Monday, July 19, 2010

What JRuby C Extension Support Means to You

As part of the Ruby Summer of Code, Tim Felgentreff has been building out C extension support for JRuby. He's already made great progress, with simple libraries like Thin and Mongrel working now and larger libraries like RMagick and Yajl starting to function. And we haven't even reached the mid-term evaluation yet. I'd say he gets an "A" so far.

I figured it was time I talked a bit about C extensions, what they mean (or don't mean) for JRuby, and how you can help.

The Promise of C Extensions

One of the "last mile" features keeping people from migrating to JRuby has been their dependence on C extensions that only work on regular Ruby. In some cases, these extensions have been written to improve performance, like the various json libraries. Some of that performance could be less of a concern under Ruby 1.9, but it's hard to claim that any implementation will be able to run Ruby as fast as C for general-purpose libraries any time soon.

However, a large number of extensions – perhaps a majority of extensions – exist only to wrap a well-known and well-trusted C library. Nokogiri, for example, wraps the excellent libxml. RMagick wraps ImageMagick. For these cases, there's no alternative on regular Ruby...it's the C library or nothing (or in the case of Nokogiri, your alternatives are only slow and buggy pure-Ruby XML libraries).

For the performance case, C extensions on JRuby don't mean a whole lot. In most cases, it would be easier and just as performant to write that code in Java, and many pure-Ruby libraries perform well enough to reduce the need for native code. In addition, there are often libraries that already do what the perf-driven extensions were written for, and it's trivial to just call those libraries directly from Ruby code.

But the library case is a bit stickier. Nokogiri does have an FFI version, but it's a maintenance headache for them and a bug report headache for us, due to the lack of a C compiler tying the two halves together. There's a pure-Java Nokogiri in progress, but building both the Ruby bindings and emulating libxml behavior takes a long time to get right. For libraries like RMagick or the native MySQL and SQLite drivers, there are basically no options on the JVM. The Google Summer of Code project RMagick4J, by Sergio Arbeo, was a monumental effort that still has a lot of work left to be done. JDBC libraries work for databases, but they provide a very different interface from the native drivers and don't support things like UNIX domain sockets.

There's a very good chance that JRuby C extension support won't perform as well as C extensions on C Ruby, but in many cases that won't matter. Where there's no equivalent library now, having something that's only 5-10x slower to call – but still runs fast and matches API – may be just fine. Think about the coarse-grained operations you feed to a MySQL or SQLite and you get the picture.

So ultimately, I think C extensions will be a good thing for JRuby, even if they only serve as a stopgap measure to help people migrate small applications over to native Java equivalents. Why should the end goal be native Java equivalents, you ask?

The Peril of C Extensions

Now that we're done with the happy, glowing discussion of how great C extension support will be, I can make a confession: I hate C extensions. No feature of C Ruby has done more to hold it back than the desire for backward compatibility with C extensions. Because they have direct pointer access, there's no easy way to build a better garbage collector or easily support multiple runtimes in the same VM, even though various research efforts have tried. I've talked with Koichi Sasada, the creator of Ruby 1.9's "YARV" VM, and there's many things he would have liked to do with YARV that he couldn't because of C extension backward compatibility.

For JRuby, supporting C extensions will limit many features that make JRuby compelling in the first place. For example, because C extensions often use a lot of global variables, you can't use them from multiple JRuby runtimes in the same process. Because they expect a Ruby-like threading model, we need to restrict concurrency when calling out from Java to C. And all the great memory tooling I've blogged about recently won't see C extensions or the libraries they call, so it introduces an unknown.

All that said, I think it's a good milestone to show that we can support C extensions, and it may make for a "better JNI" for people who really just want to write C or who simply need to wrap a native library.

How You Can Help

There's a few things I think users like you can help with.

First off, we'd love to know what extensions you are using today, so we can explore what it would take to run them under JRuby (and so we can start exploring pure-Java alternatives, too.) Post your list in the comments, and we'll see what we can come up with.

Second, anyone that knows C and the Ruby C API (like folks who work on extensions) could help us fill out bits and pieces that are missing. Set up the JRuby cext branch (I'll show you how in a moment), and try to get your extensions to build and load. Tim has already done the heavy lifting of making "gem install xyz" attempt to build the extension and "require 'xyz'" try to load the resulting native library, so you can follow the usual processes (including extconf.rb/mkmf.rb for non-gem building and testing.) If it doesn't build ok, help us figure out what's missing or incorrect. If it builds but doesn't run, help us figure out what it's doing incorrectly.

Building JRuby with C Extension Support

Like building JRuby proper, building the cext work is probably the easiest thing you'll do all day (assuming the C compiler/build/toolchain doesn't bite you.

  1. Check out (or fork and check out) the JRuby repository from http://github.com/jruby/jruby:
    git clone git://github.com/jruby/jruby.git

  2. Switch to the "cext" branch:
    git checkout -b cext origin/cext

  3. Do a clean build of JRuby plus the cext subsystem:
    ant clean build-jruby-cext-native

At this point you should have a JRuby build (run with bin/jruby) that can gem install and load native extensions.

Saturday, July 17, 2010

Browsing Memory with Ruby and Java Debug Interface

This is the third post in a series. The first two were on Browsing Memory the JRuby Way and Finding Leaks in Ruby Apps with Eclipse Memory Analyzer

Hello again, friends! I'm back with more exciting memory analysis tips and tricks! Ready? Here we go!

After my previous two posts, several folks asked if it's possible to do all this stuff from Ruby, rather than using Java or C-based apps shipped with the JVM. The answer is yes! Because of the maturity of the Java platform, there are standard Java APIs you can use to access all the same information the previous tools consumed. And since we're talking about JRuby, that means you have Ruby APIs you can use to access that information.

That's what I'm going to show you today.

