Re: [Sablevm-developer] branch permissions and tag creation

[Chris P. <chr...@ma...>]

Etienne Gagnon wrote:
> Hi,
> 
> Grzegorz B. Prokopski wrote:
> 
>>Secondly, has it been agreed that this stuff is to be merged into
>>staging? I'd like to hear from Etienne before we merge.
>>(honestly, I wish I could see the diff before asking ;-)
> 
> 
> I'd like to see a comparison of this proposal with the elimination of 
> fat locks counter proposal, on performance, before we put this code (or 
> the counter proposal) in staging.

I guess we need some benchmarks before doing that (JGF?), and also a 
working fat lock elimination implementation.

For me this change is interesting mostly from a basic research 
perspective, just because it brings your algorithm a bit closer to 
Onodera's.  In my experience, it does improve performance, but not all 
of the time (sometimes it decreases performance), and the gains are not 
terribly significant.  Processor configuration has an impact, and 
presumably so does kernel version.

The other (more involved) change that would make them really comparable 
would be to implement fat lock deflation.  I expect that deflation would 
also enhance the performance benefit of this tiny "full inflation" 
change; obviously if you inflate locks more aggressively, while this 
saves you from taking the expensive inflation path some of the time, you 
do end up with more fat locks, and so you only benefit if these locks 
are under a lot of contention.

If you only want to commit this once deflation is implemented (assuming 
that happens), that's fine with me.  As it stands, it's a pretty benign 
patch, and I don't really see any problem with it (otherwise I would not 
have suggested it for staging).  I guess I'm in favour of multiple 
different algorithms being selectable at compile-time in SableVM, just 
because I think that's neat and because it shows off the suitability of 
the VM for doing research.

I do think that a paper comparing

1) Onodera's original algorithm (if SableVM can be modified to use it)
2) your original algorithm
3) your original algorithm with full lock inflation + deflation (to 
bring it closer to Onodera's)
4) this new "fat lock elimination" algorithm (Greg talked to me briefly 
about it, but I'm not clear on the details)

would be interesting.

Cheers,
Chris