Thanks Wade. We will probably be doing something along those lines. We need
to keep this fast and simple also.
I looked back at the code and found that on all platforms other than
Windows, we use opendir()/closedir()/readdir() to handle directories.
Unicon's Open() doesn't need temp files on those platforms. On Windows
however, Windows API is used to read the content of a directory into a temp
file before passing it up at the language level. I suspect that
opendir()/closedir()/readdir() were not available in toolchains like MinGW
at the time. The good news is that these functions seem to be available
now. I just tried them and they all worked nicely. Temp files are not
needed anymore in this case.
Avoiding the temp files and all of the intermediate work involved with them
sped up my program by more than 5 times! I went from 125 seconds to 21
seconds, >5X speedup.
Cheers,
Jafar
On Fri, Jan 23, 2015 at 3:58 PM, Wade <[email protected]> wrote:
> Okay, thanks. It's been so long since I've programmed against the
> Windows API. Definitely sounds like a bug in there. Pity we have to take
> that into account.
>
> Could I suggest a prefix based on some part of the current timestamp as
> well as a random factor. That would do a fairly simple but scalable
> namespace partitioning. I don't have any facility to test it, sorry, so I'm
> just tossing ideas out here.
>
> Wade.
>
>
> On Sat, 24 Jan 2015 04:01:03 +1100, Jafar Al-Gharaibeh <[email protected]>
> wrote:
>
> Wade,
>
> _tempnam(dir, prefix) is provided by Windows, we just use it and it
> turned not to be smart at all - at least on my Windows 7 machine. However,
> our code that uses it could be made smarter to always use randomized prefix
> every time - that is one approach.
>
> Thanks,
> Jafar
>
>
> On Fri, Jan 23, 2015 at 3:44 AM, Wade <[email protected]> wrote:
>
>> Sounds like the _tempnam() function could be a lot smarter in creating
>> temporary filenames. Is that our function or is it provided by Windows?
>>
>> Wade.
>>
>>
>> On Fri, 23 Jan 2015 20:13:17 +1100, Sergey Logichev <[email protected]>
>> wrote:
>>
>> Jafar,
>>
>> I am very appreciate for your investigations! Actually, my Windows %TMP%
>> folder included ~135000 temporary files, so when I cleaned it my run time
>> decreased from ~40 secs to ~20. And the very first open() was instant, then
>> its time increased as number of temporary files increases too. My proposal
>> to purge all temporary files after program finishes or instead use virtual
>> storage at RAM, as on every searched subdirectory is created single
>> temporary file. After very short time TMP folder will contain a myriad of
>> such files.
>>
>> Nevertheless I confirm that number of threads practically do not
>> influence on execution time. Probably, it's the problem of "lazy cleanup",
>> as you mentioned. Hope you could find solution. Compared with Linux -
>> Windows is quite a bag of different bugs! Which runs from every holes :-)
>>
>> Thank you,
>> Sergey
>>
>> 23.01.2015, 10:19, "Jafar Al-Gharaibeh" <[email protected]>:
>>
>> Sergey,
>>
>> Thanks for the report. I had in mind to look at why we don't get much
>> speed up with more threads. I did look and found that the main thread was
>> grabbing most "new thread tokens" and not recycling them fast enough. I
>> have to tweak my algorithm to allow quick cleanup and reuse of threads. I
>> will do that when I get a chance.
>>
>> Now the second issue - and you've gotta love this!-, I was able to
>> confirm the slow open(). With the help of gdb and after spending a couple
>> of hours digging into the C code and the Windows API calls, I found that
>> the problem is in a call to _tempnam() to create a temporary file name. The
>> call was taking so long to finish. It creates the tmp file under your
>> system TMP folder (%TMP% on Windows). I looked in that folder and found
>> that it has more than half a million files (~2.7GB)!
>>
>> It turned out that every time my program runs, Windows was looping
>> through that huge pile of tmp files to find a name that doesn't exist so
>> that it can give it to the program. Of course I think most of those tmp
>> files were generated by my program during previous runs the last couple of
>> days.
>>
>> As a bonus, I discovered a memory leak in the process of tracking the
>> open() problem. I committed a fix for that leak. This is only affecting
>> Windows.
>>
>> Short term solution: flush your TMP folder.
>> Long term: we will into ways to improve our tmp file strategy to overcome
>> the shortcoming of Windows API. This will come in a later date! :)
>>
>> Cheers,
>> Jafar
>>
>>
>> On Thu, Jan 22, 2015 at 4:43 AM, Sergey Logichev <[email protected]>
>> wrote:
>>
>> Jafar,
>>
>> You've provided very interesting version of walk directory algorithm.
>> Communication with active threads' is a great thing!
>> I have checked your program under Windows 7. I was confused the fact that
>> execution time is negligibly depended on number of concurrent threads. I
>> dug into and discovered that the first operation open(s) takes near ALL
>> execution time! 95% at least. Check it yourself when you slightly edit
>> getdirs():
>> ...