Introducing JDI

The APIs we'll be using are part of the Java Debug Interface (JDI), a set of Java APIs for remotely inspecting a running application. It's part of the Java Platform Debugger Architecture, which also includes a C/++ API, a wire protocol, and a raw wire protocol API. Exploring those is left as an exercise for the reader...but they're also pretty cool.

We'll use the Rails app from before, inspecting it immediately after boot. JDI provides a number of ways to connect up to a running VM, using VirtualMachineManager; you can either have the debugger make the connection or the target VM make the connection, and optionally have the target VM launch the debugger or the debugger launch the target VM. For our example, we'll have the debugger attach to a target VM listening for connections.

Preparing the Target VM

The first step is to start up the application with the appropriate debugger endpoint installed. This new flag is a bit of a mouthful (and we should make a standard flag for JRuby users), but we're simply setting up a socket-based listener on port 12345, running as a server, and we don't want to suspend the JVM when the debugger connects.

jruby -J-agentlib:jdwp=transport=dt_socket,server=y,address=12345,suspend=n -J-Djruby.reify.classes=true script/server -e production

The -J-Djruby.reify.classes bit I talked about in my first post. It makes Ruby classes show up as Java classes for purposes of heap inspection.

The rest is just running the server in production mode.

As you can see, remote debugging is already baked into the JVM, which means we didn't have to write it or debug it. And that's pretty awesome.

Let's connect to our Rails process and see what we can do.

Connecting to the target VM

In order to connect to the target VM, you need to do the Java factory dance. We start with the com.sun.jdi.Bootstrap class, get a com.sun.jdi.VirtualMachineManager, and then connect to a target VM to get a com.sun.jdi.VirtualMachine object.
vmm = Bootstrap.virtual_machine_manager
sock_conn = vmm.attaching_connectors[0] # not guaranteed to be Socket
args = sock_conn.default_arguments
args['hostname].value = "localhost"
args['port'].value = "12345"
vm = sock_conn.attach(args)

Notice that I didn't dig out the socket connector explicitly here, because on my system, the first connector always appears to be the socket connector. Here's the full list for me on OS X:
➔ jruby -rjava -e "puts com.sun.jdi.Bootstrap.virtual_machine_manager.attaching_connectors
> "
[com.sun.jdi.SocketAttach (defaults: timeout=, hostname=charles-nutters-macbook-pro.local, port=),
com.sun.jdi.ProcessAttach (defaults: pid=, timeout=)]

The ProcessAttach connector there isn't as magical as it looks; all it does is query the target process to find out what transport it's using (dt_socket in our case) and then calls the right connector (e.g. SocketAttach in the case of dt_socket or SharedMemoryAttach if you use dt_shmem on Windows). In our case, we know it's listening on a socket, so we're using the SocketAttach connector directly.

The rest is pretty simple: we get the default arguments from the connector, twiddle them to have the right hostname and port number, and attach to the VM. Now we have a VirtualMachine object we can query and twiddle; we're inside the matrix.

With Great Power...

So, what can we do with this VirtualMachine object? We can:
  • walk all classes and objects on the heap
  • install breakpoints and step-debug any running code
  • inspect and modify the current state of any running thread, even manipulating in-flight arguments and variables
  • replace already-loaded classes with new definitions (such as to install custom instrumentation)
Here's the output from JRuby's ri command when we ask about VirtualMachine:
➔ ri --java com.sun.jdi.VirtualMachine
-------------------------------------- Class: com.sun.jdi.VirtualMachine
(no description...)
------------------------------------------------------------------------


Instance methods:
-----------------
allClasses, allThreads, canAddMethod, canBeModified,
canForceEarlyReturn, canGetBytecodes, canGetClassFileVersion,
canGetConstantPool, canGetCurrentContendedMonitor,
canGetInstanceInfo, canGetMethodReturnValues,
canGetMonitorFrameInfo, canGetMonitorInfo, canGetOwnedMonitorInfo,
canGetSourceDebugExtension, canGetSyntheticAttribute, canPopFrames,
canRedefineClasses, canRequestMonitorEvents,
canRequestVMDeathEvent, canUnrestrictedlyRedefineClasses,
canUseInstanceFilters, canUseSourceNameFilters,
canWatchFieldAccess, canWatchFieldModification, classesByName,
description, dispose, eventQueue, eventRequestManager, exit,
getDefaultStratum, instanceCounts, mirrorOf, mirrorOfVoid, name,
process, redefineClasses, resume, setDebugTraceMode,
setDefaultStratum, suspend, toString, topLevelThreadGroups,
version, virtualMachine

We can basically make the target VM dance any way we want, even going so far as to write our own debugger entirely in Ruby code. But that's a topic for another day. Right now, we're going to do some memory inspection.

Creating a Histogram of the Heap
The simplest heap inspection we might do is to produce a histogram of all objects on the heap. And as you might expect, this is one of the easiest things to do, because it's the first thing everyone looks for when debugging a memory issue.
classes = VM.all_classes
counts = VM.instance_counts(classes)
classes.zip(counts)

VirtualMachine.all_classes gives you a list (a java.util.List, but we make those behave mostly like a Ruby Array) of every class the JVM has loaded, including Ruby classes, JRuby core and runtime classes, and other Java classes that JRuby and the JVM use. VirtualMachine.instance_counts takes that list of classes and returns another list of instance counts. Zip the two together, and we have an array of classes and instance counts. So easy!

