New Splitting strategy: running into assertion failure in newSplitting method

[Stefan Marr]

Hi Christian:

On 29 Jul 2014, at 13:36, Christian Humer <christian.humer at gmail.com> wrote:

> Regarding the warmup time -G:-TruffleSplittingClassInstanceStamps should bring it again to a reasonable level. I think I will push this as a default because I've seen class instance stamps to have no visible influence on peak performance but horrible influence on startup time. Can you confirm that?

Sorry for the late responds.
Well, it depends on the benchmark.

For TruffleSOM, I posted the relevant results here:
http://soft.vub.ac.be/~smarr/downloads/truffle/warmup.html

What you see is for each benchmark a graph that depicts how long one benchmark iteration takes over the course of 150 or 400 iterations in the same VM.

Since most benchmarks are comparably small, I would say the results for TruffleSOM are probably not very representative, but they show of instance in the microbenchmarks that class instance stamps seem to help recursive benchmarks: Fibonacci, Recurse. On the other hand, the Storage benchmark, a recursive, array-creating GC benchmark seems to be better off without it. The Queens benchmark also is much faster without class instance stamps.

For warmup time, the version without class instance stamps seems to be better (based on visual inspection of the graphs.

Best regards
Stefan 


> 
> Thanks.
> 
> 
> - Christian Humer
> 
> 
> On Thu, Jul 24, 2014 at 10:48 PM, Stefan Marr <java at stefan-marr.de> wrote:
> Hi Christian:
> 
> I updated to the latest version of Graal and tried the new splitting heuristic.
> Overall it seems to be a win for TruffleSOM. Warmup up time seems to increase significantly, and two benchmarks suffer, but overall results seem to be positive.
> 
> However, on all benchmarks I tested manually, I was running into an assertion error.
> (at com.oracle.graal.truffle.DefaultTruffleSplittingStrategyNew.newSplitting(DefaultTruffleSplittingStrategyNew.java:102))
> 
> It doesn’t seem to cause any visible issues, but the assumption of the assertion does not hold.
> 
> Best regards
> Stefan
> 
> 
> --
> Stefan Marr
> INRIA Lille - Nord Europe
> http://stefan-marr.de/research/
> 
> 
> 
> 

-- 
Stefan Marr
INRIA Lille - Nord Europe
http://stefan-marr.de/research/