Just a quick note I would like to share.... For my PhD, I did investigate ImageSegment very very deeply:

http://dl.acm.org/citation.cfm?id=2076323 http://www.slideshare.net/MarianoMartinezPeck/2010-smalltalkspeckobject-swap...

I didn't want to write Fuel just because. I took quite a lot of time to understand how ImageSegment primitives worked. From that effort, I remember a few conclusions:

1) I found only few users of ImageSegment 2) The few users I found, were NOT using the real purpose of ImageSegment, that is, object swapping. It was used instead as an object serializer. For that, they use #writeForExportOn: which ended up using SmartRefStream for the rest of the objects. 3) I noticed I could achieve the same performance or even better with an OO serializer built at the language side, with all the benefits this means. Of course, having Cog helped here....

In the Fuel paper: http://rmod.lille.inria.fr/archives/papers/Dias12a-SPE-Fuel.pdf you can find some benchmark comparison agains IS. Also in my PhD: http://rmod.lille.inria.fr/archives/phd/PhD-2012-Martinez-Peck.pdf

Cheers,

On Mon, Oct 20, 2014 at 9:56 PM, jvuletich@dc.uba.ar wrote:

Hi Eliot,

...

Hi All,

I want to check my understanding of reference semantics for image

segments as I'm close to completing the Spur implementation.

Specifically

...

the question is whether objects reachable only through weak pointers should be included in an image segment or not.

Remember that an image segment is created from the transitive closure of an Array of root objects, the *segment roots*. i.e. we can think of an image segment as a set of objects created by tracing the object graph from the segment roots.

The segment always includes the segment roots. Except for the roots, objects are excluded from the segment that are also reachable form the roots of the system (the *system roots*, effectively the root

environment,

...

Smalltalk, and the stack of the current process).

Consider a weak array in the transitive closure that is not reachable

from

...

the system roots, and hence should be included in the segment. Objects referenced from that weak array may be in one of three categories

• reachable from the system roots (and hence not to be included in the

segment)

• *not* reachable form the system roots, but reachable from the segment

roots via strong pointers (and hence to be included in the segment)

• *not* reachable form the system roots, *not* reachable from the segment

roots via strong pointers

Should this last category be included or excluded from the segment? I think that it makes no difference, and excluding them is only an optimization. The argument is as follows. Imagine that immediately

after

...

loading the image segment there is a garbage collection. That garbage collection will collect all the objects in the last category as they are only reachable from the weak arrays in the segment. Hence we are free to follow weak references as if they are strong when we create the image segment, leaving it to subsequent events to reclaim those objects.

An analogous argument accounts for objects reachable from ephemerons. Is my reasoning sound? -- best, Eliot

I think you are right. But there is a risk of somehow, someone, gaining a strong reference to the object after the image segment was created, breaking our invariants when the segment is loaded again.

An object might be (not reachable / strongly reachable / weakely reachable) from system roots and / or segment roots. This gives us 9 possibilities. Six of them are easy (and I'll not go into them). The other three are tricky:

a- Not reachable from system roots. Weakely reachable from segment roots. Do not include them. It is best to run a GC before building the image segment, to get rid of them (run termination, etc). This is to avoid the risk of the object gaining somehow a strong reference after the segment is built, making the segment miss the weak ref to it. Doing this way would also mean that any objects affected by termination would be consistent, both in the image and in the segment.

b- Weakely reachable from system roots. Weakely reachable from segment roots. Do not include them. If the object manages to survive by gaining a strong ref from the system roots, the weak ref will be repaired on segment load (Am I right on this?) If the original object was included in the segment, then on segment load it would point to a duplicate object that is about to be collected (and maybe terminated?) In any case, doing this way would also mean that any objects affected by termination would be consistent, both in the image and in the segment.

c- Weakely reachable from system roots. Strongly reachable from segment roots. Do include them. It seems reasonable to run a GC and get rid of them after unloading the segment, to avoid the risk of the object gaining somehow a strong ref in the image, and being duplicated on segment load. But doing as I say means that we would be loading into the image an object that was already terminated, although in the state it had before running termination. Not really sure if this is ok. There could be some risk of objects in the segment being in some pre-termination state, with some objects in the image being in some after-termination state. In any case, this would suggest bad design... So perhaps it makes sense to throw an exception in these cases?

I hope this rant is of use.

