## Rakudo performance

August 2, 2013

Last week I had an amazing adventure in Perl 6. I had a specific goal I needed to accomplish for my day job: determining which face in a STEP model was disappearing when the model was tessellated.

I could determine a face I knew was connected to the face I was looking for, so I decided to try to find all the faces that shared an edge with that face. I have a set of Perl 6 tools for parsing STEP files. It was pretty easy to modify the code I already had to create a script to read & parse the file, walk the topology tree to each edge on the specified face, and then back from those edges to the faces that use them.

This worked well enough when I fed it a small sample file. It did not work so well when I fed it the actual 13M file I needed to work with. I left it running overnight, and it had a fatal error (presumably out of memory) sometime in mid-morning the next day. I think at that point I switched from trying it on Rakudo-Parrot to Rakudo-JVM, with essentially the same results. I added a bunch of say statements, and eventually determined that the line

my %reverse = $step-data.entities.keys.categorize({$step-data.entities{$_}.map(*.entity_instances) });  was the guilty party. This line built the reverse lookup table, so you could find all the entities that reference a given entity — once built, this lets you quickly walk from an edge to each face that used it. Unfortunately, it seemed to eat up all the memory on my 32G Linux box. At this point, I figured using categorize was a mistake. I wasn’t sure why it was, mind you. (Though my test program examining the question ended up leading to a separate impressive optimization from jnthn++.) But it occurred to me there was a different approach that avoided it entirely. Instead of walking back from the edges to the faces, just walk from all the faces to their edges, and then look for faces whose set of edges intersected the given face’s set of edges. Promisingly, this approach was about 33% faster on my small test file. And it didn’t use categorize. However, when I tried this on the big model, this led to the same exact same ballooning memory problem as the first script. Now, though, I had something more concrete to work with, and a few minutes later, I established that the issue was $step-data.entities.kv. On a big hash (this one had 182500 elements), calling .kv or .keys would consume all available memory on my system. You could create the hash, no problem, and if you knew the keys you could access them quite quickly. But asking it to give you the keys killed Rakudo.

With this I had a nice simple golfed test case I could give to jnthn. And before the morning was out, he gave me a patch to fix it. With the patch, the script executed correctly in 11 minutes under Rakudo-JVM. (Note that I did have to change the Java -XmX maximum memory setting in the perl6 script to get it to work.) It also executed correctly in Rakudo-Parrot -- but in 7 hours, 52 minutes.

Let me emphasize that. For this real-world task of significant size, Rakudo-JVM was 40 times faster than Rakudo-Parrot.

Note that I don't know why. But I do know that jnthn++'s fix which was needed to make this work on both Rakudos was in no way JVM-specific. This isn't an example of optimized Rakudo-JVM versus non-optimized Rakudo-Parrot.

But the important thing here is that this represents a very significant improvement in Rakudo's performance by switching to JVM. The script is pretty basic core stuff, mostly file I/O, grammar parsing, and hashes. The improvement is much smaller on a small data-set -- on my small test file, Rakudo-JVM is not even twice as fast as Rakudo-Parrot. But throw a big task (well, this big task, anyway) at Rakudo, and Rakudo-JVM crushes Rakudo-Parrot.

## Set Operations

June 21, 2013

A month or so ago, while working on ABC, I tried to use [∪] (reduce on set union) in my grammar actions. Unfortunately, it blew up when I tried it in Niecza. I assumed the problem was that I hadn’t implemented zero or one arguments forms of set union, and made a quick effort to add them. It didn’t work, so I worked around the problem by using the set constructor directly and added it to my to-do list.

I finally got around to tackling the problem again this week, and my earlier difficulties were quickly explained when I looked at the stack trace for the error message. ∪ turns out to be list associative. That means if you write $a ∪$b ∪ $c, the actual call generated is infix:<∪> ($a, $b,$c), and the reduce meta-op generates that call internally. So I needed to change infix:<∪> from a binary sub to an N-ary sub.