Let's take these two pieces and put them together in an easy-to-use class
require 'java'

module JRuby
class Debugger
VMM = com.sun.jdi.Bootstrap.virtual_machine_manager

attr_accessor :vm

def initialize(options = {})
connectors = VMM.attaching_connectors
if options[:port]
connector = connectors.find {|ac| ac.name =~ /Socket/}
elsif options[:pid]
connector = connectors.find {|ac| ac.name =~ /Process/}
end

args = connector.default_arguments
for k, v in options
args[k.to_s].value = v.to_s
end

@vm = connector.attach(args)
end

# Generate a histogram of all classes in the system
def histogram
classes = @vm.all_classes
counts = @vm.instance_counts(classes)
classes.zip(counts)
end
end
end

I've taken the liberty of expanding the connection process to handle pids and other arguments passed in. So to get a histogram from a VM listening on localhost port 12345, we can simply do:
JRuby::Debugger.new(:hostname => 'localhost', :port => 12345).histogram

Now of course this list is going to have a lot of JRuby and Java objects that we might not be interested in, so we'll want to filter it to just the Ruby classes. On JRuby master, all the generated Ruby classes start with a package name "ruby". Unfortunately, jitted Ruby methods start with a package of "ruby.jit" right now, so we'll want to filter those out too (unless you're interested in them, of course...JRuby is an open book!)
require 'jruby_debugger'

# connect to the VM
debugr = JRuby::Debugger.new(:hostname => 'localhost', :port => 12345)
histo = debugr.histogram
# sort by count
histo.sort! {|a,b| b[1] <=> a[1]}
# filter to only user-created Ruby classes with >0 instances
histo.each do |cls,num|
next if num == 0 || cls.name[0..4] != 'ruby.' || cls.name[5..7] == 'jit'
puts "#{num} instances of #{cls.name[5..-1].gsub('.', '::')}"
end

If we run this short script against our Rails application, we see similar results to the previous posts (but it's cooler, because we're doing it all from Ruby!)
➔ jruby ruby_histogram.rb | head -10
11685 instances of TZInfo::TimezoneTransitionInfo
1071 instances of Gem::Version
1012 instances of Gem::Requirement
592 instances of TZInfo::TimezoneOffsetInfo
432 instances of Gem::Dependency
289 instances of Gem::Specification
142 instances of ActiveSupport::TimeZone
118 instances of TZInfo::DataTimezoneInfo
118 instances of TZInfo::DataTimezone
45 instances of Gem::Platform

Just so we're all on the same page, it's important to know what we're actually dealing with here. VirtualMachine.all_classes returns a list of com.sun.jdi.ReferenceType objects. Let's ri that.
➔ ri --java com.sun.jdi.ReferenceType
--------------------------------------- Class: com.sun.jdi.ReferenceType
(no description...)
------------------------------------------------------------------------


Instance methods:
-----------------
allFields, allLineLocations, allMethods, availableStrata,
classLoader, classObject, compareTo, constantPool,
constantPoolCount, defaultStratum, equals, failedToInitialize,
fieldByName, fields, genericSignature, getValue, getValues,
hashCode, instances, isAbstract, isFinal, isInitialized,
isPackagePrivate, isPrepared, isPrivate, isProtected, isPublic,
isStatic, isVerified, locationsOfLine, majorVersion, methods,
methodsByName, minorVersion, modifiers, name, nestedTypes,
signature, sourceDebugExtension, sourceName, sourceNames,
sourcePaths, toString, virtualMachine, visibleFields,
visibleMethods

You can see there's quite a bit more you can do with a ReferenceType. Let's try something.

Digging Deeper Into TimezoneTransitionInfo

Let's actually take some time to explore our old friend TimezoneTransitionInfo (hereafter referred to as TTI). Instead of walking all classes in the system, we'll want to just grab TTI directly. For that we use VirtualMachine.classes_by_name, which returns a list of classes on the target VM of that name. There should be only one, since we only have a single JRuby instance in our server, so we'll grab that class and request exactly one instance of it...any old instance.
tti_class = debugr.vm.classes_by_name('ruby.TZInfo.TimezoneTransitionInfo')[0]
tti_obj = tti_class.instances(1)[0]
puts tti_obj

Running this we can see we've got the reference we're looking for.
➔ jruby tti_digger.rb
instance of ruby.TZInfo.TimezoneTransitionInfo(id=2)

ReferenceType.instances returns a list (no larger than the specified size, or all instances if you specify 0) of com.sun.jdi.ObjectReference objects.
➔ ri --java com.sun.jdi.ObjectReference
------------------------------------- Class: com.sun.jdi.ObjectReference
(no description...)
------------------------------------------------------------------------


Instance methods:
-----------------
disableCollection, enableCollection, entryCount, equals, getValue,
getValues, hashCode, invokeMethod, isCollected, owningThread,
referenceType, referringObjects, setValue, toString, type,
uniqueID, virtualMachine, waitingThreads

Among the weirder things like disabling garbage collection for this object or listing all threads waiting on this object's monitor (a la 'synchronize' in Java), we can access the object's fields through getValue and setValue.

Let's examine the instance variables TTI contains. You may recall from previous posts that all Ruby objects in JRuby store their instance variables in an array, to avoid the large memory and cpu cost of storing them in a map. We can grab a reference to that array and display its contents.
var_table_field = tti_class.field_by_name('varTable')
tti_vars = tti_obj.get_value(var_table_field)
puts "varTable: #{tti_vars}"
puts tti_vars.values.map(&:to_s)

And the new output:
➔ jruby tti_digger.rb
varTable: instance of java.lang.Object[7] (id=13)
instance of ruby.TZInfo.TimezoneOffsetInfo(id=15)
instance of ruby.TZInfo.TimezoneOffsetInfo(id=16)
instance of org.jruby.RubyFixnum(id=17)
instance of org.jruby.RubyFixnum(id=18)
instance of org.jruby.RubyNil(id=19)
instance of org.jruby.RubyNil(id=19)
instance of org.jruby.RubyNil(id=19)

Since the varTable field is a simple Object[] in Java, the reference we get to it is of type com.sun.jdi.ArrayReference.
➔ ri --java com.sun.jdi.ArrayReference
-------------------------------------- Class: com.sun.jdi.ArrayReference
(no description...)
------------------------------------------------------------------------


Instance methods:
-----------------
disableCollection, enableCollection, entryCount, equals, getValue,
getValues, hashCode, invokeMethod, isCollected, length,
owningThread, referenceType, referringObjects, setValue, setValues,
toString, type, uniqueID, virtualMachine, waitingThreads

Of course each of these references can be further explored, but already we can see that this TTI instance has seven instance variables: two TimezoneOffsetInfo objects, two Fixnums, and three nils. But we don't have instance variable names!

