A tour on Spur for non-VM experts

A Tour to Spur for
Non-VM Experts
Guille Polito, Christophe Demarey
ESUG 2016, 22/08, Praha

From a user point of view
We are working on the new Pharo Kernel
● Bootstrap: create an image from scratch
- Classes
- Global objects
- Processes and contexts
● Image Initialization: What is the correct order?
BTW, see our talk on this ;)
Mission Pharo Kernel, Thursday 10 am

What is this talk about?
Dec 14, 2015; 11:08am
[IMPORTANT] Starting migration to Spur
VM

Initial Motivation to look into Spur
1) How do we move to Spur
as fast as possible?
2) Should we /
How do we adapt to it?
3) What are the risks?

Motivation of this talk #1: Education
Explain what is Spur
Determine if a problem comes
from image side or VM side?

Motivation of this talk #2:
Understanding the Impact
Is my application compatible?
Will It break? Do I have to port it?

Spur is not a new Virtual Machine
Its underlying execution engine is the same as in Cog
(same bytecode, same interpreter, same JIT compiler)

Spur is not a new Garbage Collector
It just implements a new garbage collector
(which, BTW, is not new...)

Spur is not a new Object Format.
It just implements a new object format
(which, BTW, is just the means to an end)

So... what is Spur?
Spur is a new Memory Manager for Cog VM.
- New object representation in memory
(that allows ephemerons, pinned objects,...)
- New memory organization of Pharo images
(that allows to better manage resources)

A tour on Spur for non-VM experts

Spur in a Nutshell
It's a Cog VM
+ 64 bits support
+ faster: x1.8 speedup
+ larger images (> 2 Go)
+ ephemeron support
and more ...

Spur > 64-bits support
● No more need to install 32-bits libraries
● Images with size > 2 Go

Spur > faster access to classes
Class Table
● Direct access to
class objects
0
1
2
3
4
5
6
7
8
Array (hash=0)
String (hash=6)
Next page...

Spur > faster
Garbage Collector
● “Young objects die young (and fast)”
● Added survivor segments (future and past) to
the young space
=> allows more minor GC instead of major GC
future
Old space Young Space
past
eden
Major GC (mark & sweep)
Minor GC
(scavenger)

Spur > Fast become
No more hangs in large images when using #become:
(e.g. Moose with a big famix model)
Why? Spur introduces forwarders
● prevents to scan the whole memory
● replaces pointers when they are accessed
● implemented by a partial read barrier1
Cheap in most cases (just one indirection)
Costly if you rely a lot on primitives fallback
1. Eliot Miranda, Clément Bera. A Partial Read Barrier for Efficient Support
of Live Object-Oriented Programming. ISMM'15

Spur > faster
Immediate objects
New immediate objects
● Character
● Float (only 64-bits)
Speed-up in wide strings
Speed-up in float arithmetic and memory saving

Spur > other features
Spur object format:
● All classes are compact
=> only two kind of headers (3 before Spur)
● Support for pinned-objects (see UFFI talk on Friday)
● Ongoing support of read-only objects
● Still 2 unused bits

Spur > scalability
● Memory is now divided in several segments
● No more need to have a contiguous chunk of
memory
Memory
segment
Memory
segment
Memory
segment
Memory
segment

Spur > reliability
● Ephemeron finalization support
● Avoid memory leaks
BTW, see our OTHER talk on this ;)
A Weak Pharo story, Thursday, 3 pm

Part 2:
Porting applications and
frameworks to Spur
Cog
Spur

How do I port
my application
to Spur?

Porting Applications
Okay, maybe just wait that your developer friends
port your favorite frameworks.

Porting Basics #1
The number hierarchy changed
● Beware if you have visitors
● Beware if you have overrides
Porting Frameworks/Libraries

Porting Basics #2
Character is now immediate
● Beware if you have overrides that use the
internal state

Porting Basics #3
New (enhanced) ephemeron finalization
● If you need finalization you'll probably want to
use the new one
BTW, see our OTHER talk on this ;)
A Weak Pharo story, Thursday, 3 pm

Porting Basics #4
Native Boost is being deprecated
● If you are using FFI, you will need to review your
bindings
Native Boost UFFI

VM development hosted on GitHub:
OpenSmalltalk / opensmalltalk-vm

Why is it a good news?
● Brings together the VM community
● Easier to contribute
– Pull requests
– Issue tracker
– Documentation:
https://github.com/OpenSmalltalk/opensmalltalk-
vm/blob/Cog/CONTRIBUTING.md

VM build all flavors through
Travis CI
Still missing VM tests. Upcoming?

Where to find VM binaries?
● Pharo
http://files.pharo.org/vm/
Squeak, NewSpeak
https://bintray.com/opensmalltalk/vm/cog

Should I move to Spur?
Bounce
BubbleSort
DeltaBlue
Fannkuch
GraphSearch
Json
Mandelbrot
Nbody
PageRank
Permute
Queens
QuickSort
Sieve
Storage
Towers
0
0,2
0,4
0,6
0,8
1
1,2
1,4
1,6
1,8
Pre-Spur
Spur
by Stefan Marr, Apr 06, 2015

Should I move to Spur?
✔ 64-bits support
✔ Increased performances: x1.8 speedup
✔ Scalability, Reliability and open to new
features
– image not compacting anymore (will be fixed soon)

1) Class tables
2) Forwarders
3) Ephemeron Finalization
4) The Scavenger GC
Dissecting Spur...

