Re^2: selecting columns from a tab-separated-values file

[sundialsvc4]

in reply to Re: selecting columns from a tab-separated-values file
in thread selecting columns from a tab-separated-values file

Take hammer.   See nail.   Hit nail on the head.

Physical latencies – seek-time, basically – is really the only thing that matters here.   Although the operating-system’s buffer pools will mitigate the seeking somewhat, you still want to hold that to a minimum ... taking care that I/O doesn’t sneak up on you from the backside in the form of virtual-memory paging.

I’m personally not sure that threads would help here ... although I do yield to your greater expertise on this matter ... but I think that simply buffering a few hundred thousand extracted records, e.g. in an array of hashrefs, just might hit pay dirt.   Instead of writing each record out immediately, push it onto an array (of hashrefs).   When the array reaches some agreed-upon and easily-tweakable threshhold, pause to go shift them all off and write them to the output file, then continue reading.   (The read/write head moves from over-here to over-there, stays there a while to write the stuff out, then moves back.)   Set the threshhold to some point where you can be fairly sure that there will be enough physical RAM available to hold everything without paging.   I suspect that you will be astonished at what just-this does for the program.

[BrowserUk]

taking care that I/O doesn’t sneak up on you from the backside in the form of virtual-memory paging.

You and your hobby horses. Virtual memory limits don't enter into it.

When running, those two programs I posted use 2.7 MB and 3.2 MB respectively when using the 256 times larger read size than standard that I suggest. Even if I increase that 10-fold to 10MB each -- which slows the processing down -- they use a whole 12 MB & 16 MB. The programmer that runs my heating system could handle that.

I’m personally not sure that threads would help here

I'm not sure that standing on one leg whilst drinking blood from a freshly decapitated chicken would help; so I don;t mention it.

And your instincts prejudiced guessing is wrong!

Approximately 2/3rds of the throughput gains from my posted solution come exactly because the CPU intensive part of the processing -- the spliting and joining of the records -- can run flat out (100% utilisation) on one CPU, whilst the first thread doing the Input is instantly ready to respond to the completion of each read because it isn't having to perform any CPU intensive processing on each record.

push it onto an array (of hashrefs).

The input is a stream; the output is a stream; What on earth do you need an array of hashrefs for?

I suspect that you will be astonished at what just-this does for the program.

No. I can pretty much call it without trying it. It will run 3 to 5 times slower than a simple:

perl -F"\t" -anle"print join chr(9), @F[2,0,5]" in.tsv > out.tsv

That is, instead of taking the OPs 5+ hours to run it will take closer to 24 hours.

Why? You'll never know unless you try it. And you won't do that.

(And even if you get someone else to do it for you, you won't post the results, because it will show your 'advice' to be nothing more than groundless guessing.)

[sundialsvc4]

Bah, humbug ...   ;-)

There’s one million-pound elephant in this room, if there is one at all, and that elephant certainly won’t consist of the mere-microseconds that a CPU will require to split-up a chunk of data that came in from a disk drive.   We’re not reading graphic-images here and resizing them ... the task to be performed against each unit of data is quite inconsequential.   The advantage comes from being able to gulp large amounts of data at a time from a disk device that (probably ...) does not have to spend too much time dong it.   (Although, depending on the peculiarities of physical distribution of the file-in-question across the disk real estate, that might not turn out to be so.)

“What is the array of hashrefs for?”   Why, you yourself said it!   Even though “the input is a stream, and the output is a stream,” there is probably only one device.   A physical operation consisting of seek + read or seek + write, which we desire to reduce to the point that no seek is required.   We want:   seek($$) + read + read + read + ... + read + seek($$) + write + write .... + write + seek ($$).

The specific reason why I questioned the validity of threads in this scenario, was that I judged the amount of CPU-intensive processing required to be pragmatically inconsequential relative to the amount of time that would be spent in an I/O-wait.   Also, it was because I judged that the physical latency cost of the disk device would be the “above all, ruling constraint” no matter what.   Pragmatically, we can afford to spend a few microseconds crunching numbers because the disk platter won’t have rotated too far during that time anyway.

But there is only so far that the point can be argued before it becomes merely an argument.   If the per-record CPU load is indeed slight to the point of being inconsequential, then memory is a big fat available buffer that costs nothing ... if, indeed paging is not occurring, so that it really does cost “nothing.”   But if a page-fault that must be resolved to disk does take place, then you just moved that read/write head after all.   Overlapping I/O with CPU-processing might prove to be beneficial on the OP’s system, not yours... or not.   If it were me, I would start by using the RAM as a buffer, see how much bang that buck got me, then pragmatically move forward from there.