Cheers, Juan Vuletich

On Wed, Oct 22, 2014 at 1:42 AM, Casey Ransberger casey.obrien.r@gmail.com wrote:

Hi Mariano,

I've stripped some context here (it can be found by reading the original thread on squeak-dev) but I did want to respond to one or two things you said, so...

...

On Oct 20, 2014, at 6:55 PM, Mariano Martinez Peck <

marianopeck@gmail.com> wrote:

...

1. I found only few users of ImageSegment

If by users, you mean people (as opposed to something like senders or implementors,) did you count the people using EToys? I'm afraid my memory is telling me that it's used in saving and sharing projects (I'm not sure though.)

By users I mean code...senders... and yeah, Etoys Projects was one of them.

...

2. The few users I found, were NOT using the real purpose of

ImageSegment, that is, object swapping. It was used instead as an object serializer. For that, they use #writeForExportOn: which ended up using SmartRefStream for the rest of the objects.

Seems legit. Fair enough.

...

3. I noticed I could achieve the same performance or even better with an

OO serializer built at the language side, with all the benefits this means. Of course, having Cog helped here....

I suppose I'll be reading your paper then, won't I? :)

Cheers,

Casey

On Wed, Oct 22, 2014 at 2:05 AM, Bert Freudenberg bert@freudenbergs.de wrote:

On 20.10.2014, at 18:55, Mariano Martinez Peck marianopeck@gmail.com wrote:

...

The few users I found, were NOT using the real purpose of ImageSegment,

that is, object swapping. It was used instead as an object serializer. For that, they use #writeForExportOn: which ended up using SmartRefStream for the rest of the objects.

Well, if you look closer, you will see that projects use image segments in two completely different ways. One is, as you say, for serialization, which is not the best use of image segments, agreed, especially with all the other logic wrapped around it.

But if you enable projectsSentToDisk then entering a project will swap the previous project to disk as an image segment, allowing you to have images with very large projects without having to hold all in main memory at the same time.

This uses a completely different code path and file format than regular project export. The same technique could be used to swap out arbitrary chunks of an image.

Totally agree. So it seems we agree that the key and good part of ImageSegment is that one, swapping out, but not as a general object graph serializer.

On 22.10.2014, at 08:47, Craig Latta craig@netjam.org wrote:

...

So it seems we agree that the key and good part of ImageSegment is that one, swapping out, but not as a general object graph serializer.

Even for swapping out objects, I think putting them in their own

normal object memory is a better idea, since object memories can be minimal and small. This lets you perform more sophisticated reasoning about what to do when class formats change between swap-out and swap-in (as well as other meta-operations).

That's a different thing. Swapping is strictly about cutting up a single image into multiple segments. Proper mutation code would have to walk the whole object memory, meaning it needs to swap in all segments in turn. Which fortunately is extremely efficient, but likely still hairy enough that we're not actually using it given today's abundance of main memory. It might, however, still make a lot of sense e.g. as a deployment mechanism on mobile platforms, which still are severely memory-limited. If used strictly for deployment you don't have to worry about mutation.

- Bert -

Hi Stephane, Hi All,

let me talk a little about the ParcPlace experience, which led to David Leibs' parcels, whose architecture Fuel uses.

In the late 80's 90's Peter Deutsch write BOSS (Binary Object Storage System), a traditional interpretive pickling system defined by a little bytecoded language. Think of a bytecode as something like "What follows is an object definition, which is its id followed by size info followed by the definitions or ids of its sub-parts, including its class", or "What follows is the id of an already defined object". So the loading interpreter looks at the next byte in the stream and that tells it what to do. So the storage is a recursive definition of a graph, much like a recursive grammar for a programming language.

This approach is slow (its a bytecode interpreter) and fragile (structures in the process of being built aren't valid yet, imagine trying to take the hash of a Set that is only half-way through being materialized). But this architecture was very common at the time (I wrote something very similar). The advantage BOSS had was a clumsy hack for versioning. One could specify blocks that were supplied with the version and state of older objects, and these blocks could effect shape change etc to bring loaded instances up-to-date.