Instance variable names are only stored on the object's class. There, a table of names to offsets is kept up-to-date as new instance variable names are discovered. We can access this from the TTI class reference and combine it with the variable table to get the output we want to see.
# get the metaclass object and class reference
metaclass_field = tti_class.field_by_name('metaClass')
tti_class_obj = tti_obj.get_value(metaclass_field)
tti_class_class = tti_class_obj.reference_type

# get the variable names from the metaclass object
var_names_field = tti_class_class.field_by_name('variableNames')
var_names = tti_class_obj.get_value(var_names_field)

# splice the names and values together
table = var_names.values.zip(tti_vars.values)
puts table

This looks a bit complicated, but there's actually a lot of boilerplate here we could put into a utility class. For example, the metaClass and variableNames fields are standard on all (J)Ruby objects and classes, respectively. But considering that we're actually walking a remote VM's *live* heap...this is pretty simple code.

Here's what our script outputs now:
➔ jruby tti_digger.rb
"@offset"
instance of ruby.TZInfo.TimezoneOffsetInfo(id=25)
"@previous_offset"
instance of ruby.TZInfo.TimezoneOffsetInfo(id=26)
"@numerator_or_time"
instance of org.jruby.RubyFixnum(id=27)
"@denominator"
instance of org.jruby.RubyFixnum(id=28)
"@at"
instance of org.jruby.RubyNil(id=29)
"@local_end"
instance of org.jruby.RubyNil(id=29)
"@local_start"
instance of org.jruby.RubyNil(id=29)

We could go even deeper, but I think you get the idea.

Your Turn

Here's a gist of the three scripts we've created, so you can refer to and build off of them. And of course the javadocs and ri docs will help you as well, plus everything we've done here you can do in a jirb session.

There's a lot to the JDI API, but once you've got the VirtualMachine object in hand it's pretty easy to follow. As you'd expect from any debugger API, you need to know a bit about how things work on the inside, but through the magic of JRuby it's actually possible to write most of those fancy memory and debugging tools entirely in Ruby. Perhaps this article has peaked your interest in exploring JRuby internals using JDI and you might start to write debugging tools. Perhaps we can ship a few utilities to make some of the boilerplate go away. In any case, I hope this series of articles shows that JRuby users have an amazing library of tools available to them, and you don't even have to leave your comfort zone if you don't want to.

Note: The variableNames field is a recent addition to JRuby master, so if you'd like to play with that you'll probably want to build JRuby yourself or wait for a nightly build that picks it up. But you can certainly do a lot of exploring even without that patch.

Monday, July 12, 2010

Finding Leaks in Ruby Apps with Eclipse Memory Analyzer

After my post on Browsing Memory the JRuby Way, one commenter and several other folks suggested I actually show using Eclipse MAT with JRuby. So without further ado...

The Eclipse Memory Analyzer, like many Eclipse-based applications, starts up with a "for dummies" page linking to various actions.



The most interesting use of MAT is to analyze a heap dump in a bit more interactive way than with the "jhat" tool. The analysis supports the "jmap" dump format, so we'll proceed to get a jmap dump of a "leaky" Rails application.

I've added this controller to a simple application:

class LeakyController < ApplicationController
class MyData
def initialize(params)
@params = params
end
end

LEAKING_ARRAY = {}
def index
LEAKING_ARRAY[Time.now] = MyData.new(params)
render :text => "There are #{LEAKING_ARRAY.size} elements now!"
end
end


Some genius has decided to save all recent request parameters into a constant on the LeakyController, keyed by time, wrapped in a custom type, and never cleaned out. Perhaps this was done temporarily for debugging, or perhaps we have a moron on staff. Either way, we need to find this problem and fix it.

We'll run this application and crank 10000 requests through the /leaky index, so the final request should output "There are 10000 elements now!"

~ ➔ ab -n 10000 http://localhost:3000/leaky
This is ApacheBench, Version 2.3 <$Revision: 655654 $>
Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
Licensed to The Apache Software Foundation, http://www.apache.org/

Benchmarking localhost (be patient)
Completed 1000 requests
Completed 2000 requests
...


After 10000 requests have completed, we notice this application seems to grow and grow until it maxes out the heap (JRuby, being on the JVM, automatically limits heap sizes for you). Let's start by using jmap to investigate the problem.

~ ➔ jps -l
61976 org/jruby/Main
61999 sun.tools.jps.Jps
61837

~ ➔ jmap -histo 61976 | grep " ruby\." | head -5
37: 11685 280440 ruby.TZInfo.TimezoneTransitionInfo
40: 10000 240000 ruby.LeakyController.MyData
133: 970 23280 ruby.Gem.Version
137: 914 21936 ruby.Gem.Requirement
170: 592 14208 ruby.TZInfo.TimezoneOffsetInfo


We can see our old friend TimezoneTransitionInfo in there, but of course we've learned to accept that one. But what's this LeakyController::MyData object we've apparently got 10000 instances of? Where are they coming from? Who's holding on to them?

At this point, we can proceed to get a memory dump and move over to MAT, or have MAT acquire and open the dump in one shot, similar to VisualVM. Let's have MAT do it for us.

Getting Our Heap Into MAT

(Caveat: While preparing this post, I discovered that the jmap tool for the current OS X Java 6 (build 1.6.0_20-b02-279-10M3065) is not properly dumping all information. As a result, many fields and objects don't show up in dump analysis tools like MAT. Fortunately, there's a way out; on OS X, you can grab Soylatte or OpenJDK builds from various sources that work properly. In my case, I'm using a local build of OpenJDK 7.)