Chapter 1
Classes are in Tables
(and they hide in tables)

1.1 Compact classes
Smalltalk compactClassesArray

1.1 Cons of the old object header
● A full word is used to indicate an object's class
– 4G classes in 32 bits
– 16E (2^60) classes in 64 bits (!!)
● Three different headers
=> checks for the header type are common

1.2 Class table
0
1
2
3
4
5
6
7
8
Array (hash=0)
String (hash=6)
Next page...

1.2 Class table
class idx = 0
0
1
2
3
4
5
6
7
8
Array (hash=0)
String (hash=6)
Next page...

1.2 Pros of the new object header
● 2^22 classes (4M). Still enough and efficient.
● Compatible with 64bits
● All classes are compact => only two kind of headers

1.2 Hidden objects
0
1
2
3
4
5
6
7
8
Array (hash=0)
String (hash=6)
Next page...
Class idx = 1

1.2 Hidden objects?
● The class table is an object (and lives in the heap)
● Its class index is “hidden”:
– Array allInstances
will iterate objects by class index
● In the class table:
– Indexes 0-15 are reserved for tagged objects
– Indexes 16-32 are reserved for hidden classes

1.3 Maintaining the class table
Classes are normal objects...
They are created with no special primitives...
But...
How does the VM know an object is a
class to put it into the class table?

1.3 Identifying classes by definition
A class is an object that is instantiated:
A class enters the class table
upon instantiation

1.3 But the index is the hash!
But... hashes are assigned lazily for all objects:
Classes, on instance-side,
define a special hash method
Behavior >> basicIdentityHash
<primitive: 175>
self primitiveFailed
Object >> basicIdentityHash
<primitive: 75>
self primitiveFailed

Chapter 1 - Conclusions
● Classes are organized in tables
● All classes are compact
● Simpler object header
● Still place for 4M classes
● On the image side, is almost transparent

Chapter 2
The forwarder plague

2.1 Become
● Swaps two objects
– (actually, swaps two object's identity)
● Useful for:
– Updating and migrating objects
– Install proxies
– Replace an object's behavior

2.1 The old become
● Full scanned all memory
● And was SLOOOOOW

2.1 Lazy become
ba
Eliot Miranda, Clément Bera. A Partial Read Barrier for Efficient
Support of Live Object-Oriented Programming. ISMM'15

2.1 Lazy become
b becomeForward: a.
User's code
ba

2.1 Lazy become
b becomeForward: a.
User's code
a forwarder!

2.1 Lazy Become
b becomeForward: a.
b doSomething
User's code
a forwarder!

2.2 The read barrier
● A full read barrier would be too expensive
– (on every read, on every primitive, on every
message send...)
● The read barrier is implemented in two places:
– Message send lookup failure
– Primitive failure

2.2 Message send lookup failure
method := (self lookupSelector: selector inClass: class).
method ifNil: [
(receiver isForwarder) ifTrue: [
receiver := receiver forward.
“scan also the objects in the stack” ].
method := (self lookupSelector: selector inClass: class).
].

2.2 Primitive failure
self performPrimitive: primitiveNumber.
self primitiveFailed ifTrue: [
“scan the stack looking for forwarders and retry”
self performPrimitive: primitiveNumber.
].

2 Conclusions
● Become does not need full scan anymore
● A forwarder replaces the object in place
● Two-way become copies object at the end
● Forwarders are bypassed using a partial read barrier:
– Message lookup failure
– Primitive failure
● No noticeable overhead

3.5 Scavenger GC
● “Young Objects Die Young (and quick)”
● Young objects are created in eden
● Objects are “tenured” after surviving several
generations
● Tenured objects go to old space

3.5 Scavenger GC
Old space New Space
future past
eden

3.5 Scavenger GC
Old space New Space
future past
eden
● Mark and Sweep (marking collector)
● Runs “every blue moon” on the entire memory
● Slow
● Scavenger (copying collector)
● Runs often, only in new space
● Object tenure (to old space)
depends on the ratio of
allocation

3.5 Scavenger GC
New Space
future past eden
1) Future is always empty during execution

3.5 Scavenger GC
New Space
future past eden
2) On a GC, past and eden objects that are referenced are copied to future

3.5 Scavenger GC
New Space
past future eden
2) On a GC, past and eden objects that are referenced are copied to future
3) Then, future and past spaces are swapped

3.5 Scavenger GC
Two questions remain:
● How does the scavenger do a GC without
iterating the entire heap?
● How does he know object ages?

3.5 Scavenger GC
Two questions remain:
● How does the scavenger do a GC without
iterating the entire heap?
● How does he know object ages?
It maintains a set of “objects in new space referenced from old space”
By their addresses! Lower addresses are younger....

Is that all?
● Pinned objects?
● The finalization queue?
● Memory segments, bridges, …?
● (The not working) Memory compaction?
● New immediate objects?
● ...

A tour on Spur for non-VM experts

More Related Content

What's hot (20)

Viewers also liked (15)

Similar to A tour on Spur for non-VM experts (20)

More from ESUG (20)

Recently uploaded (20)

A tour on Spur for non-VM experts