David Leibs has an epiphany as, in the early 90's, ParcPlae was trying to decompose the VW image (chainsaw was the code name of the VW 2.5 release). If one groups instances by class, one can instantiate in bulk, creating all the instances of a particular class in one go, followed by all the instances of a different class, etc. Then the arc information (the pointers to objects to be stored in the loaded objects inst vars) can follow the instance information. So now the file looks like header, names of classes that are referenced (not defined), definitions of classes, definitions of instances (essentially class id, count pairs), arc information. And materializing means finding the classes in the image, creating the classes in the file, creating the instances, stitching the graph together, and then performing any post-load actions (rehashing instances, etc).

Within months we merged with Digitalk (to form DarcPlace-Dodgytalk) and were introduced to TeamV's loading model which was very much like ImageSegments, being based on the VM's object format. Because an ImageSegment also has imports (references to classes and globals taken from the host system, not defined in the file) performance doesn't just depend on loading the segment into memorty. It also depends on how long it takes to search the system to find imports, etc. In practice we found that a) Parcels were 4 times faster than BOSS, and b) they were no slower than Digitalk's image segments. But being independent of the VM's heap format Parcels had BOSS's flexibility and could support shape change on load, something ImageSegments *cannot do*. I went on to extend parcels with support for shape change, plus support for partial loading of code, but I won't describe that here. Too detailed, even thought its very important.

Mariano spent time talking with me and Fuel's basic architecture is that of parcels, but reimplemented to be nicer, more flexible etc. But essentially Parcels and Fuel are at their core David Leibs' invention. He came up with the ideas of a) grouping objects by class and b) separating the arcs from the nodes.

Now, where ImageSegments are faster than Parcels is *not* loading. Our experience with VW vs TeamV showed us that. But they are faster in collecting the graph of objects to be included. ImageSegments are dead simple. So IMO the right architecture is to use Parcels' segregation, and Parcels' "abstract" format (independent of the heap object format) with ImageSegment's computation of the object graph. Igor Stasenko has suggested providing the tracing part of ImageSegments (Dan Ingalls' cool invention of mark the segment root objects, then mark the heap, leaving the objects to be stored unmarked in the shadow of the marked segment roots) as a separate primitive. Then this can be quickly partitioned by class and then written by Smalltalk code.

The loader can then materialize objects using Smalltalk code, can deal with shape change, and not be significantly slower than image segments. Crucially this means that one has a portable, long-lived object storage format; freeing the VM to evolve its object format without breaking image segments with every change to the object format.

I'd be happy to help people working on Fuel by providing that primitive for anyone who wants to try and reimplement the ImageSegment functonality (project saving, class faulting, etc) above Fuel.

On Wed, Oct 22, 2014 at 11:56 AM, Stéphane Ducasse < stephane.ducasse@inria.fr> wrote:

What I can tell you is that instability raised by just having one single pointer not in the root objects pointing to an element in the segment and the implication of this pointer on reloaded segments, (yes I do not want to have two objects in memory after loading) makes sure that we will not use IS primitive in Pharo in any future. For us this is a non feature.

IS was a nice trick but since having a pointer to an object is so cheap and the basis of our computational model so this is lead fo unpredictable side effects. We saw that when mariano worked during the first year of his PhD (which is a kind of LOOM revisit).

Stef

Hi David,

On Oct 22, 2014, at 5:52 PM, "David T. Lewis" lewis@mail.msen.com wrote:

Eliot,

Thanks for this background, it is very helpful and interesting.

I would also like to put in a good word for Fuel. It is well designed, well documented, and well supported on Squeak and Pharo. Very high quality work.

I use Fuel in RemoteTask (in package CommandShell) for inter-image communication. ReferenceStream also works, and both are supported in RemoteTask. But if you want to have a serializer that you can read and understand, I'd say that Fuel is hard to beat.

I am not advocating anything with respect to image segments, project saving, and so forth, I'm just saying that Fuel is a very good thing. It works well in Squeak, and I suspect that many folks may not be aware of this.

Oh I agree. If only ImageSegments weren't used... :-). We use an early version of Fuel at Cadence which is essential to our system. We haven't upgraded as it "just works".

Dave

On Wed, Oct 22, 2014 at 12:53:15PM -0700, Eliot Miranda wrote:

...

Hi Stephane, Hi All,

let me talk a little about the ParcPlace experience, which led to David Leibs' parcels, whose architecture Fuel uses.