From the File menu, we select Acquire Heap Dump.



The resulting dialog should be familiar, since it lists the same JVM processes the "jps" command listed above. (If you had to specify a specific JDK home, like me, you'll need to click the "Configure" button and set the "jdkhome" flag" for "HPROF jmap dump provider".)



We'll pick our Rails instance (pid 61976) and proceed.

MAT connects to the process, pulls a heap dump to disk, and immediately proceeds to parse and open it.



Once it has completed parsing, we're presented with a few different paths to follow.



On other days, we might be interested in doing some component-by-component browsing to look for fat objects or minor leaks, or we might want to revisit the results of previous analyses against this heap. But today, we really need to figure out this MyData leak, so we'll run the Leak Suspects Report.

Leak Suspects?



Are you kidding? A tool that can search out and report possible leaks in a running system? Yes, Virginia, there is a Santa Claus!

This is the good side of the "plague of choices" we have on the JVM. Because there's so many tools for almost every basic purpose (like the dozen – at least – memory inspection tools), tool developers have moved on to more specific needs like leak analysis. MAT is my favorite tool for leak-hunting (and it uses less memory than jhat for heap-browsing, which is great for larger dumps).

Once MAT has finished chewing on our heap, it presents a pie chart of possible leak suspects. The logic used essentially seeks out data structures whose accumulated size is large in comparison to the rest of the heap. In this case, MAT has identified three suspects that in total comprise over half of the live heap data.



Scrolling down we start to get details about these leak candidates.



So there's a Hash, a Module, and 711 Class objects in our list of suspects. The Class objects are probably just loaded classes, since the JRuby core classes and additional classes loaded from Rails and its dependent libraries will easily number in the hundreds. We'll ignore that one for now. There's also an unusually large Module taking up almost 4MB of memory. We'll come back to that.

The Hash seems like the most likely candidate. Let's expand that.

The first new block of information gives us a list of "shortest paths" to the "accumulation point", or the point at which all this potentially-leaking data is gathering. There's more to this in the actual application, but I'm showing the top of the "path" here.



At the top of this list, we see the RubyHash object originally reported as a suspect, and a tree of objects that lead to it. In this case, we go from the Hash itself into a ConcurrentHashMap (note that we're hiding nothing here; you can literally browse anything in memory) which in turn is referenced by the "constants" field of a Class. So already we know that this hash is being referenced in some class's constant table. Pretty cool, eh?

Let's make sure we've got the right Hash and not some harmless data structure inside Rails. If we scroll down a bit more, we see a listing of all the objects this Hash has accumulated. Let's see what's in there.



Ok, so it's a hashtable structure with a table of entries. Can we get more out of this?

Of course like most of these tools, just about everything is clickable. We can dive into one of the hash entries and see what's in there. Clicking on an entry gives us several new ways to display the downstream objects we've managed to aggregate. In this case, we'll just do "List Objects", and the suboption "With Outgoing References" for downstream data.



Now finally in the resulting view of this particular RubyHashEntry, we can see that our MyData object is happily tucked away inside.



Ok, so we definitely have the right data structure. Not only that, but we can see that the entry's "key" is a Time object (org.jruby.RubyTime). Let's go back to the "Shortest Paths" view and examine the ConcurrentHashMap entry that's holding this Hash object. Each entry in this hash maps a constant name to a value, so we should be able to see which constant is holding the leaking references.

(At this point you'll see the side effects of my switch to OpenJDK 7; the memory addresses have changed, but the structure is the same.)



We'll do another "List Objects" "with outgoing references" against the the HashEntry object immediately referencing our RubyHash.



And there it is! In the "key" field of the HashEntry, we see our constant name "LEAKING_ARRAY".

What About That Module?

Oh yeah, what about that Module that showed up in the leak suspects? It was responsible for almost 4MB of the heap. Let's go back and check it out.



A-ha! Eclipse MAT has flagged the Gem module as being a potential leak suspect. But why? Let's go back to the suspect report and look at the Accumulated Objects by Class table, toward the bottom.



Ok, so the Gem module eventually references nearly 6000 Gem::Specification objects, which makes up the bulk of our 3.8MB. I guarantee I don't have 6000 gem versions installed. Perhaps that's something that RubyGems should endeavor to fix? Perhaps we've just used JRuby and Eclipse MAT to discover either a leak or wasteful memory use in RubyGems?

Evan Phoenix pointed out that I misread the columns. It's actually 249 Specification objects, their "self" size is almost 6000 bytes, and their "retained" size is 3.8MB. But that gives me an opportunity to show off another feature of MAT: Customized Retained Set calculation.

In this case, the retained size seems a bit suspect. Could there really be 3.8MB of data kept alive by Gem::Specification objects? It seems like a bit much, to be sure, but digging through the tree of references from the Gem module down shows there's several references to classes and modules, which in turn reference constant tables, method tables, and so on. How can we filter out that extra noise?

First we'll return to the view of the Gem module (two screenshots up) by going back to leak suspect #2, expanding "Shortest Paths". The topmost RubyModule in that list is the Gem module, so we're all set to calculate a Customized Retained Set.



The resulting dialog provides a list of options through which you can specify classes or fields to ignore when calculating the retained set from a given starting point. In this case, it's simple enough to filter out org.jruby.RubyClass and org.jruby.RubyModule, so that references from Gem::Specification back into the class/module hierarchy don't get included in calculations.



Which results in a similar view to those we've seen, but with objects sorted by retained heap.



Well what the heck? It looks like it's all String data?

JRuby's String implementation is an org.jruby.RubyString object, aggregating an org.jruby.util.ByteList object, aggregating a byte array, so the top three entries there in total are essentially all String memory. The best way to investigate where they're coming from is to do "List Objects" on RubyString, but instead of "with outgoing references" we'll use "with incoming references" to show where all those Strings are coming from.