It doesn’t advance the technical validity of your arguments to belittle the opinions others, you know... no matter how personally self-assured you might be.

[BrowserUk]

It doesn’t advance the technical validity of your arguments to belittle the opinions others, you know... no matter how personally self-assured you might be.

I am not belittleing your opinion. There is no matter of opinion involved. This is not a subjective matter; it is a quantifiable, easily measured, matter of right and wrong; fact and fiction. And you provided fiction, whilst I provided fact.

I am telling directly and definitively and unambiguously, that the technical content -- such as it is -- in both your first post and this one are both theoretically -- and demonstrably -- factually incorrect. Ie. You are wrong!

[Lotus1]

It turns out that BrowserUK's approach was an order of magnitude quicker than the other approaches presented here. I took the liberty of coding up Sundialsvc4's suggestion of buffering a few hundred thousand lines worth of data before printing them out to a file. I used array refs since there was no reason to use hash refs.

I used a Powershell script to time the different approaches.

echo 'BrowserUK 1024' > bench.txt
Measure-Command { .\buk.bat } >> bench.txt
echo 'BrowserUK 2048' >> bench.txt
Measure-Command { .\buk2048.bat } >> bench.txt
echo 'Sundial approach by Lotus1' >> bench.txt
Measure-Command { perl sundial.pl test.csv > dialout.csv} >> bench.txt
echo spacebar >> bench.txt
Measure-Command { .\sed -n "s/^\([^\t]*\t\)[^\t]*\t\([^\t]*\t\)[^\t]*\
+t[^\t]*\t\([^\t]*\)\t.*/\2\1\3/p" test.csv > spaceout.csv} >> bench.t
+xt
echo 'kenosis/choroba' >> bench.txt
Measure-Command { perl kc.pl test.csv > kcout.csv} >> bench.txt
echo 'mildside -- wrong order but testing anyway' >> bench.txt
Measure-Command { .\cut '-f1,3,6' test.csv > cutout.csv} >> bench.txt

Here are the results using a 1Gb test file on an idle server with 16 cores and 8Gb RAM: update: (Windows Server 2008 R2)

BrowserUK 1024
Minutes           : 1
Seconds           : 54
Milliseconds      : 103
Ticks             : 1141033211

BrowserUK 2048
Minutes           : 1
Seconds           : 55
Milliseconds      : 124
Ticks             : 1151241665

Sundial approach by Lotus1
Minutes           : 21
Seconds           : 53
Milliseconds      : 28
Ticks             : 13130283183

spacebar
Minutes           : 21
Seconds           : 24
Milliseconds      : 215
Ticks             : 12842154788

kenosis/choroba
Minutes           : 22
Seconds           : 4
Milliseconds      : 865
Ticks             : 13248658887

mildside -- wrong order but testing anyway
Minutes           : 22
Seconds           : 19
Milliseconds      : 295
Ticks             : 13392954883

Here is the sundialsvc4 approach that I put together for the test:

#! perl -sw
use strict;

my $count = 0;
my @lines;
while( <> ) {
    $count++;
    my @f = ( split /\t/, $_, 7 )[ 0, 2, 5 ];
    push @lines, \@f;
    
    if( $count >= 300000 ) {  #size of buffer
        $count = 0;
        print_buffer( \@lines );
    }
}
print_buffer( \@lines );

sub print_buffer {
    my ($aref) = @_;
    foreach (@$aref) {
        print join( "\t", @$_ ) . "\n";
    }
    splice( @$aref );
}

Here is the Kenosis/choroba approach.

#! perl -sw
use strict;

#kenosis/choroba approach

while( <> ) {
    print join( "\t", ( split /\t/, $_, 7 )[ 0, 2, 5 ]), "\n";
}

[mildside]

Great job there Lotus1.

I'm curious about the use of splice to clear the array in your code, as below:

    splice( @$aref );

I would probably have used the below:

    @$aref = ();

Is splice faster or better in some other way?

[Lotus1]

Thanks. I started with assigning the empty array but I had a bug somewhere so I stuck the splice in and got it working. I think I forgot the '@' in the first try but put it in with splice. I don't know which is faster but it is only called a handful of times in this approach anyway.