In the late 80's 90's Peter Deutsch write BOSS (Binary Object Storage System), a traditional interpretive pickling system defined by a little bytecoded language. Think of a bytecode as something like "What follows is an object definition, which is its id followed by size info followed by the definitions or ids of its sub-parts, including its class", or "What follows is the id of an already defined object". So the loading interpreter looks at the next byte in the stream and that tells it what to do. So the storage is a recursive definition of a graph, much like a recursive grammar for a programming language.

This approach is slow (its a bytecode interpreter) and fragile (structures in the process of being built aren't valid yet, imagine trying to take the hash of a Set that is only half-way through being materialized). But this architecture was very common at the time (I wrote something very similar). The advantage BOSS had was a clumsy hack for versioning. One could specify blocks that were supplied with the version and state of older objects, and these blocks could effect shape change etc to bring loaded instances up-to-date.

David Leibs has an epiphany as, in the early 90's, ParcPlae was trying to decompose the VW image (chainsaw was the code name of the VW 2.5 release). If one groups instances by class, one can instantiate in bulk, creating all the instances of a particular class in one go, followed by all the instances of a different class, etc. Then the arc information (the pointers to objects to be stored in the loaded objects inst vars) can follow the instance information. So now the file looks like header, names of classes that are referenced (not defined), definitions of classes, definitions of instances (essentially class id, count pairs), arc information. And materializing means finding the classes in the image, creating the classes in the file, creating the instances, stitching the graph together, and then performing any post-load actions (rehashing instances, etc).

Within months we merged with Digitalk (to form DarcPlace-Dodgytalk) and were introduced to TeamV's loading model which was very much like ImageSegments, being based on the VM's object format. Because an ImageSegment also has imports (references to classes and globals taken from the host system, not defined in the file) performance doesn't just depend on loading the segment into memorty. It also depends on how long it takes to search the system to find imports, etc. In practice we found that a) Parcels were 4 times faster than BOSS, and b) they were no slower than Digitalk's image segments. But being independent of the VM's heap format Parcels had BOSS's flexibility and could support shape change on load, something ImageSegments *cannot do*. I went on to extend parcels with support for shape change, plus support for partial loading of code, but I won't describe that here. Too detailed, even thought its very important.

Mariano spent time talking with me and Fuel's basic architecture is that of parcels, but reimplemented to be nicer, more flexible etc. But essentially Parcels and Fuel are at their core David Leibs' invention. He came up with the ideas of a) grouping objects by class and b) separating the arcs from the nodes.

Now, where ImageSegments are faster than Parcels is *not* loading. Our experience with VW vs TeamV showed us that. But they are faster in collecting the graph of objects to be included. ImageSegments are dead simple. So IMO the right architecture is to use Parcels' segregation, and Parcels' "abstract" format (independent of the heap object format) with ImageSegment's computation of the object graph. Igor Stasenko has suggested providing the tracing part of ImageSegments (Dan Ingalls' cool invention of mark the segment root objects, then mark the heap, leaving the objects to be stored unmarked in the shadow of the marked segment roots) as a separate primitive. Then this can be quickly partitioned by class and then written by Smalltalk code.

The loader can then materialize objects using Smalltalk code, can deal with shape change, and not be significantly slower than image segments. Crucially this means that one has a portable, long-lived object storage format; freeing the VM to evolve its object format without breaking image segments with every change to the object format.

I'd be happy to help people working on Fuel by providing that primitive for anyone who wants to try and reimplement the ImageSegment functonality (project saving, class faulting, etc) above Fuel.

On Wed, Oct 22, 2014 at 11:56 AM, St??phane Ducasse < stephane.ducasse@inria.fr> wrote:

...

What I can tell you is that instability raised by just having one single pointer not in the root objects pointing to an element in the segment and the implication of this pointer on reloaded segments, (yes I do not want to have two objects in memory after loading) makes sure that we will not use IS primitive in Pharo in any future. For us this is a non feature.

IS was a nice trick but since having a pointer to an object is so cheap and the basis of our computational model so this is lead fo unpredictable side effects. We saw that when mariano worked during the first year of his PhD (which is a kind of LOOM revisit).

Stef

-- best, Eliot

If one groups instances by class, one can instantiate in bulk, creating all the instances of a particular class in one go

What does "instantiate in bulk" mean? Doesn't that mean one still must send #new (or #basicNew) to the class for each instance? Why would that be faster?

That is not a limitation of ImageSegments per se, but just how they are used in Etoys.