Finally we have a view that lets us hunt through all these strings and see where they're coming from. Poking at the first few shows they're stored in constant tables of the Gem module (that last RubyModule I haven't expanded in). That's probably not a big deal. But if we sort the the list of RubyString objects by their retained sizes, we get a different picture of the system.



If we dig into the *largest* String objects, they appear to be referenced by Gem::Specification instance variables! So there's probably something worth investigating here.

It's also worth noting that any Ruby application is going to have a lot of Strings in it, so this isn't all that unusual to see. But it's nice to have a tool that lets you investigate potential inefficiencies (even down to the raw bytes!), and it's nice to know that at least some of that retained data for the Gem module is "real" and not just references back into the class hierarchy.

(And I'm not convinced all those Strings really *need* to be alive...but you're welcome to take it from here!)

Your Turn

Eclipse MAT is probably one of the nicest of the free tools. In addition to object browsing, leak detection, GC root analysis, and object query language support, there's a ton of other features, both in the main distribution and available from third parties. If you're hunting for memory leaks, or just want to investigate the memory usage of your (J)Ruby application, MAT is a tool worth playing with (and as always, I hope you will blog and report your experiences!)

Friday, July 09, 2010

Browsing Memory the JRuby Way

There's been a lot of fuss made lately over memory inspection and profiling tools for Ruby implementations. And it's not without reason; inspecting a Ruby application's memory profile, much less diagnosing problems, has traditionally been very difficult. At least, difficult if you don't use JRuby.


Because JRuby runs on the JVM, we benefit from the dozens of tools that have been written for the JVM. Among these tools are numerous memory inspection, profiling, and reporting tools, some built into the JDK itself. Want a heap dump? Check out the jmap (Java memory map) and jhat (Java heap analysis tool) shipped with Hotspot-based JVMs (Sun, OpenJDK). Looking for a bit more? There's the Memory Analysis Tool based on Eclipse, the YourKit memory and CPU profiling app, VisualVM, now also shipped with Hotspot JVMs...and many more. There's literally dozens of these tools, and they provide just about everything you can imagine for investigating memory.

In this post, I'll show how you can use two of these tools: VisualVM, a simple, graphical tool for exploring a running JVM; and the jmap/jhat combination, which allows you to dump the memory heap to disk for inspection offline.

Getting JRuby Prepared

All these tools work with any version of JRuby, but as part of JRuby 1.6 development I've been adding some enhancements. Specifically, I've made some modifications that allow Ruby objects to show up side-by-side with Java objects in memory profiles. A little explanation is in order.

In JRuby, all the core classes are represented by "native" Java classes. Object is represented by org.jruby.RubyObject, String is org.jruby.RubyString, and so on. Normally, if you extend one of the core classes, we don't actually create a new "native" class to represent it; instead, all user-created classes that extend Object simply show up as RubyObject in memory. This is still incredibly useful; you can look into RubyObject and see the metaClass field, which indicates the actual Ruby type.

Let's see what that looks like, so we know where we're starting from. We'll run a simple script that creates a custom class, instantiates and saves 10000 instances of it, and then sleeps.


~/projects/jruby ➔ cat foo_heap_example.rb 
class Foo
end

ary = []
10000.times { ary << Foo.new }

puts "ready for analysis!"
sleep

~/projects/jruby ➔ jruby foo_heap_example.rb
ready for analysis!


So we have our test subject ready to go. To use the jmap tool, we need the pid of this process. Of course we can use the usual shell tricks to get it, but the JDK comes with a nice tool for finding all JVM pids active on the system: jps

~/projects/jruby ➔ jps -l
52862 sun.tools.jps.Jps
52857 org/jruby/Main
48716 com.sun.enterprise.glassfish.bootstrap.ASMain


From this, you can see I have three JVMs running on my system right now: jps itself; our JRuby instance; and a GlassFish server I used for testing earlier today. We're interested in the JRuby instance, pid 52857. Let's see what jmap can do with that.

~/projects/jruby ➔ jmap
Usage:
jmap [option] <pid>
(to connect to running process)
jmap [option] <executable <core>
(to connect to a core file)
jmap [option] [server_id@]<remote server IP or hostname>
(to connect to remote debug server)

where <option> is one of:
<none> to print same info as Solaris pmap
-heap to print java heap summary
-histo[:live] to print histogram of java object heap; if the "live"
suboption is specified, only count live objects
-permstat to print permanent generation statistics
-finalizerinfo to print information on objects awaiting finalization
-dump:<dump-options> to dump java heap in hprof binary format
dump-options:
live dump only live objects; if not specified,
all objects in the heap are dumped.
format=b binary format
file=<file> dump heap to <file>
Example: jmap -dump:live,format=b,file=heap.bin <pid>
-F force. Use with -dump:<dump-options> <pid> or -histo
to force a heap dump or histogram when <pid> does not
respond. The "live" suboption is not supported
in this mode.
-h | -help to print this help message
-J<flag> to pass <flag> directly to the runtime system

<


The simplest option here is -histo, to print out a histogram of the objects on the heap. Let's run that against our JRuby instance.

~/projects/jruby ➔ jmap -histo:live 52857

num #instances #bytes class name
----------------------------------------------
1: 22677 3192816 <constMethodKlass>
2: 22677 1816952 <methodKlass>
3: 35089 1492992 <symbolKlass>
4: 2860 1389352 <instanceKlassKlass>
5: 2860 1193536 <constantPoolKlass>
6: 2798 739264 <constantPoolCacheKlass>
7: 5861 465408 [B
8: 5399 298120 [C
9: 3042 292032 java.lang.Class
10: 4037 261712 [S
11: 10002 240048 org.jruby.RubyObject
12: 3994 179928 [[I
13: 5474 131376 java.lang.String
14: 1661 95912 [I
...


