[PATCH] Reduce Chance Of Mistakenly Early Backing Memory Cleanup

[Paul Sandoz]

Hi Ben,

Thanks for looking into this.


> On Feb 5, 2018, at 8:52 AM, Ben Walsh <ben_walsh at uk.ibm.com> wrote:
> 
> Running with the following test under the JMH benchmark framework (never 
> used this before, so please do point out any issues with the test) …
> 

Your benchmark looks good, i would return the byte buffer thereby avoiding any risk of dead code elimination (generally a best practice even if not absolutely required). As a follow on you might wanna try measuring a loop e.g.:

  ByteBuffer bb = ...
  for (int i = 0; i < L; i++) {
    bb.put((byte)i)
  }
  return bb;

where L is parameterized (see JMH’s @Param annotation), then you can easily vary from 1 upwards. Performance might be affected in loops if unrolling and/or vectorization is perturbed by the fence (i doubt the compiler will hoist the fence out of the loop given the JIT is forced to *not* inline it).

To test for vectorization you could write a test method to get int values from the buffer and sum ‘em up returning the sum.

Alas a 7% hit on simple access is not good :-( we really need the JIT to track the argument passed to the method.

Thanks,
Paul.

> 
> -------------------------------------------------------------------------------------------------------------------


> Result "org.sample.ByteBufferBenchmark.benchmark_byte_buffer_put":
>  33100911.857 ±(99.9%) 747461.951 ops/s [Average]
>  (min, avg, max) = (25373082.559, 33100911.857, 38885170.177), stdev = 
> 3164800.705
>  CI (99.9%): [32353449.906, 33848373.808] (assumes normal distribution)
> 
> 
> # Run complete. Total time: 00:27:27
> 
> Benchmark                                       Mode  Cnt         Score   
> Error  Units
> ByteBufferBenchmark.benchmark_byte_buffer_put  thrpt  200  33100911.857 ± 
> 747461.951  ops/s
> 
> 


> 
> Result "org.sample.ByteBufferBenchmark.benchmark_byte_buffer_put":
>  35604933.518 ±(99.9%) 654975.515 ops/s [Average]
>  (min, avg, max) = (25558172.378, 35604933.518, 39524804.534), stdev = 
> 2773207.341
>  CI (99.9%): [34949958.003, 36259909.033] (assumes normal distribution)
> 
> 
> # Run complete. Total time: 00:27:51
> 
> Benchmark                                       Mode  Cnt         Score   
> Error  Units
> ByteBufferBenchmark.benchmark_byte_buffer_put  thrpt  200  35604933.518 ± 
> 654975.515  ops/s
> 
> 
> ... So a performance degradation of roughly 7%.
> 
> 
> Regards,
> Ben Walsh
>