- Bert -

On 26.10.2014, at 05:30, karl ramberg karlramberg@gmail.com wrote:

One aspect with Etoys projects is that they can not extend the system. It works nicely if you just use Etoys tile scripting. But if you introduce a new class in a project, loading the project in a system that do not have that class will fail. So the use of project as a distribution system of applications will be limited to a certain version of images.

Karl

...

On Wed, Oct 22, 2014 at 9:53 PM, Eliot Miranda eliot.miranda@gmail.com wrote:

Hi Stephane, Hi All,

let me talk a little about the ParcPlace experience, which led to David Leibs' parcels, whose architecture Fuel uses.

In the late 80's 90's Peter Deutsch write BOSS (Binary Object Storage System), a traditional interpretive pickling system defined by a little bytecoded language. Think of a bytecode as something like "What follows is an object definition, which is its id followed by size info followed by the definitions or ids of its sub-parts, including its class", or "What follows is the id of an already defined object". So the loading interpreter looks at the next byte in the stream and that tells it what to do. So the storage is a recursive definition of a graph, much like a recursive grammar for a programming language.

This approach is slow (its a bytecode interpreter) and fragile (structures in the process of being built aren't valid yet, imagine trying to take the hash of a Set that is only half-way through being materialized). But this architecture was very common at the time (I wrote something very similar). The advantage BOSS had was a clumsy hack for versioning. One could specify blocks that were supplied with the version and state of older objects, and these blocks could effect shape change etc to bring loaded instances up-to-date.

David Leibs has an epiphany as, in the early 90's, ParcPlae was trying to decompose the VW image (chainsaw was the code name of the VW 2.5 release). If one groups instances by class, one can instantiate in bulk, creating all the instances of a particular class in one go, followed by all the instances of a different class, etc. Then the arc information (the pointers to objects to be stored in the loaded objects inst vars) can follow the instance information. So now the file looks like header, names of classes that are referenced (not defined), definitions of classes, definitions of instances (essentially class id, count pairs), arc information. And materializing means finding the classes in the image, creating the classes in the file, creating the instances, stitching the graph together, and then performing any post-load actions (rehashing instances, etc).

Within months we merged with Digitalk (to form DarcPlace-Dodgytalk) and were introduced to TeamV's loading model which was very much like ImageSegments, being based on the VM's object format. Because an ImageSegment also has imports (references to classes and globals taken from the host system, not defined in the file) performance doesn't just depend on loading the segment into memorty. It also depends on how long it takes to search the system to find imports, etc. In practice we found that a) Parcels were 4 times faster than BOSS, and b) they were no slower than Digitalk's image segments. But being independent of the VM's heap format Parcels had BOSS's flexibility and could support shape change on load, something ImageSegments *cannot do*. I went on to extend parcels with support for shape change, plus support for partial loading of code, but I won't describe that here. Too detailed, even thought its very important.

Mariano spent time talking with me and Fuel's basic architecture is that of parcels, but reimplemented to be nicer, more flexible etc. But essentially Parcels and Fuel are at their core David Leibs' invention. He came up with the ideas of a) grouping objects by class and b) separating the arcs from the nodes.

Now, where ImageSegments are faster than Parcels is *not* loading. Our experience with VW vs TeamV showed us that. But they are faster in collecting the graph of objects to be included. ImageSegments are dead simple. So IMO the right architecture is to use Parcels' segregation, and Parcels' "abstract" format (independent of the heap object format) with ImageSegment's computation of the object graph. Igor Stasenko has suggested providing the tracing part of ImageSegments (Dan Ingalls' cool invention of mark the segment root objects, then mark the heap, leaving the objects to be stored unmarked in the shadow of the marked segment roots) as a separate primitive. Then this can be quickly partitioned by class and then written by Smalltalk code.

The loader can then materialize objects using Smalltalk code, can deal with shape change, and not be significantly slower than image segments. Crucially this means that one has a portable, long-lived object storage format; freeing the VM to evolve its object format without breaking image segments with every change to the object format.

I'd be happy to help people working on Fuel by providing that primitive for anyone who wants to try and reimplement the ImageSegment functonality (project saving, class faulting, etc) above Fuel.

...

On Wed, Oct 22, 2014 at 11:56 AM, Stéphane Ducasse stephane.ducasse@inria.fr wrote: What I can tell you is that instability raised by just having one single pointer not in the root objects pointing to an element in the segment and the implication of this pointer on reloaded segments, (yes I do not want to have two objects in memory after loading) makes sure that we will not use IS primitive in Pharo in any future. For us this is a non feature.