Previously the code looked like this

proto sub infix:<∪>($,$ --> Set) is equiv(&infix:<|>) {*}
multi sub infix:<∪>(Any $a, Any$b --> Set) { $a.Set ∪$b.Set }
multi sub infix:<∪>(Set $a, Set$b --> Set) { Set.new: $a.keys,$b.keys }
multi sub infix:<∪>(Baggy $a, Any$b --> Bag) { $a ∪$b.Bag }
multi sub infix:<∪>(Any $a, Baggy$b --> Bag) { $a.Bag ∪$b }
multi sub infix:<∪>(Baggy $a, Baggy$b --> Bag) {
Bag.new-from-pairs(($a.Set ∪$b.Set).map({ ; $_ =>$a{$_} max$b{$_} })) }  So the rules were, if you had a Baggy (ie Bag or KeyBag), you promoted both arguments to Bag. Otherwise both arguments were promoted to Set. Also note that, because of this, the proto signature was wrong — it could generate a Set or a Bag. I replaced them with a single sub: only sub infix:<∪>(\|$p) is equiv(&infix:<|>) {
my $set = Set.new:$p.map(*.Set.keys);
if $p.grep(Baggy) { my @bags =$p.map(*.Bag);
Bag.new-from-pairs($set.map({ ;$_ => [max] @bags>>.{$_} })); } else {$set;
}
}


Line 2 creates the Set version of the union. Then we check to see if any of the arguments are Baggy using $p.grep(Baggy). (That filters out all the non-Baggy from the argument list, and then (because it is used in a boolean context) returns true if the resulting list has any elements in it.) If there is a Baggy, then we convert all the arguments to Bag. Line 5 is a spiffy bit of code that creates a new Bag using the maximum count for each element found in $set. Finally, if there were no Baggyarguments, just return $set. Once I had that working, I did the same for infix:<∩>. Then I started looking at infix:<(-)>. I’ve only ever seen set difference used as a binary operation, so my first step was to think about how to generalize it to N arguments. My thought was $a (-) $b (-)$c should be ($a (-)$b) (-) $c. If someone has a good reason this shouldn’t be the case, please let me know! Then the next question was, what should Baggy objects do here? Previously set difference always converted its arguments to Set. But it seemed to me there was a pretty obvious way to do it. After consideration, I concluded it should only use the Bag form if the first argument (the “set” things are being subtracted from) was a Baggy. Here’s my current code for infix:<∖> (note the new name, which is the ISO symbol for set difference rather than the common backslash): only sub infix:<∖>(\|$p) is equiv(&infix:<^>) {
return ∅ unless $p; if$p[0] ~~ Baggy {
my @bags = $p.map(*.Bag); my$base = @bags.shift;
Bag.new-from-pairs($base.keys.map({ ;$_ => $base{$_} - [+] @bags>>.{$_} })); } else { my @sets =$p.map(*.Set);
my $base = @sets.shift; Set.new:$base.keys.grep(* ∉ @sets.any );
}
}


This works (I think — I haven’t really tested the Baggy functionality, I’ve only run the existing Set tests). But it breaks some spectests (which assume set difference always returns a Set). And it involves the unspecified details mentioned above, so I’m stepping out on a bit of a limb. So I’m posting here looking for feedback before I commit these changes.

(Edited to add: I’ve also added infix:<⊖> for symmetric difference; that’s one of the standard symbols used, but it’s not ISO standard.)

## Philosophical Issues with Rakudo’s .parse

March 6, 2013

For a little bit now I’ve had the ABC module working on both Niecza and Rakudo. I got a new Linux box yesterday; this morning I was playing around with my ABC tools there, trying to see if I could get them to work. I installed them with panda, so I was using ABC with Niecza by default on my MBP and with Rakudo by default on my new machine. I tried it with a file of my own compositions. Apparently I hadn’t tried this in a while, because it had several issues. And the two systems handled them completely differently.