The resulting output is a listing of literally every object in the system...not just Ruby objects even! The value of this should be apparent; not only can you start to investigate the memory overhead of code you've written, you'll also be able to investigate the memory overhead of every library and every piece of code running in the same process, right down to byte arrays (the "[B" above) and "native" Java strings ("java.lang.String" above). And so far we haven't had to do anything special to JRuby. Nice, eh?

So, back to the matter at hand: the Foo class from our example. Where is it?

Well, the answer is that it's right there; 10000 of those 10002 org.jruby.RubyObject instances are our Foo objects; the other two are probably objects constructed for JRuby runtime purposes. But obviously, there's nothing in this output that tells us how to find our Foo instances. This is what I'm remedying in JRuby 1.6.

On JRuby master, there's now a flag you can pass that will stand up a JVM class for every user-created Ruby class. Among the many benefits of doing this, we also get a more useful profile. Let's see how to use the flag (which will either be default or very easy to access by the time we release JRuby 1.6).

~/projects/jruby ➔ jruby -J-Djruby.reify.classes=true foo_heap_example.rb 
ready for analysis!


If we run jmap against this new instance, we see a more interesting result.

 num     #instances         #bytes  class name
----------------------------------------------
1: 22677 3192816 <constMethodKlass>
2: 22677 1816952 <methodKlass>
3: 35089 1492992 <symbolKlass>
4: 2860 1389352 <instanceKlassKlass>
5: 2860 1193536 <constantPoolKlass>
6: 2798 739264 <constantPoolCacheKlass>
7: 5863 465456 [B
8: 5401 298208 [C
9: 3042 292032 java.lang.Class
10: 4037 261712 [S
11: 10000 240000 ruby.Foo
12: 3994 179928 [[I
13: 5476 131424 java.lang.String
14: 1661 95912 [I


A-ha! There's our Foo instances! The "reify classes" option generates a JVM class of the same name as the Ruby class, prefixed by "ruby." to separate it from other JVM classes. Now we can start to see the real power of the tools, and we're just at the beginning. Let's see what a simple Rails application looks like.

~/projects/jruby ➔ jmap -histo:live 52926 | grep " ruby."
29: 11685 280440 ruby.TZInfo.TimezoneTransitionInfo
97: 970 23280 ruby.Gem.Version
98: 914 21936 ruby.Gem.Requirement
122: 592 14208 ruby.TZInfo.TimezoneOffsetInfo
138: 382 9168 ruby.Gem.Dependency
159: 265 6360 ruby.Gem.Specification
201: 142 3408 ruby.ActiveSupport.TimeZone
205: 118 2832 ruby.TZInfo.DataTimezoneInfo
206: 118 2832 ruby.TZInfo.DataTimezone
273: 41 984 ruby.Gem.Platform
383: 14 336 ruby.Mime.Type
403: 13 312 ruby.Set
467: 8 192 ruby.ActionController.MiddlewareStack.Middleware
476: 8 192 ruby.ActionView.Template
487: 7 168 ruby.ActionController.Routing.DividerSegment
508: 6 144 ruby.TZInfo.LinkedTimezoneInfo
523: 6 144 ruby.TZInfo.LinkedTimezone
810: 4 96 ruby.ActionController.Routing.DynamicSegment
2291: 2 48 ruby.ActionController.Routing.Route
2292: 2 48 ruby.I18n.Config
2293: 2 48 ruby.ActiveSupport.Deprecation.DeprecatedConstantProxy
2298: 2 48 ruby.ActionController.Routing.ControllerSegment
...


This time I've opted to grep out just the "ruby." items in the histogram, and the results are pretty impressive! We can see the baffling fact that there's 970 instance of Gem::Version, using at least 23280 bytes of memory. We can see the even more depressing fact that there's 11685 live instances of TZInfo::TimezoneTransitionInfo, using at least 280440 bytes.

Now that we're getting useful data, let's look at the first of our tools in more detail: jmap and jhat.

jmap and jhat

As you might guess, I do a lot of profiling in the process of developing JRuby. I've used probably a dozen different tools at different times. But the first tool I always reach for is the jmap/jhat combination.

You've seen the simple case of using jmap above, generating a histogram of the live heap. Let's take a look at an offline heap dump.

~/projects/jruby ➔ jmap -dump:live,format=b,file=heap.bin 52926
Dumping heap to /Users/headius/projects/jruby/heap.bin ...
Heap dump file created


That's how easy it is! The binary dump in heap.bin is supported by several tools: jhat (obviously), VisualVM, the Eclipse Memory Analysis Tool, and others. It's not officially a "standard" format, but it hasn't changed in a long time. Let's have a look at jhat options.

~/projects/jruby ➔ jhat
ERROR: No arguments supplied
Usage: jhat [-stack <bool>] [-refs <bool>] [-port <port>] [-baseline <file>] [-debug <int>] [-version] [-h|-help] <file>

-J<flag> Pass <flag> directly to the runtime system. For
example, -J-mx512m to use a maximum heap size of 512MB
-stack false: Turn off tracking object allocation call stack.
-refs false: Turn off tracking of references to objects
-port <port>: Set the port for the HTTP server. Defaults to 7000
-exclude <file>: Specify a file that lists data members that should
be excluded from the reachableFrom query.
-baseline <file>: Specify a baseline object dump. Objects in
both heap dumps with the same ID and same class will
be marked as not being "new".
-debug <int>: Set debug level.
0: No debug output
1: Debug hprof file parsing
2: Debug hprof file parsing, no server
-version Report version number
-h|-help Print this help and exit
<file> The file to read

For a dump file that contains multiple heap dumps,
you may specify which dump in the file
by appending "#<number>" to the file name, i.e. "foo.hprof#3".

All boolean options default to "true"


Generally you can just point jhat at a heap dump and away it goes. Occasionally if the heap is large, you may need to use the -J option to increase the maximum heap size of the JVM jhat runs in. Since we're running a Rails app, we'll bump the heap up a little bit.