IS was a nice trick but since having a pointer to an object is so cheap and the basis of our computational model so this is lead fo unpredictable side effects. We saw that when mariano worked during the first year of his PhD (which is a kind of LOOM revisit).

Stef

-- best, Eliot

On Oct 26, 2014, at 8:41 AM, Mariano Martinez Peck marianopeck@gmail.com wrote:

On Sun, Oct 26, 2014 at 12:35 PM, Bert Freudenberg bert@freudenbergs.de wrote:

...

That is not a limitation of ImageSegments per se, but just how they are used in Etoys.

I agree at some point. But what if the serializer were able to serialize classes as well? Fuel is able to serialize classes, traits, closures, compiled methods, etc. Of course there are scenarios when this becomes complicated, but for the average it works.

There's no restriction in the kinds if object ImageSegments can store either. Including classes, contexts, etc presents no problem (*). The only real difference is that ImageSegments are tied to the VM's object representation whereas systems like Fuel are portable. That's a key advantage.

(*) I realize that Cog's implementations are currently broken w.r.t. Contexts and should be fixed. Because of context-to-stack mapping the segment writer should be careful to store a married context as an in married context. Luckily this is just a small matter if programming ;-)

...

• Bert -

On 26.10.2014, at 05:30, karl ramberg karlramberg@gmail.com wrote:

...

One aspect with Etoys projects is that they can not extend the system. It works nicely if you just use Etoys tile scripting. But if you introduce a new class in a project, loading the project in a system that do not have that class will fail. So the use of project as a distribution system of applications will be limited to a certain version of images.

Karl

On Wed, Oct 22, 2014 at 9:53 PM, Eliot Miranda eliot.miranda@gmail.com wrote:

...

Hi Stephane, Hi All,

let me talk a little about the ParcPlace experience, which led to David Leibs' parcels, whose architecture Fuel uses.

In the late 80's 90's Peter Deutsch write BOSS (Binary Object Storage System), a traditional interpretive pickling system defined by a little bytecoded language. Think of a bytecode as something like "What follows is an object definition, which is its id followed by size info followed by the definitions or ids of its sub-parts, including its class", or "What follows is the id of an already defined object". So the loading interpreter looks at the next byte in the stream and that tells it what to do. So the storage is a recursive definition of a graph, much like a recursive grammar for a programming language.

This approach is slow (its a bytecode interpreter) and fragile (structures in the process of being built aren't valid yet, imagine trying to take the hash of a Set that is only half-way through being materialized). But this architecture was very common at the time (I wrote something very similar). The advantage BOSS had was a clumsy hack for versioning. One could specify blocks that were supplied with the version and state of older objects, and these blocks could effect shape change etc to bring loaded instances up-to-date.

David Leibs has an epiphany as, in the early 90's, ParcPlae was trying to decompose the VW image (chainsaw was the code name of the VW 2.5 release). If one groups instances by class, one can instantiate in bulk, creating all the instances of a particular class in one go, followed by all the instances of a different class, etc. Then the arc information (the pointers to objects to be stored in the loaded objects inst vars) can follow the instance information. So now the file looks like header, names of classes that are referenced (not defined), definitions of classes, definitions of instances (essentially class id, count pairs), arc information. And materializing means finding the classes in the image, creating the classes in the file, creating the instances, stitching the graph together, and then performing any post-load actions (rehashing instances, etc).

Within months we merged with Digitalk (to form DarcPlace-Dodgytalk) and were introduced to TeamV's loading model which was very much like ImageSegments, being based on the VM's object format. Because an ImageSegment also has imports (references to classes and globals taken from the host system, not defined in the file) performance doesn't just depend on loading the segment into memorty. It also depends on how long it takes to search the system to find imports, etc. In practice we found that a) Parcels were 4 times faster than BOSS, and b) they were no slower than Digitalk's image segments. But being independent of the VM's heap format Parcels had BOSS's flexibility and could support shape change on load, something ImageSegments *cannot do*. I went on to extend parcels with support for shape change, plus support for partial loading of code, but I won't describe that here. Too detailed, even thought its very important.