>> if ( stat(s).mode ? ="d" ) & ( tm := &time, d := open(s) ) then {
>> if n=1 then write(s," : ",&time-tm)
>> ...
>>
>>
>> So, if first open() is so long then all other enhancements have no sense.
>> Please clarify if I am wrong.
>>
>> Best regards,
>> Sergey
>>
>> 22.01.2015, 00:58, "Jafar Al-Gharaibeh" <[email protected]>:
>>
>> Here is a slightly tweaked/reformatted version. It now by default
>> auto-detect the number of available cores in the machine and launch twice
>> as many threads.
>>
>> --Jafar
>>
>> On Wed, Jan 21, 2015 at 12:17 PM, Jafar Al-Gharaibeh <[email protected]>
>> wrote:
>>
>> David,
>>
>> I added a threaded solution @
>> http://rosettacode.org/wiki/Walk_a_directory/Recursively#Icon_and_Unicon
>> Please review/edit as you see fit. (The source file is attached).
>> Combining recursion with thread might not be the best solution for this
>> problem. If I were to put this in real use I'd go with an iterative
>> approach using master/workers model. Anyway, this is a excellent
>> demonstration on how to use threads!. The key features are:
>>
>> 1- How to create threads, limit their numbers, self-load balanced (new
>> threads are spawned at the time/place where needed. One they are done,
>> they vanish allowing new threads to pop up in new places in the directory
>> structure)
>> 2- pass data and collect results to/from the threads using the new
>> language features.
>>
>>
>> Here is some sample output from my desktop machine (quad-core with
>> mechanical HDD. I will try another machine with an SSD and see if more
>> threads scale better).
>>
>> the first argument to the program is the target directory. The second is
>> the maximum number of concurrent threads to use at any given moment. (soft
>> limit! my counters are "unmutexed", so the actual number might deviate).
>> Note that this is different from the actual number of threads used during
>> the run which is reported at the end. The program can create/destroy
>> threads as needed, but cannot use more than "max" # of threads at any
>> given moment, and again "max" is "soft". :)
>>
>> Cheers,
>> Jafar
>>
>> c:\proj>tdir c:\ 1
>> 39708 directories in 99867 ms using 1 threads
>>
>> c:\proj>tdir c:\ 4
>> 39708 directories in 62222 ms using 4 threads
>>
>> c:\proj>tdir c:\ 4
>> 39708 directories in 87650 ms using 4 threads
>>
>> c:\proj>tdir c:\ 1
>> 39708 directories in 92525 ms using 1 threads
>>
>> c:\proj>tdir c:\ 4
>> 39708 directories in 95655 ms using 4 threads
>>
>> c:\proj>tdir c:\ 16
>> 39708 directories in 66138 ms using 21 threads
>>
>> c:\proj>tdir c:\ 8
>> 39708 directories in 69307 ms using 8 threads
>>
>> c:\proj>tdir c:\ 4
>> 39708 directories in 70539 ms using 4 threads
>>
>> c:\proj>tdir c:\ 16
>> 39708 directories in 76392 ms using 32 threads
>>
>>
>>
>> On Sun, Jan 11, 2015 at 1:25 PM, David Gamey <[email protected]>
>> wrote:
>>
>> Sergey,
>>
>> I am responsible for much of the Rosetta code contributions (thanks also
>> to Steve, Andrew, Matt, Peter, and about 4 others) and this one in
>> particular dating from 2010. As I recall this was before the
>> multi-threading versions were widely available. I think multi-threading is
>> underrepresented in Rosetta/Unicon.
>>
>> If you come up with a multi-threading version, we should add it to the
>> post as an alternative version. If you don't feel comfortable doing this,
>> post the code and I can add it.
>>
>> David
>>
>>
>> ------------------------------
>> *From:* Sergey Logichev <[email protected]>
>> *To:* Jafar Al-Gharaibeh <[email protected]>
>> *Cc:* Unicon group <[email protected]>
>> *Sent:* Sunday, January 11, 2015 1:16 AM
>> *Subject:* Re: [Unicon-group] Walk of file directory
>>
>> Jafar,
>>
>> Thank you for a whole bundle of advices and suggestions! Threads are
>> worth to try. The thought of search by file attributes is very useful too.
>> Your suggestion about slow I/O partly is right. For UNIX I tried the
>> program on Raspberry Pi with 6 Class microSD as HDD (it's slow, agree). But
>> for Windows it was quite fast HDD. It would be interesting to compare
>> performance of the program on Windows with classic approach based on Win32
>> _FINDFIRST, _FINDNEXT functions. I have threaded Delphi/Lazarus
>> implementations of this algorithm. Feel that it will be faster but in which
>> degree?
>>
>> Sergey
>>
>> 10.01.2015, 21:50, "Jafar Al-Gharaibeh" <[email protected]>:
>>
>>
>> Sergey,
>>
>> There are so many things that came to mind when I saw your program.
>>
>> 1- At the end of your email, sourceforge ad says "Go Parallel", Which is
>> not a bad idea for this highly parallel application.