On Niecza, I got this message:

Unhandled exception: Did not match ABC grammar: last tune understood:
A Hat for the Whales
d|:e~A3 BGEF|G2BG DG (3ABd|eBdB ~A3G|Bdef gfed|
e~A3 BGEF|G2BG DG (3ABd|eBdB AAGA|[1Beed e3d:|


On Rakudo, there were no error messages, but "A Hat for the Whales" was the last tune processed, even though there were a dozen more tunes in the ABC file after it.

Without thinking about it too much I corrected this batch of errors. Now on Niecza the entire file parsed, but I got the error

Unhandled exception: Illegal key signature B minor

Under Rakudo, the same ABC file just appeared to work.

What's going on here? Well, key signature processing looks something like this (simplified):

    token key-def { <basenote> <chord_accidental>? <mode>? }
token mode { <minor> | <major> }
token minor { "m" ["in" ["o" ["r"]?]?]? }
token major { "maj" ["o" ["r"]?]? }


No space allowed between basenote and mode. So when you invoke parse with that rule on “B minor”, in Niecza it fails, because the entire string was not successfully parsed. In Rakudo it succeeds, because “B” is a valid key signature, and it doesn’t care that the entire string was not parsed. In practice, this means that three of the tunes in the collection got the wrong key signature: B major instead of the correct B minor.

To my mind, Niecza’s behavior is more intuitive. If you’re trying to parse a file, parsing half of it and then silently ignoring the rest is wildly unhelpful behavior. In every use I’ve had for parse so far, I’ve wanted it to parse everything in the string I sent to the function.

But hey, I understand that some people might have different needs. Can we at least add a flag that will give you a parsing failure if the entire string is not parsed? :all, perhaps?

## STEP Actions

October 16, 2012

I’ve started adding actions to the STEP code, targeted (at least at first) at collecting up the entity instance names in a parameter list to make it easier to analyze what’s in a STEP file.

I need to do this more often, because I spent about an hour fumbling about making stupid mistakes. And even now that I’ve got it working, I’m haunted by the notion there should be a simpler approach. Anyway, here’s what I’ve got at the moment:

