(discussion branch moved to Magma list).
Well, I doubt it's faster than just using Avi's regular WriteBarrier. Understand the overrides are only compiled once per class, not once per object. In the real world compilation is not a significant factor on performance.
The overridden methods follow the pattern:
someMethod | t1 t2 newT1 newT2 returnValue | t1 := self instVarAt: 1. t2 := self instVarAt: 2. returnValue := super someMethod. newT1==t1 ifFalse: [ self modified. ^ returnValue ]. newT2==t2 ifFalse: [ self modified. ^ returnValue ]. ^ returnValue
which is probably even faster than unwinding stack all the way to the Immutable handler every time an attribute is set. Not sure, but the above is fast enough to accomplish the overall goal and substantially improve app performance.
Ok, what if we call it a "fast write-barrier provided by the VM". Would that change your view? Igor?
in your case (Magma) you can do even better: at the beginning of transaction make a deep copy of object graph. Upon finishing transaction, you walk the modified graph and compare objects if they are changed or not. This could be somewhat memory-hungry (but do we care?), but you don't have to use anything like write barrier or overriding compiled methods.
And with having a per-object properties, it just another piece of cake:
Object>>maIsChanged ^ (self propertyAt: #originalCopy) = self
and for nested transactions, you can make it even more hungry
Object>>maIsChangedIn: transaction ^ (self propertyAt: transaction) = self
:) but nested (and not even) transactions implying anyways that you hold the original somewhere, to be able to roll back the changes.
So in case of rollback, all you need to do is to send #restore to the root object:
Object>>restore self befomeForward: (self propertyAt:#originalCopy). self allIvarsDo: [:each | each restore]
Now, add there a VM primitive which can quickly mark a graph starting from your roots, and stockpile marked objects into some array, and you will have a super fast thingy :)
On 12 June 2012 00:00, Chris Muller ma.chris.m@gmail.com wrote:
(discussion branch moved to Magma list).
Well, I doubt it's faster than just using Avi's regular WriteBarrier. Understand the overrides are only compiled once per class, not once per object. In the real world compilation is not a significant factor on performance.
The overridden methods follow the pattern:
someMethod | t1 t2 newT1 newT2 returnValue | t1 := self instVarAt: 1. t2 := self instVarAt: 2. returnValue := super someMethod. newT1==t1 ifFalse: [ self modified. ^ returnValue ]. newT2==t2 ifFalse: [ self modified. ^ returnValue ]. ^ returnValue
which is probably even faster than unwinding stack all the way to the Immutable handler every time an attribute is set. Not sure, but the above is fast enough to accomplish the overall goal and substantially improve app performance.
Ok, what if we call it a "fast write-barrier provided by the VM". Would that change your view? Igor?
Magma mailing list Magma@lists.squeakfoundation.org http://lists.squeakfoundation.org/mailman/listinfo/magma
On Tue, Jun 12, 2012 at 8:13 PM, Igor Stasenko siguctua@gmail.com wrote: [snip]
This could be somewhat memory-hungry (but do we care?), but you don't
[snip]
As long as the Windows VM is severely limited in how much memory it can handle, some of us care a lot. It just depends on how memory-hungry it is, and how much memory we're already using.
On 13 June 2012 18:13, Chris Cunningham cunningham.cb@gmail.com wrote:
On Tue, Jun 12, 2012 at 8:13 PM, Igor Stasenko siguctua@gmail.com wrote: [snip]
This could be somewhat memory-hungry (but do we care?), but you don't
[snip]
As long as the Windows VM is severely limited in how much memory it can handle, some of us care a lot. It just depends on how memory-hungry it is, and how much memory we're already using.
no, it just about keeping a copy of working set of your transaction. And nothing will save you from memory problems if you start to abuse DB to load huge blobs of data, required for your working set..
and as i said, you have to use the copy anyways in case if you need to rollback transaction.
and nobody said that you cannot employ a hybrid approach: you can keep a copies of objects up to certain size/or for certain kind of classes and for the large ones you can still use the same write barrier approach. so you will have a best of both worlds :)
In terms of implementation, this is quite easy to do: since graph comes from database in serialized form, creating an initial copy is just about materializing same object twice (or materializing and then copying) instead of once. One will be used in working set, and another will be used to detect the changes upon committing transaction.