~/projects/jruby ➔ jhat -J-Xmx200M heap.bin
Reading from heap.bin...
Dump file created Fri Jul 09 02:07:46 CDT 2010
Snapshot read, resolving...
Resolving 604115 objects...
[much verbose logging elided for brevity]

Chasing references, expect 120 dots........................................................................................................................
Eliminating duplicate references........................................................................................................................
Snapshot resolved.
Started HTTP server on port 7000
Server is ready.


"Server is ready"? Damn you Java people! Does everything have to be a server with you?

In this case, it's actually an incredibly useful tool. jhat starts up a small web application on port 7000 that allows you to click through the dump file. Let's see what that looks like.



Here's the front page of the tool. We see a listing of all JVM classes in the system. If you scroll to the bottom, there's a few more general functions.



Let's go with what we know and view the heap histogram again.



Here we can see that there's lots of objects taking up memory, and they're a mix of JVM-native types, JRuby implementation classes, and actual Ruby classes. In fact, here we can see our friend TZInfo::TimezoneTransitionInfo again. Let's click through.



Pretty mundane stuff so far; basically just information about the class itself. But you see at the bottom of this screenshot that we can go from here to viewing all instances of TimezoneTransitionInfo. Let's try that.



Ahh, that's more like it! Now we can see that there's a heck of a lot of these things floating around. Let's investigate a bit more and click through the first instance.



Now this is some cool stuff!

We can see that the JVM class generated for TimezoneTransitionInfo has three fields: metaClass, which points at the Ruby Class object; varTable, which is an array of Object references used for instance variables and other "internal" variables; and a flags field containing runtime flags for the object, like whether it's frozen, tainted, and so on. We can see that this object has no special flags set, and we can dig deeper into those fields if we like. We'll skip that today.

Moving further down, we see a few more amazing links. First, there's a list of all references to this object. Ahh, now we can start to investigate why they're staying in memory, even though we're not using them. We can even have jhat show us the full chains of references keeping these objects alive; a series of objects leading all the way back to one "rooted" by a thread or by global JVM state. And we can explore the other direction as well, walking all objects reachable from this one.

This is only a small part of what you can do with jmap and jhat, and they're so simple to use it feels almost criminal. But what if we want to inspect an application while it's running? Dumping heaps and analyzing them offline can tell you much of the story, but sometimes you just want to see the objects coming and going yourself. Let's move on to VisualVM.

VisualVM

VisualVM spawned out of the NetBeans profiling tools. One of the biggest complaints about the JVMs of old were that all the built-in tooling seemed to be designed for JVM engineers alone. Because Sun had the foresight to build and own their own IDE and related modules, it eventually became a natural fit to pull out the profiling tools for use by everyone. And so VisualVM was born.

On most systems with Java 6 installed, you should have a "jvisualvm" command. Let's run it now.



When you start up VisualVM, you're presented with a list of running JVMs, similar to using the 'jps' command. You can also connect to remote machines, browse offline heap and core dump files, and look through memory and CPU profiling snapshots from previous runs. Today, we'll just open up our running Rails app and see what we can see.



VisualVM connects to the running process and brings up a basic information pane with process information, JVM information, and so on. We're interested in monitoring heap usage, so let's move to the "Monitor" tab.



Already we're getting some useful information. This view shows CPU usage (currently zero, since it's an idle Rails app), Heap usage over time, and the number of JVM classes and threads that are active. We can trigger a full GC, if we'd like to tidy things up before we start poking around. But most importantly, we can do the jmap/jhat dance in one step, by clicking the Heap Dump button. Tantalizing, isn't it?



Initially, we see a basic summary of the heap: total size, number of classes and GC roots, and so on. We're looking for our friend TimezoneTransitionInfo, so let's look for it in the "Classes" pane.



Ahh, there it is, just a little ways down the list. The counts are as we expect, so let's double-click and dig a bit deeper.



Here we have a lot of the same information about object instances that we did with jhat, but presented in a much richer format. Almost everything is active; you can jump around the heap and do analysis that would take a lot of manual work very easily. Let's try another tool: the Retained Size calculator.

Because our JVM tools see all objects equally, the reported size for a Ruby object on the heap is only part of the story. There's also the variable table, the object's instance variables, and objects they reference to consider. Let's jump to a different object now, Gem::Version.

We don't want to have to scroll through the list of classes to find ruby.Gem.Version, so let's make use of the Object Query Language console. With the OQL console, you can write SQL-like queries to retrieve listings of objects in the heap. We'll search for all instances of ruby.Gem.Version.



The query runs and we get a listing of Gem::Version objects. Let's dig deeper and see how much retained memory each Version object is keeping alive.



Clicking on the "Compute Retained Sizes" link in the "Instances" pane prompts us with this dialog. We're tough...we can take it.



Reticulating splines...



So it looks like each of the Version objects take from 125 to 190 bytes for a total of 19400 bytes, most of which is from the variable table. What's in there?



Ahh...looks like there's a String and an Array. And of course we can poke around the heap ad infinatum, into and out of "native" JRuby and JVM classes, and truly get a complete picture of what our running applications look like. Now you're playing with power.

Your Turn

This is obviously only the tip of the iceberg. Tools like Eclipse Memory Analysis Tool include features for detecting leaks; VisualVM and NetBeans both allow you to turn on allocation tracing, to show where in your code all those objects are being created. There's tools for monitoring live GC behavior, and many of these tools even allow you to dig into a running heap and modify live objects. If you can dream it, there's a tool that can do it. And you get all that for free by using JRuby.

If you'd like to play with this, it all works with JRuby 1.5.1 but you won't get the nice JVM classes for Ruby classes. For that, you can pull and build JRuby master, download a 1.6.0.dev snapshot, or just wait for JRuby 1.6. And if you do play with these or other tools, I hope you'll let us know and blog about your experience!

In the future, I'll try to show some of the other tools plus some of the CPU profiling capabilities they bring to the table. For now, rest assured that if you're using JRuby, you really do have the best tools available to you.