class ISO_10303_21::Actions {
method entity_instance_name($/) { make [~$/] }
method parameter($/) { for <typed_parameter untyped_parameter omitted_parameter> ->$s {
return make $/{$s}.ast if $/{$s}.defined;
}
}
method omitted_parameter($/) { make [] } method untyped_parameter($/) {
for <entity_instance_name list_of_parameters> -> $s { return make$/{$s}.ast if$/{$s}.defined; } make []; } method typed_parameter($/) { make $<parameter>.ast } method list_of_parameters($/) { make merge-arrays(@($<parameter>)».ast); } method parameter_list($/) { make merge-arrays(@($<parameter>)».ast); } }  What’s bugging me are the parameter and untyped_parameter definitions — those seem like they should be easier to express somehow. At least, that basic functionality is going to be pretty common. Oh, and I just found (and fixed) a bug in the code looking at this blog post. Feel like I had something more to say, but I’m tired and heading to bed. ## Parsing STEP October 10, 2012 A month ago I started perl6-ISO_10303-21. In plain English, it’s the beginnings of a toolset for working with CAD files in the STEP format. All it does is parse so far. The STEP (that’s ISO-10303) Part 21 standard provides a BNF (I think) grammar for ASCII STEP files. It was downright trivial to translate it to a Perl 6 grammar — looking at the commits, it took me about two days to get it to read the first file. THen it sat there for a long time until I starting throwing files at it. Since then I’ve made a couple of simple changes. First, I replace C-style comments in the file with spaces before parsing; the spec’s grammar doesn’t mention them at all. Second, I discovered three of the basic sample files I was using were actually illegal, with newlines embedded in strings. I fretted about what to do for a while: should I expand the grammar to accept illegal but “standard” files, or keep the grammar strict so that I could use it to test files my own code exports for correctness? Then I realized I could have my cake and eat it too: grammar ISO_10303_21::LooseGrammar is ISO_10303_21::Grammar { token non_q_char { <special> | <digit> | <space> | <lower> | <upper> | \v } }  I just defined a new subclass, changing the non_q_char token to also accept vertical whitespaces, too. Now ISO_10303_21::Grammar is strict, and ISO_10303_21::LooseGrammar is loose. The next obstacle I’ve run into was speed. The 1.3 meg sample file took about five minutes to prase in Rakudo, and 10+ hours (!) in Niecza. Luckily jnthn++ was on the job, and sent me a patch about fifteen minutes after I sent him the profile of the parse today. (Rakudo’s got --profile built in.) That shaved off 1/6th of the running time. He also pointed out that tokens like token standard_keyword { <upper> [ <upper> | <digit> ]* }  were unnecessarily creating a Match object for each character in the keyword. Since all we want is the full text of the keyword, this could be changed to token standard_keyword { <.upper> [ <.upper> | <.digit> ]* }  and it would suppress those Match objects. That reduced the running time by another 10%. Another few days like this and it will be fast enough to do some useful work for me… ## Sums of Fourth Powers October 5, 2012 Another fun John Cook post has sent me off into playing with numbers and Perl 6. He gives a footnote saying that Euler discovered 635318657 = 158^4 + 59^4 = 134^4 + 133^4 and that this was the smallest number known to be the sum of two fourth powers in two ways. It seems odd now to think of such questions being unresolved. Today we’d ask Hardy “What do you mean 635318657 is the smallest known example? Why didn’t you write a little program to find out whether it really is the smallest?” Perl 6 has features that make that fun to explore, so let’s go to it! Here was my first attempt: my @fourth-powers = (1..*).map(* ** 4) ...^ * > 635318657; my %counts; for @fourth-powers X+ @fourth-powers ->$sum {
%counts{$sum}++; } for %counts.keys.grep({ %counts{$_} > 2 }) -> $sum { say "$sum has { %counts{$sum} / 2 } fourth power sums"; }  This works, and is quite fast in Niecza. (It’s much slower in Rakudo for some reason, though it still finishes in about ten seconds.) I love that first line, which constructs a lazy infinite list of fourth powers, then truncates it at 635318657. Then it uses X+ to build all the sums of two fourth powers less than (or equal to) 635318657. The rest of the code is pretty mundane, I fear. I decided to try to use classify to get rid of the explicit loops, resulting in this code: my @fourth-powers = (1..*).map(* ** 4) ...^ * > 635318657; my @results = (@fourth-powers X+ @fourth-powers).classify({$_ }).pairs.grep(*.value > 2);
say @results.map({ $_.key ~ " " ~$_.value.Int / 2 ~ " times" });


Unfortunately, this doesn’t work in Niecza. So now I’ve got some bugs to fix…

July 2, 2012

So, I’ve hit on something extremely annoying working with Niecza on my MacBook Pro. I upgraded to Mono 2.10.9 a few months back. Since then, Glib# has not worked. Even after I uninstalled Mono 2.10.9. It is incredibly frustrating.

Here’s the error I’m getting:

Unhandled exception: System.TypeInitializationException: An exception was thrown by the type initializer for GLib.GType ---> System.DllNotFoundException: glibsharpglue-2


