[LLVMdev] [Polly] Update of Polly compile-time performance on LLVM test-suite

Thu Aug 1 08:29:14 PDT 2013

On 07/31/2013 09:23 PM, Star Tan wrote:
> At 2013-07-31 22:50:57,"Tobias Grosser" <tobias at grosser.es
> <mailto:tobias at grosser.es>> wrote:
>
>>On 07/30/2013 10:03 AM, Star Tan wrote:
>>> Hi Tobias and all Polly developers,
>>>
>>> I have re-evaluated the Polly compile-time performance using newest
>>> LLVM/Polly source code.  You can view the results on
>>> http://188.40.87.11:8000
>>> <http://188.40.87.11:8000/db_default/v4/nts/16?compare_to=9&baseline=9&aggregation_fn=median>.
>>>
>>> Especially, I also evaluated ourr187102 patch file that avoids expensive
>>> failure string operations in normal execution. Specifically, I evaluated
>>> two cases for it:
>>>
>>> Polly-NoCodeGen: clang -O3 -load LLVMPolly.so -mllvm
>>> -polly-optimizer=none -mllvm -polly-code-generator=none
>>> http://188.40.87.11:8000/db_default/v4/nts/16?compare_to=9&baseline=9&aggregation_fn=median
>>> Polly-Opt: clang -O3 -load LLVMPolly.so -mllvm -polly
>>> http://188.40.87.11:8000/db_default/v4/nts/18?compare_to=11&baseline=11&aggregation_fn=median
>>>
>>> The "Polly-NoCodeGen" case is mainly used to compare the compile-time
>>> performance for the polly-detect pass. As shown in the results, our
>>> patch file could significantly reduce the compile-time overhead for some
>>> benchmarks such as tramp3dv4
>>> <http://188.40.87.11:8000/db_default/v4/nts/16/graph?test.355=2> (24.2%), simple_types_constant_folding
>>> <http://188.40.87.11:8000/db_default/v4/nts/16/graph?test.366=2>(12.6%),
>>> oggenc
>>> <http://188.40.87.11:8000/db_default/v4/nts/16/graph?test.331=2>(9.1%),
>>> loop_unroll
>>> <http://188.40.87.11:8000/db_default/v4/nts/16/graph?test.235=2>(7.8%)
>>
>>Very nice!
>>
>>Though I am surprised to also see performance regressions. They are all
>>in very shortly executing kernels, so they may very well be measuring
>>noice. Is this really the case?
>
> Yes, it seems that shortly executing benchmarks always show huge unexpected noise even we run 10 samples for a test.
>
> I have changed the ignore_small abs value to 0.05 from the original 0.01, which means benchmarks with the performance delta less then 0.05s would be skipped. In that case,the results seem to be much more stable.
> However, I have noticed that there are many other Polly patches between the two version r185399 and r187116. They may also affect the compile-time performance. I would re-evaluate LLVM-testsuite to see the performance improvements caused only by our

I doubt the Polly changes changed performance a much. However, there 
have been huge numbers of patches to LLVM/clang. Those obviously changed 
performance. The rerun test show that our results in fact filter noise 
out effectively. Can you check if this also holds for the original 0.01?

