Hello,

Where I work [1] we are thinking about using Protocol Buffers.  However we
have found that the current perl implementation has some performance
issues.  There has been at least one message to this list that touches
on this:

http://groups.google.com/group/protobuf-perl/browse_thread/thread/364...

We are seeing the same issues.  Even for smaller messages (12k-20k) we
are seeing 5-10 seconds for deserialization.  For messages with embedded
messages the serialization also becomes very slow.

So I ask: is there currently a plan or idea of how these performance
issues can be resolved?  If so, what is it?

We are willing to help and at least one of our developers has already
submitted a number of patches to this project.  Before we commit
significant time to this we'd like to know what thought or effort has
already gone into resolving these issues.

[1] I am writing from my personal e-mail address, not my work e-mail
address, fyi.  I work for a company named Oversee.net.  

--
seth /\ sethdaniel.org

Yeah, it appears that Moose is slow.  :-/
Patches welcome, or maybe we need to do a without-Moose rewrite.  At least
the test suite would be reusable, and probably a lot of the wire-level
stuff.

Or perhaps the Moose crew can fix it from their side?

Unfortunately I have about zero time to work on this myself.

On Thu, Sep 18, 2008 at 4:57 AM, Seth Daniel <proto...@sethdaniel.org>wrote:

Yes, most of it is very easy to fix using the same techniques we use
to make simple Moose accessors fast.

Likewise in the short term, but I might be able to put some work
into it this weekend and probably more later.

I'd be more than happy to help out an interested person on IRC or
email.

--
  Yuval Kogman <nothingm...@woobling.org>
http://nothingmuch.woobling.org  0xEBD27418

Since writing that e-mail this has been released:

http://code.google.com/p/protobuf-perlxs/

I tested it out and it's very fast.  For those concerned primarily with
speed I'd recommend looking at it.  

--
seth /\ sethdaniel.org

I don't have time to help much but I'd be very interested in studying the
results of running the code under Devel::NYTProf. If someone could put
up the html reports somewhere (or just email a tarball to me) I may be
able to help in some way. Even if only to help improve NYTProf to give
more insight into the performance of Moose based modules.

Tim.

http://sethd.org/prof/simple/30s/index.html
http://sethd.org/prof/embed/54s/index.html

The results posted in the above links were achieved after applying the
following very simple patch to Class::MOP (patch is also attached):

--- MOP.pm  2008-08-29 15:24:01.000000000 -0700
+++ MOP.pm.new  2008-08-29 15:23:38.000000000 -0700
@@ -128,6 +128,7 @@

     if (ref($class) || !defined($class) || !length($class)) {
         my $display = defined($class) ? $class : 'undef';
+        die if $^S;
         confess "Invalid class name ($display)";
     }

@@ -137,12 +138,14 @@
         my $file = $class . '.pm';
         $file =~ s{::}{/}g;
         eval { CORE::require($file) };
+        die if( $^S && $@ );
         confess "Could not load class ($class) because : $@" if $@;
     }

     # initialize a metaclass if necessary
     unless (does_metaclass_exist($class)) {
         eval { Class::MOP::Class->initialize($class) };
+        die if( $^S && $@ );
         confess "Could not initialize class ($class) because : $@" if $@;
     }

Essentially an enormous amount of time was spent 'confess'ing.  Without this
patch deserialization would take *much* longer.  A 500k message was taking
around 10 minutes to deserialize.  With this patch the time went down to 30s
for a non-embedded message and 54s for an embedded message.  Of course, those
times are still really slow.

For comparison the perlxs generator can serialize a 1.5M message in around
.04s on my machine.

If you need un-patched results I can provide those easily enough.

--
seth /\ sethdaniel.org

class_mop.patch < 1K Download

Almost all that time is spent within this code:
http://sethd.org/prof/simple/30s/Moose-Meta-Attribute.pm-line.html#58

due to this one call to does():
http://sethd.org/prof/simple/30s/Protobuf-Meta-Message.pm-line.html#161

    if ($attr->does('Protobuf::Attribute::Field::Scalar')) {
        $attr->set_value($self, $value);
    } else {
        $attr->push_value($self, $value);
    }

Find some way to avoid calling does() and the performance will be *much*
better.

I don't know if anyone on the Moose list is here but it's clear that Moose
needs a good sprinkling of caching. Nothing too complex.  Invalidate
whenever anything structural changes. Just a simple matter of programming.

Tim.

p.s. NYTProf is amazing! :)