Here’s the call:

constant $GLIB = "glib-sharp, Version=2.12.0.0, Culture=neutral, PublicKeyToken=35e10195dab3c99f"; constant$G_TYPE_STRING = CLR::("GLib.GType,$GLIB").String;  Here’s what gacutil says about glib-sharp: glib-sharp, Version=2.12.0.0, Culture=neutral, PublicKeyToken=35e10195dab3c99f  If anyone has any suggestions on how to fix this, I’m all ears. ## Handling Errors in ABC May 24, 2012 When I started thinking I should write yesterday’s post and started testing the Ed Reavy ABC file, I quickly ran into one of the module’s limitations that I thought I should fix before claiming ABC was a useful program. When I started trying the Reavy file, if the ABC file ran into an error in the ABC file, it would frequently just hang. Now, it was relatively easy to track down the erroneous ABC at that point, because the ABC Actions class prints all the title fields it finds as it finds them. You’d watch the list of tune names going by, and if it stopped the list before the tune file was done and just sat there, you knew something was wrong with that tune. That’s okay for a toy program that only I play with, but it’s unacceptable for a serious program other people can use. I had a theory about what caused it: when the grammar hit something it didn’t understand, it started backtracking to try to find a correct way of interpreting the file. Since there was no correct way, that resulted in a painfully long attempt to repeatedly re-interpret the bad data. So I asked how to stop backtracking grammar regexes on #perl6. They helpfully suggested changing the ABC grammar’s regex methods to token, because token suppresses backtracking. I tried just changing one method, and that didn’t help. I tried changing them all, and that failed some tests. I changed one method back, and suddenly all the tests passed, and the grammar no longer hung on ABC file errors! That was a huge improvement, but it was still only marginally helpful at figuring out what the problem in ABC file was. Once I was done with yesterday’s post, I tried to figure out how to get better information about where the parsing problem was. I hit upon a simple method. I added a $.current-tune attribute to the ABC::Actions class. When the action for a tune title fires, it clears this attribute and sets it to the tune title:

    method header_field($/) { if$<header_field_name> eq "T" {
$*ERR.say: "Parsing " ~$<header_field_data>;
$.current-tune =$<header_field_data> ~ "\n";
}