>>Also, it may be interesting to compare against the non-polly case to see
>>how much overhead there is still due to our scop detetion.
>>
>>> The "Polly-opt" case is used to compare the whole compile-time
>>> performance of Polly. Since our patch file mainly affects the
>>> Polly-Detect pass, it shows similar performance to "Polly-NoCodeGen". As
>>> shown in results, it reduces the compile-time overhead of some
>>> benchmarks such as tramp3dv4
>>> <http://188.40.87.11:8000/db_default/v4/nts/16/graph?test.355=2> (23.7%), simple_types_constant_folding
>>> <http://188.40.87.11:8000/db_default/v4/nts/16/graph?test.366=2>(12.9%),
>>> oggenc
>>> <http://188.40.87.11:8000/db_default/v4/nts/16/graph?test.331=2>(8.3%),
>>> loop_unroll
>>> <http://188.40.87.11:8000/db_default/v4/nts/16/graph?test.235=2>(7.5%)
>>>
>>> At last, I also evaluated the performance of the ScopBottomUp patch that
>>> changes the up-down scop detection into bottom-up scop detection.
>>> Results can be viewed by:
>>> pNoCodeGen-ScopBottomUp: clang -O3 -load LLVMPolly.so (v.s.
>>> LLVMPolly-ScopBottomUp.so)  -mllvm -polly-optimizer=none -mllvm
>>> -polly-code-generator=none
>>> http://188.40.87.11:8000/db_default/v4/nts/21?compare_to=16&baseline=16&aggregation_fn=median
>>> pOpt-ScopBottomUp: clang -O3 -load LLVMPolly.so (v.s.
>>> LLVMPolly-ScopBottomUp.so)  -mllvm -polly
>>> http://188.40.87.11:8000/db_default/v4/nts/19?compare_to=18&baseline=18&aggregation_fn=median
>>> (*Both of these results are based on LLVM r187116, which has included
>>> the r187102 patch file that we discussed above)
>>>
>>> Please notice that this patch file will lead to some errors in
>>> Polly-tests, so the data shown here can not be regards as confident
>>> results. For example, this patch can significantly reduce the
>>> compile-time overhead of SingleSource/Benchmarks/Shootout/nestedloop
>>> <http://188.40.87.11:8000/db_default/v4/nts/19/graph?test.17=2> only
>>> because it regards the nested loop as an invalid scop and skips all
>>> following transformations and optimizations. However, I evaluated it
>>> here to see its potential performance impact.  Based on the results
>>> shown on
>>> http://188.40.87.11:8000/db_default/v4/nts/21?compare_to=16&baseline=16&aggregation_fn=median,
>>> we can see detecting scops bottom-up may further reduce Polly
>>> compile-time by more than 10%.
>>
>>Interesting. For some reason it also regresses huffbench quite a bit.
>
> This is because the ScopBottomUp patch file invalids the scop detection for huffbench. The run-time of huffbench with different options are shown as follows:
>
> clang: 19.1680s  (see runid=14)
>
> polly without ScopBottomUp patch file: 14.8340s (see runid=16)
>
> polly with ScopBottomUp patch file: 19.2920s (see runid=21)
>
> As you can see, the ScopBottomUp patch file shows almost the same execution performance with clang. That is because no invalid scops is detected with this patch file at all.

I am still confused. So you are saying Polly reduces the run-time from 
19 to 14 secs for huffbench? This is nice, but very surprising for the 
no-codgen runs, no?

>>:-( I think here an up-to-date non-polly to polly comparision would come
>>handy to see which benchmarks we still see larger performance
>>regressions. And if the bottom-up scop detection actually helps here.
>>As this is a larger patch, we should really have a need for it before
>>switching to it.
>>
> I have evaluated Polly compile-time performance for the following options:
>
>    clang: clang -O3  (runid: 14)
>
>    pBasic: clang -O3 -load LLVMPolly.so (runid:15)
>
>    pNoGen: pollycc -O3 -mllvm -polly-optimizer=none -mllvm -polly-code-generator=none (runid:16)
>
>    pNoOpt: pollycc -O3 -mllvm -polly-optimizer=none (runid:17)
>
>    pOpt: pollycc -O3 (runid:18)
>
> For example, you can view the comparison between "clang" and "pNoGen" with:
>
> http://188.40.87.11:8000/db_default/v4/nts/16?compare_to=14&baseline=14
>
> It shows that without optimizer and code generator, Polly would lead to less then 30% extra compile-time overhead.

This is a step in the right direction, especially as most runs show a 
lot less overhead.

Still, we need to improve on this. Ideally, we should not see more than 
5% slowdown. I suspect we can get some general speed-ups by reducing the 
set of passes we schedule for canonicalization.  However, before, it may 
be good to look at some of the slow kernels. lemon e.g. looks 
interesting - 20% slowdown.

> For the execution performance, it is interesting that pNoGen not only significantly improves the execution performance for some benchmarks (nestedloop/huffbench) but also significantly reduces the execution performance for another set of benchmarks (gcc-loops/lpbench).

Yes, that is really interesting. I suspect a couple of our 
canonicalization passes enabled/blocked additional optimizations in 
LLVM. The huffbench kernel seems especially interesting. This is not 
your number one priority in GSoC, but understanding why the gcc-loops 
got so much worse may be interesting. I suspect this may some kind of 
generic LLVM issue we expose and we should report a bug explaining the 
issue.

Tobi