>>
>> There is a similar program "wordcount" listed in my dissertation
>> (available on unicon.org) that go through directories and count words in
>> every file using threads (Chapter 7, page 107)
>>
>> 2- Unicon open() already supports " pattern matching that would greatly
>> (I believe) speedup your program. For example you can do this:
>> L := open("*.icn")
>>
>> to get a list of all of Unicon source files in the current directory.
>>
>> Note: It would be nice if there were a way to tell open() to return
>> files not only based on a pattern, but also on file attribute to allow
>> something like "get me all directories in the current directory", or "get
>> me all read only file". There are a lot of situations where filtering
>> directory names for example is very useful - like this program
>>
>> 3- The program on Rosetta Code is not optimized for speed. You can
>> minimize the number of lists created and put() by careful rewriting of the
>> code.
>>
>> 4- Depending on how deep the directory tree is, there might be a lot of
>> I/O going on. A slow disk might limit how fast you can go regardless of how
>> optimized your code is.
>>
>> I will share results if get around trying any of these options.
>>
>> Cheers,
>> Jafar
>>
>>
>>
>> On Sat, Jan 10, 2015 at 5:51 AM, Sergey Logichev <[email protected]>
>> wrote:
>>
>> Hello all!
>>
>> Now I investigate the best approach to get list of files in specified
>> directory and beneath in Unicon.
>> I found excellent example at rosettacode.org:
>> http://rosettacode.org/wiki/Walk_a_directory/Recursively#Icon_and_Unicon
>>
>> I reconstructed this one to implement matching of filenames to specified
>> pattern (regular expression). My program recursively walks a directory and
>> prints appropriate filenames. The same as dir (ls) does. All working fine
>> except performance. If directory has a lot of subdirs the search may took
>> 10-20 seconds before starting output. Could you provide some advices how to
>> enchance the performance?
>>
>> Some notes how to make and use. Unpack content of udir.zip to your local
>> directory. Define which environment you use in env.icn file - uncomment
>> line "$define _UNIX 1" in the case of UNIX. Nothing to do in the case of
>> Windows.
>> Make udir program:
>> unicon -c futils.icn
>> unicon -c options.icn
>> unicon -c regexp.icn
>> unicon udir.icn
>>
>> Usage: udir -f<filemask>
>> for example: udir -f*.icn
>> shall list of icn files in the current dir and all its subdirectories.
>>
>> Best regards,
>> Sergey Logichev
>>
>>
>> ------------------------------------------------------------------------------
>> Dive into the World of Parallel Programming! The Go Parallel Website,
>> sponsored by Intel and developed in partnership with Slashdot Media, is
>> your
>> hub for all things parallel software development, from weekly thought
>> leadership blogs to news, videos, case studies, tutorials and more. Take a
>> look and join the conversation now. http://goparallel.sourceforge.net
>> _______________________________________________
>> Unicon-group mailing list
>> [email protected]
>> https://lists.sourceforge.net/lists/listinfo/unicon-group
>>
>>
>>
>> ------------------------------------------------------------------------------
>> Dive into the World of Parallel Programming! The Go Parallel Website,
>> sponsored by Intel and developed in partnership with Slashdot Media, is
>> your
>> hub for all things parallel software development, from weekly thought
>> leadership blogs to news, videos, case studies, tutorials and more. Take a
>> look and join the conversation now. http://goparallel.sourceforge.net
>>
>> _______________________________________________
>> Unicon-group mailing list
>> [email protected]
>> https://lists.sourceforge.net/lists/listinfo/unicon-group
>>
>>
>>
>>
>>
>>
>>
>> ------------------------------------------------------------------------------
>> New Year. New Location. New Benefits. New Data Center in Ashburn, VA.
>> GigeNET is offering a free month of service with a new server in Ashburn.
>> Choose from 2 high performing configs, both with 100TB of bandwidth.
>> Higher redundancy.Lower latency.Increased capacity.Completely compliant.
>> http://p.sf.net/sfu/gigenet
>> _______________________________________________
>> Unicon-group mailing list
>> [email protected]
>> https://lists.sourceforge.net/lists/listinfo/unicon-group
>>
>>
>
>
>
>
>
> ------------------------------------------------------------------------------
> New Year. New Location. New Benefits. New Data Center in Ashburn, VA.
> GigeNET is offering a free month of service with a new server in Ashburn.
> Choose from 2 high performing configs, both with 100TB of bandwidth.
> Higher redundancy.Lower latency.Increased capacity.Completely compliant.
> http://p.sf.net/sfu/gigenet
> _______________________________________________
> Unicon-group mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/unicon-group
>
>
------------------------------------------------------------------------------
New Year. New Location. New Benefits. New Data Center in Ashburn, VA.
GigeNET is offering a free month of service with a new server in Ashburn.
Choose from 2 high performing configs, both with 100TB of bandwidth.
Higher redundancy.Lower latency.Increased capacity.Completely compliant.
http://p.sf.net/sfu/gigenet
_______________________________________________
Unicon-group mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/unicon-group