make ~$<header_field_name> => ~$<header_field_data>;
}


When the action method for the bar token fires, it adds the string of the bar to $.current-tune. When the action method for line_of_music fires, it adds a newline. And when a fatal error occurs parsing an ABC file, it prints out this attribute, resulting in error messages that look like this: Unhandled exception: Did not match ABC grammar: last tune understood: Silent The Lonely Glen g/2f/2|ed/2B/2 AG/2A/2|B/2d^c/2 dg/2f/2|ed/2B/2 AG/2A/2| B/2>A/2G/2  While this isn’t an exact reproduction of what’s in the tune, it’s close enough to easily show where the problem is: g/2f/2|ed/2B/2 AG/2A/2|B/2d^c/2 dg/2f/2|ed/2B/2 AG/2A/2| B/2>A/2G/2/F/2 eg/2f/2|ed/2B/2 ba/2g/2|g/2f/2e/2^d/2|  You just look at where the grammar stopped understanding the file, and the error pops right out if you know ABC: B/2>A/2G/2/F/2. Delete that duration not attached to a note, and the grammar will zip right through the tune. I’m sure there are more sophisticated ways to handle errors in parsing available in Perl 6, because the STD grammar uses them to give you great error messages. But this is a simple technique which is pretty handy. ## About ABC May 23, 2012 I’ve referenced the ABC module a number of times over the years on this blog, but I don’t think I’ve ever properly explained it. I think it’s particularly relevant with the recurring discussions going on about Perl 6′s “production readiness”, because at this point the abc2ly.pl script in the module is a productive tool I am using on a regular basis. One that would have been harder to write in almost any other programming language, I should add. ABC format is a designed to make it very easy to enter single-line music. It’s designed to be a terse ASCII representation of the music, and it’s absolutely terrific for notating jigs and reels and things like that. Here’s an example ABC hornpipe: X:20 T:Two Weeks To Wait M:4/4 L:1/8 C:Solomon Foster R:Hornpipe K:Edor GF|:E2GB AGFA|(3Bcd ed gedB|A2GB ABGA|BgdB A2BG| E2GB AGFA|(3Bcd ed gedB|A2GB ABGF|E2ED E2GF:| |:E2GB edef|g2fg edBG|A2GB ABGA|BgdB e2ef| g2fg edBG|~A3B AGEG|~A3B AGFG|E2ED E2GF:|  A tune considers of a header section with information about the tune and the basic glue to define the tune, followed by a list of notes, barlines, repeat signs, and other symbols. The ABC module’s abc2ly.pl script uses the ABC module’s ABC library to read the file, and outputs the music in the Lilypond format. Lilypond produces very nice output, but the file format is a pain. For instance, the (3Bcd triplet in the above file is translated to this: \times 2/3 { b'8[ cis''8 d''8] }. But the PDF Lilypond outputs from it is beautiful. I’ve been using this tool to handle my sheet music creation needs for about a year now, but in the last six weeks I’ve really stepped things up. I’ve added multi-note stems, chord names, text annotations, ties, slurs, and many other markings I needed, as well as fixing a number of bugs. As a test I threw the Ed Reavy ABC collection at it today. (After deleting the comment lines in the beginning, because I haven’t added them to the parser yet.) In the process I found four mistakes in the ABC file which threw the parser. Once they were fixed, it processed all 127 tunes in about 22 seconds. I haven’t examined the 44 pages of sheet music output by Lilypond closely yet, but it looks like all the problems I see with the output are actually errors (that parse) in the ABC file. In conclusion, while it doesn’t implement the entire ABC standard yet, this is a perfectly usable ABC to Lilypond translator, capable of handling fairly complex single-staff ABC files quickly and accurately. It’s definitely reached the point where it’s more than reasonable for other people to start using it. ## Optimizing abc2ly.pl May 21, 2012 Shortly after I wrote my last blog post, I started actively working on the ABC module again. First I had to enhance a few features so it could handle some whistle parts for a little musical I was asked to play for. Next I needed some enhancements to handle whistle / fiddle parts for the standard wedding music pieces for an upcoming gig. Finally getting obsessed with ABC updates, I decided to throw the code at a collection of 159 tunes I put together a few years ago. Suddenly I was running into hard limitations of the ABC module. It was always slow, but taking 40 seconds to process my file of three tunes wasn’t really a problem for me. I did add some progress notification messages, and was shocked to learn it wasn’t the parsing of the ABC files that was slow, it was processing the parsed result. But when I got things working well enough it was actually trying to process all 159 tunes, it would run for 15 or 20 minutes and then crash. (This is under Niecza, BTW.) I spent Friday and Saturday trying to optimize bits of code in there that I knew were terrible, with no noticeable improvement. Finally I did a profile run under mono. I found the results didn’t really mean anything to me, but I sent it to sorear++ to see what he thought. And he said, “Why do you need to use .eval in abc2ly.pl?” Well, I quickly checked, and found all the uses were intended to convert strings like "3/4" to Rats. You may say “Why were you doing that? Shouldn’t .Numeric work just as well?” And the answer is yes, it does. (In my defense, I’m pretty sure that didn’t work when I started working on the ABC module.) But now here is the tricky bit. Here’s a simple script to demonstrate the issue: my$a = 0;
for 1..400 {
$a += "1/$_";
}
say $a;  On my machine, that takes 1.3 seconds to run under Niecza. Now let’s make a simple change: my$a = 0;
for 1..400 {
$a += "1/$_".eval;
}
say \$a;


This version, with just the .eval added, takes 39.0 seconds to run. Huge difference, eh?

It turned out that even though the evals were happening per measure of music and not per note, they were so slow they were completely swamping the performance of my script. With the evals, processing the wedding.abc file took about 28 seconds. Without them, 3.7 seconds. And remember the big ABC file that would run for 15 minutes and then croak? Without the evals, the entire file processes in 31.8 seconds. That’s easily 30 times faster.

Lesson? Don’t use eval unless you really, really need to.

This advice holds for Rakudo as well: the above fraction-summing scripts took .3 and 8.5 seconds to run, respectively. (Yes, Rakudo is doing a lot better performance-wise on those scripts than Niecza did!)