New fast path for proxy loading in Voyage

On 08 May 2015, at 2:12 , Mariano Martinez Peck <[hidden email]> wrote:

On Fri, May 8, 2015 at 9:00 AM, Henrik Johansen <[hidden email]> wrote:

Hi Esteban!

As we talked about on IRC yesterday, I had hoped to employ the fast become discussed in 

http://forum.world.st/Igor-s-fast-become-for-CompiledMethods-in-Cog-td4345568.html

to bring faster proxy loading untill Spur arrives.

Sadly, it's not present in any current VM's, for reasons unknown*.

Luckily, the alternative proposed by Levente in the original thread (http://forum.world.st/A-trick-to-speedup-become-but-not-becomeForward-tp3707064p3708128.html) is appropriate for our case.

As I don't have write-access to Voyage repo, I've attached the package with the changes (which work in both 3.0 and 4.0, but I assume not older, due to using the SlotBuilder for anon subclasses)

hahahahah excellent. It was similar to the same idea I have for my proxies in Marea. In my case, what I did is to have 3 types of Proxies (each with each different object "format" -> normal, compact, long), and each of those proxy classes was a "variableSubclass". So then when I received the object to proxify, I would create an instance (the correct one from one of those 3 proxy classes) of the size of the object to proxify. Therefore, both, the proxy and the target would have same size and format, and therefore I could simply apply Igor idea.

Yeah, it's a really nice fast-path, when the VM's become implementation usually involves a heap scan...

In the case of Voyage, we're restricted in that we don't start out with a database ID, and not the actual original object, so perfectly pre-shaping the proxy is impossible.

Variably sized objects are right out, so I settled on the compromise that was easy to do; fixed size objects with equal or greater number of instvars as the proxy class.

To do fast swapping with objects of less than 3 slots, the current data held in proxy slots would have to be held somewhere external, which is doable, but ugly.

(You can also work around this easily if the performance hit is noticeable by adding dummy slots...)

Compact proxies is a really limited use case, since you never store any of the builtin ones in buckets directly, you'd have to manually make your mapped classes compact.

Seeing as how it's really just a holdover performance hack till Spur is ready for use and become: becomes fast, I don't think it makes sense implementing either in the context of Voyage. 

Cheers,

Henry

-- 
-----------------------------------------------------------------------
Objektfabrik Joachim Tuchel          [hidden email]
Fliederweg 1                         http://www.objektfabrik.de
D-71640 Ludwigsburg                  http://joachimtuchel.wordpress.com
Telefon: +49 7141 56 10 86 0         Fax: +49 7141 56 10 86 1

On 09 May 2015, at 07:54, [hidden email] wrote:

Hi guys,

I am fascinated by Voyage and was about to take a look at how much effort it might be to port it to VA Smalltalk for a project that doesn't have the option of moving away from it.

that’s cool, I never hear about people using it in other dialects :)

If I understand correctly, I can now give up on the idea, because Voyage will be using Slots from now on. Am I right?

well… eventually, that will happen, I suppose I will replace the magritte package with slots… but:

1) that will not happen any time soon (at least, just now in Pharo5 we are starting to create the tools to take advantage of slots… so I guess it will be ready “for general consumption” in pharo6… not that we cannot do some experimental stuff, and we will, but I do not thing there will be a replacement for magritte right now)

2) even if I do that, nothing forbids other dialect users to keep maintaining/using the magritte branch 

now… recent changes from Henrik are meant to take advantage not from slots but the fast-become made by Igor, a become who just works in

the case of two objects of same size (in that case, it just swap the objects and do not update the full system). This is hopefully a temporal change (while we wait for spur, who implements a general fast become)… I guess this can be rewritten without using the slot builder, but in any case this is too pharo specific to help you… nothing prevents you to do a “compatibility package” that overrides that method and allows you to continue using Voyage. 

Please let’s me know if you do it and you make your packages public for VA users. 

thanks, glad you find it useful :)

Esteban

Joachim

Am 08.05.15 um 15:14 schrieb Henrik Johansen:

On 08 May 2015, at 2:12 , Mariano Martinez Peck <[hidden email]> wrote:

On Fri, May 8, 2015 at 9:00 AM, Henrik Johansen <[hidden email]> wrote:

Hi Esteban!

As we talked about on IRC yesterday, I had hoped to employ the fast become discussed in 

http://forum.world.st/Igor-s-fast-become-for-CompiledMethods-in-Cog-td4345568.html

to bring faster proxy loading untill Spur arrives.

Sadly, it's not present in any current VM's, for reasons unknown*.

Luckily, the alternative proposed by Levente in the original thread (http://forum.world.st/A-trick-to-speedup-become-but-not-becomeForward-tp3707064p3708128.html) is appropriate for our case.

As I don't have write-access to Voyage repo, I've attached the package with the changes (which work in both 3.0 and 4.0, but I assume not older, due to using the SlotBuilder for anon subclasses)

hahahahah excellent. It was similar to the same idea I have for my proxies in Marea. In my case, what I did is to have 3 types of Proxies (each with each different object "format" -> normal, compact, long), and each of those proxy classes was a "variableSubclass". So then when I received the object to proxify, I would create an instance (the correct one from one of those 3 proxy classes) of the size of the object to proxify. Therefore, both, the proxy and the target would have same size and format, and therefore I could simply apply Igor idea.

Yeah, it's a really nice fast-path, when the VM's become implementation usually involves a heap scan...

In the case of Voyage, we're restricted in that we don't start out with a database ID, and not the actual original object, so perfectly pre-shaping the proxy is impossible.

Variably sized objects are right out, so I settled on the compromise that was easy to do; fixed size objects with equal or greater number of instvars as the proxy class.

To do fast swapping with objects of less than 3 slots, the current data held in proxy slots would have to be held somewhere external, which is doable, but ugly.

(You can also work around this easily if the performance hit is noticeable by adding dummy slots...)

Compact proxies is a really limited use case, since you never store any of the builtin ones in buckets directly, you'd have to manually make your mapped classes compact.

Seeing as how it's really just a holdover performance hack till Spur is ready for use and become: becomes fast, I don't think it makes sense implementing either in the context of Voyage. 

Cheers,

Henry

-- 
-----------------------------------------------------------------------
Objektfabrik Joachim Tuchel          [hidden email]
Fliederweg 1                         http://www.objektfabrik.de
D-71640 Ludwigsburg                  http://joachimtuchel.wordpress.com
Telefon: +49 7141 56 10 86 0         Fax: +49 7141 56 10 86 1

-- 
-----------------------------------------------------------------------
Objektfabrik Joachim Tuchel          [hidden email]
Fliederweg 1                         http://www.objektfabrik.de
D-71640 Ludwigsburg                  http://joachimtuchel.wordpress.com
Telefon: +49 7141 56 10 86 0         Fax: +49 7141 56 10 86 1