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.

Indeed it was never our intention to say that it was our idea. I still remember the first time I loaded RB in VW30.... 2 s while normally loading
code was taking the time to cook pasta. I remember that I was still waiting but the code was already loaded. It was a cool feeling.
So I always wanted to experiment with that and one day mariano came and needed a fast loader and martin was working on ... a pickle format...
What a coincidence :)
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.
may be. For me if the use of IS is tructured (ie you control the fact that there will no pointer to the graph from elements that are not in the roots)
then you may have a stable system on reload else you will have to decide what to do on reload and this can be a real pain.

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.
Oh yes! This was what was also worrying to me.

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.

We do not have the resources for that now and will get probably less in the future because student cost doubled for internships :(


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).