Subj : Re: STL Containers w/ Virtually Zero-Overhead Lock-Free Reads...
To   : (Usenet)
From : Chris Thomasson
Date : Wed Sep 21 2005 06:13 am

>> Perhaps I should create a demo virtually
>> zero-overhead container class...
>
> Please do, it would be a perfect feasibility study.
>
> But be careful to implement the existing interface exactly and
> completely, especially the guarantees about iterators lifetime and
> validity - this will be needed to actually apply some algorithms to

IMHO, this is the most important issue of all.

I know virtually zero-overhead STL is possible; however, there are a couple 
of subtitle “features” that come along with “state-of-the-art solutions” to 
the classic reader/writer problem.

Unfortunately, these “features” are not very forgiving wrt algorithms that 
incorporate object membership in a particular collection into their standard 
logic.

Basically, objects can legitimately exist in a collection “after other 
threads” have atomically removed them.

It is safe, and “normal” for iterating threads to access a “recently 
 removed” object, simply because the objects memory is being meticulously 
managed by the “collector”.




>> Would anybody be interested in such a thing?
>
> Certainly.

Okay. I thought that there might be a little interest. I am currently 
polishing a proof-of-concept library right now. However, I am very busy 
incorporating it parts of it into some in-house server software. I will have 
the time to provide the demo library in the near future (most likely before 
next year). Humm... Perhaps a “new collection and iteration model” should be 
created, and the STL could stay the same. Something like

std::fast_map< ... >  ect.

I believe something has to be done in C++ that can address the 
high-performance computers that are becoming available to everyday 
customers...


Thank you for your comments.




P.S.

When I use the term “features”, I am referring to an inherent side-effect 
that is common to algorithms that allow writes, and memory barrier / atomic 
operation free reads to execute in parallel. This type of behavior is key to 
scalability wrt SMP, NUMA, ect... BTW, there is an odd “solution” to the 
iteration problem; however, I am really not allowed to post any information. 
;(... 



      [ See http://www.gotw.ca/resources/clcm.htm for info about ]
      [ comp.lang.c++.moderated.    First time posters: Do this! ]

.