pattern matching, it is not at all easy to determine whether that list is a "string".
>>> Further, since strings are "only" a subset of the set of lists of integers, it can be impossible to determine programmatically whether the list is a list of integers or is meant to represent a string. Determining whether a particular list even qualifies as a string in a program requires non-trivial processing of the entire list.
>>>
>>> It's rather unfortunate that Erlang has earned this reputation, since the truth is that Erlang is truly excellent at text processing. However, to benefit from this excellence, you need to do two things:
>>> 1) Represent and process text as binaries.
>>> 2) Assume that the text binary is UTF-8 encoded, unless otherwise stated (meaning, e.g. #text{encoding = cstring, value = <<116,101,120,116,0>>}).
>>>
>>> Suddenly, thanks to binary syntax and pattern matching, processing text in your programs becomes deterministic and easy. (Note that part of the reason for this is that binaries are "expected" to be opaque, whereas general list processing is fundamental to writing any program in Erlang).
>>>
>>> There's a couple of minor drawbacks, both of which are the result of the initial decisions about "strings":
>>> 1) The code is littered with additional angle brackets <<"string">> (annoying, but definitely worth the inconvenience)
>>> 2) The standard Erlang/OTP library functions require textual arguments as lists (requiring overuse of binary_to_list)
>>>
>>> And there are further benefits:
>>> 1) Parsing/transcoding different charset encodings is far more straightforward
>>> 2) Internationalization/localization is far more straightfoward
>>>
>>> I wonder if, had the current binary pattern matching/comprehensions been available "way-back-when", whether the decision about "string" representation in Erlang may have been different. (i.e. <<116,101,120,116>> = "text").
>>>
>>> Finally, here's my two questions:
>>> 1) Is there any benefit at all to the "list representation" of strings above binary text?
>>> 2) If not, I wonder if there's any way to change our minds about "strings" as we enter 2013?
>>>
>>> regs,
>>> /s
>>>
>>> _______________________________________________
>>> erlang-questions mailing list
>>> erlang-questions@REDACTED
>>> http://erlang.org/mailman/listinfo/erlang-questions
>>
>>_______________________________________________
>>erlang-questions mailing list
>>erlang-questions@REDACTED
>>http://erlang.org/mailman/listinfo/erlang-questions
>>
>
>
>_______________________________________________
>erlang-questions mailing list
>erlang-questions@REDACTED
>http://erlang.org/mailman/listinfo/erlang-questions
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20130109/7e6c300d/attachment.htm>

More information about the erlang-questions mailing list