In terms of implementation, this is quite easy to do: since graph comes from database in serialized form, creating an initial copy is just about materializing same object twice (or materializing and then copying) instead of once. One will be used in working set, and another will be used to detect the changes upon committing transaction.
Magma essentially does this today. Back in the early 2000's, it made shallow copies of each materialized object and identity-mapped them to the "working copy" of the object. It would compare them upon transaction commit to see if they should be part of the commit package.
Since then, it was changed to just keep the original buffers that were already part of the materialization (so not making copies when just reading from the db). On transaction commit it now compares the referenced oids of each to determine if different.
BTW: I just realized a BIG problem with the WriteBarrier approach to change-detection: If the method which modifies the inst-variable ALSO happens to perform the commit, then the commit will not notice the changes!
For example, consider a simple setter:
name: aString myDbSession commit: [ name = aString ]
WB will override this method in its generated anonymous subclass to:
name: aString | t1 | t1 := name. returnVal := super name: aString. t1 == name ifFalse: [ self writeBarrier modified ]. ^ returnVal
Do you see the problem? Super call already tried the commit before it even knew it was changed. It emphasizes the limitation of WB: That it can only work via method calls, not direct assignments like a immutability-bit could.
The only solution is to NEVER co-locate the updating of a persistent inst-var in the same method as performing/signaling the commit. That sucks bad enough that I've made Magma sessions now default allowWriteBarrier to false. :-(
On 15 June 2012 17:57, Chris Muller ma.chris.m@gmail.com wrote:
In terms of implementation, this is quite easy to do: since graph comes from database in serialized form, creating an initial copy is just about materializing same object twice (or materializing and then copying) instead of once. One will be used in working set, and another will be used to detect the changes upon committing transaction.
Magma essentially does this today. Back in the early 2000's, it made shallow copies of each materialized object and identity-mapped them to the "working copy" of the object. It would compare them upon transaction commit to see if they should be part of the commit package.
Ah.. cool.. so even less work. So, i presume you miss one little primitive which scans the memory subgraph from the roots which you provide and gives you back all objects which inside it. Like that you can very quickly discover what is a new shape of a working set and then iterate over it and see what objects were changed and which were added/removed.
Since then, it was changed to just keep the original buffers that were already part of the materialization (so not making copies when just reading from the db). On transaction commit it now compares the referenced oids of each to determine if different.
BTW: I just realized a BIG problem with the WriteBarrier approach to change-detection: If the method which modifies the inst-variable ALSO happens to perform the commit, then the commit will not notice the changes!
For example, consider a simple setter:
name: aString myDbSession commit: [ name = aString ]
WB will override this method in its generated anonymous subclass to:
name: aString | t1 | t1 := name. returnVal := super name: aString. t1 == name ifFalse: [ self writeBarrier modified ]. ^ returnVal
Do you see the problem?
yes, i do :) A correct implementation should modify the existing methods, but not use subclassing. But that's of course harder to do without decent support from tools (btw Opal Compiler will allow to do that easily).
but i think you can easily fix it by implementing #maValue in block and use aBlock maValue in commit method, instead of #value.
A sketch implementaiton of #maValue is something like:
maValue | rcvr receiverStateBefore result | rcvr := self home receiver. receiverStateBefore :=rcrvr copyState. result := self value. (rcvr isSameAs: receiverStateBefore) ifFalse: [ rcvr writeBarrier modified ]. ^ result
Super call already tried the commit before it even knew it was changed. It emphasizes the limitation of WB: That it can only work via method calls, not direct assignments like a immutability-bit could.
The only solution is to NEVER co-locate the updating of a persistent inst-var in the same method as performing/signaling the commit. That sucks bad enough that I've made Magma sessions now default allowWriteBarrier to false. :-(
magma@lists.squeakfoundation.org
-
Chris Cunningham -
Chris Muller -
Igor Stasenko