Mariano spent time talking with me and Fuel's basic architecture is that of parcels, but reimplemented to be nicer, more flexible etc. But essentially Parcels and Fuel are at their core David Leibs' invention. He came up with the ideas of a) grouping objects by class and b) separating the arcs from the nodes.

Now, where ImageSegments are faster than Parcels is *not* loading. Our experience with VW vs TeamV showed us that. But they are faster in collecting the graph of objects to be included. ImageSegments are dead simple. So IMO the right architecture is to use Parcels' segregation, and Parcels' "abstract" format (independent of the heap object format) with ImageSegment's computation of the object graph. Igor Stasenko has suggested providing the tracing part of ImageSegments (Dan Ingalls' cool invention of mark the segment root objects, then mark the heap, leaving the objects to be stored unmarked in the shadow of the marked segment roots) as a separate primitive. Then this can be quickly partitioned by class and then written by Smalltalk code.

The loader can then materialize objects using Smalltalk code, can deal with shape change, and not be significantly slower than image segments. Crucially this means that one has a portable, long-lived object storage format; freeing the VM to evolve its object format without breaking image segments with every change to the object format.

I'd be happy to help people working on Fuel by providing that primitive for anyone who wants to try and reimplement the ImageSegment functonality (project saving, class faulting, etc) above Fuel.

On Wed, Oct 22, 2014 at 11:56 AM, Stéphane Ducasse stephane.ducasse@inria.fr wrote:

...

What I can tell you is that instability raised by just having one single pointer not in the root objects pointing to an element in the segment and the implication of this pointer on reloaded segments, (yes I do not want to have two objects in memory after loading) makes sure that we will not use IS primitive in Pharo in any future. For us this is a non feature.

IS was a nice trick but since having a pointer to an object is so cheap and the basis of our computational model so this is lead fo unpredictable side effects. We saw that when mariano worked during the first year of his PhD (which is a kind of LOOM revisit).

Stef

-- best, Eliot

-- Mariano http://marianopeck.wordpress.com

On Tue, Oct 21, 2014 at 04:29:53PM -0700, Yoshiki Ohshima wrote:

I'm not fully following the discussion here, but I do remember seeing the following email from Dan in 1999:

http://lists.squeakfoundation.org/pipermail/squeak-dev/1999-March.txt

and search for: "From DanI at wdi.disney.com Fri Mar 26 07:17:09 1999"

It does not require two bits to mark.

(Hopefully this email has some relevance to the discussion at hand...

I don't know if it is directly relevant to the specific discussion, but it is definitely worth rereading this post from Dan. I remember being amazed at its simplicity when I first read it in fifteen years ago, and what stands out to in retrospect is that we seem to have largely overlooked what he apparently considered to be its main potential application: "Steps to Modularity - Incremental Snapshots"

Thanks for the pointer :-)

Dave

Here is a copy of the squeak-dev post from 1999:

From DanI at wdi.disney.com Fri Mar 26 07:17:09 1999

From: DanI at wdi.disney.com (Dan Ingalls) Date: Sat Jan 28 04:56:45 2012 Subject: Steps to Modularity - Incremental Snapshots Message-ID: <v0300780cb320dafbd7ab@[206.16.10.26]>

Folks -

A week or so ago, I sent out a message describing a technique for extracting segments from the Squeak image. What I want to know is,

Does anyone know of such a technique having being used previously?

I figure it must be known, but I have certainly never heard of it. Please reply directly to me.

To recap, here's how it works:

1. Mark the root (or roots) of the subtree desired.

2. Do a GC mark pass. SInce this stops at any marked objects, the subtree will be unmarked, "in the shadow" of its roots.

3. Copy the roots and the unmarked subtree into a byteArray (the image segment) Relocate internal pointers as you go Copy external pointers into a table of outpointers.

Reinstalling a segment is incredibly simple -- all you do is remap any pointers in one pass and throw away the byteArray header!

Thanks - Dan

PS... Ted and I have just completed an implementation and it is great. (It will be out in updates and release 2.4 within a week). It can trace and copy a 520kb tree of over 15000 objects in about 390ms. Used for deepCopy, it is about 20 times faster than what we do currently. Used to swap segments in and out, it finally offers a realistic vehicle for breaking down Squeak's monolithic images.

It's even faster than you would guess from the above. There is a fixed overhead for the full GC mark and unmark. This is 350ms on my machine (could surely be improved). It can then copy out the 520kb segment or reinstall it in about 40ms either way.