(cs) performance regressions for 1.4.2_08b01

Hao Dong wrote:

I believe the footprint increase is caused by the new nio encoder/decoder introduced in 1.4.2_08. Those JIS0208 and JIS0212 De(En)coders are just big maps that occupy large amount of memory. To make it worse, 1.4.2_08 has two copies of those, both in sun.io and sun.nio package. PCK charset is using old sun.io converters and euc_JP is using sun.nio de/encoders. Before 1.4.2_08, there are only sun.io converters. And in Tiger, there are only sun.nio converters. Only in 1.4.2_08, it has the both.

 We did move euc_JP only to nio and leave pck in sun.io is because there is some difference in pck charset between 1.4.2 and 1.5. We wanted to be safe and did not touch pck converter (the escalatio is about euc_JP only).

 If the memory footprint regression is a BIG issue, we can modify the fix to eliminate one set of converters.

 Thanks,
 -Hao


Yumin Qi <###@###.###> wrote:


>> I tested footprint3 with following versions:
>> 142fcs, 142_04, 142_05, 142_06, 142_07, 142_08b01, result:
>>
>> ============================================================================
>> results/footresult.142
>> Benchmark Samples Mean Stdev Geomean Weight
>> footprint3 5 73040.74 0.00
>> ============================================================================
>> results/footresult.142_04
>> Benchmark Samples Mean Stdev %Diff P Significant
>> footprint3 5 73157.05 0.00 -0.16 0.000 Yes
>> ============================================================================
>> results/footresult.142_05
>> Benchmark Samples Mean Stdev %Diff P Significant
>> footprint3 5 73295.16 0.00 -0.35 0.000 Yes
>> ============================================================================
>> results/footresult.142_06
>> Benchmark Samples Mean Stdev %Diff P Significant
>> footprint3 5 73309.54 0.00 -0.37 0.000 Yes
>> ============================================================================
>> results/footresult.142_07
>> Benchmark Samples Mean Stdev %Diff P Significant
>> footprint3 5 73341.70 0.00 -0.41 0.000 Yes
>> ============================================================================
>> results/footresult.142_08
>> Benchmark Samples Mean Stdev %Diff P Significant
>> footprint3 5 73621.10 0.00 -0.79 0.000 Yes
>> ============================================================================
>> * - Not Significant: The %Diff for the mean could be noise. If there is
>> an actual difference, more samples would be needed to detect it.
>>
>>
>> It can be seen the regression is acumulative. _05 and _06 are at same level. There are jumps:
>> 1) from _04 to _05
>> 2) from _06 to _07
>> 3) from _07 to _08-b01 (_08 above is refered to _0b-b01)
>>
>> The cause from _07 fcs to _08-b01 is bug fix: 6196407
>> The fix affected Solaris and Linux on sparc and x86.
>> The new charset encoder/decoder code consume more time than before. This is an escalation: 1-4955711, and fixed in 1.4.2_08-b01. I dont think we can back this fix out, or we shoudl consult PDE of this fix to see if there is a better way to avoid this performance regression.
>>
>> Without the fix:
>>
>> ============================================================================
>> results/footresult.142_07
>> Benchmark Samples Mean Stdev Geomean Weight
>> footprint3 5 73341.70 0.00
>> ============================================================================
>> results/footresult.142_08_07charsetjar
>> Benchmark Samples Mean Stdev %Diff P Significant
>> footprint3 5 73339.33 0.00 0.00 0.000 Yes
>> ============================================================================

###@###.### 2005-2-28 23:32:53 GMT