RFR: 8247532: Records deserialization is slow

Claes Redestad claes.redestad at oracle.com
Tue Jun 23 08:23:10 UTC 2020



On 2020-06-23 10:06, Claes Redestad wrote:
> Hi,
> 
> On 2020-06-23 09:49, Peter Levart wrote:
>> Hi Chris, Claes,
>>
>>
>> Ok then, here's with benchmark included:
>>
>>
>> http://cr.openjdk.java.net/~plevart/jdk-dev/RecordsDeserialization/webrev.07/ 
>>
>>
>>
>> I haven't been able to run the benchmark with "make test" though. I 
>> have tried various ways to pass javac options to build like:
>>
>>
>> make test TEST='micro:org.openjdk.bench.java.io.RecordDeserialization' 
>> TEST_OPTS="VM_OPTIONS=--enable-preview --release=16"
>>
>>
>> ...but javac doesn't seem to get them. Is there some secret option to 
>> achieve that?
> 
> Hmm, we might as well have the microbenchmarks build with
> --enable-preview on by default. Try this:

Fixed:

diff -r f2e1cd498381 make/test/BuildMicrobenchmark.gmk
--- a/make/test/BuildMicrobenchmark.gmk Tue Jun 23 10:08:35 2020 +0200
+++ b/make/test/BuildMicrobenchmark.gmk Tue Jun 23 10:33:17 2020 +0200
@@ -90,10 +90,11 @@
      TARGET_RELEASE := $(TARGET_RELEASE_NEWJDK_UPGRADED), \
      SMALL_JAVA := false, \
      CLASSPATH := $(MICROBENCHMARK_CLASSPATH), \
-    DISABLED_WARNINGS := processing rawtypes cast serial, \
+    DISABLED_WARNINGS := processing rawtypes cast serial preview, \
      SRC := $(MICROBENCHMARK_SRC), \
      BIN := $(MICROBENCHMARK_CLASSES), \
      JAVA_FLAGS := --add-modules jdk.unsupported --limit-modules 
java.management, \
+    JAVAC_FLAGS := --enable-preview, \
  ))

  $(BUILD_JDK_MICROBENCHMARK): $(JMH_COMPILE_JARS)

I verified this works with your micro, and doesn't seem to cause
any issues elsewhere:

  make test TEST=micro:RecordDeserialization

I can shepherd this as a separate fix for documentation purposes, but
feel free to include it in your patch and ping build-dev at ..

/Claes

> 
> diff -r 52741f85bf23 make/test/BuildMicrobenchmark.gmk
> --- a/make/test/BuildMicrobenchmark.gmk    Tue Jun 23 09:54:42 2020 +0200
> +++ b/make/test/BuildMicrobenchmark.gmk    Tue Jun 23 09:59:29 2020 +0200
> @@ -93,7 +93,7 @@
>       DISABLED_WARNINGS := processing rawtypes cast serial, \
>       SRC := $(MICROBENCHMARK_SRC), \
>       BIN := $(MICROBENCHMARK_CLASSES), \
> -    JAVA_FLAGS := --add-modules jdk.unsupported --limit-modules 
> java.management, \
> +    JAVA_FLAGS := --enable-preview --add-modules jdk.unsupported 
> --limit-modules java.management, \
>   ))
> 
>   $(BUILD_JDK_MICROBENCHMARK): $(JMH_COMPILE_JARS)
> 
>>
>>
>> Otherwise, I simulated what would happen when there are more then 10 
>> ObjectStreamClass layouts for same class rapidly interchanging, so 
>> that they push each other out of the cache, by temporarily setting 
>> cache's MAX_SIZE = 0. Note that this is worst case scenario:
>>
>>
>> Benchmark                                      (length)  Mode  Cnt    
>> Score    Error  Units
>> RecordDeserializationBench.deserializeClasses        10  avgt   10    
>> 9.393 ±  0.287  us/op
>> RecordDeserializationBench.deserializeClasses       100  avgt   10   
>> 35.642 ±  0.977  us/op
>> RecordDeserializationBench.deserializeClasses      1000  avgt   10  
>> 293.769 ±  7.321  us/op
>> RecordDeserializationBench.deserializeRecords        10  avgt   10   
>> 15.335 ±  0.496  us/op
>> RecordDeserializationBench.deserializeRecords       100  avgt   10  
>> 211.427 ± 11.908  us/op
>> RecordDeserializationBench.deserializeRecords      1000  avgt   10  
>> 990.398 ± 26.681  us/op
>>
>>
>> This is using JMH option '-gc true' to force GC after each iteration 
>> of benchmark. Without it, I get a big (~4s) full-GC pause just in the 
>> middle of a run with length=100:
>>
>>
>> Iteration   1: 528.577 us/op
>> Iteration   2: 580.404 us/op
>> Iteration   3: 4438.228 us/op
>> Iteration   4: 644.532 us/op
>> Iteration   5: 698.493 us/op
>> Iteration   6: 800.738 us/op
>> Iteration   7: 929.791 us/op
>> Iteration   8: 870.946 us/op
>> Iteration   9: 863.416 us/op
>> Iteration  10: 916.508 us/op
>>
>>
>> ...so results are a bit off because of that:
>>
>>
>> Benchmark                                      (length)  Mode  Cnt     
>> Score      Error  Units
>> RecordDeserializationBench.deserializeClasses        10  avgt   10     
>> 8.263 ±    0.043  us/op
>> RecordDeserializationBench.deserializeClasses       100  avgt   10    
>> 33.406 ±    0.160  us/op
>> RecordDeserializationBench.deserializeClasses      1000  avgt   10   
>> 287.595 ±    0.960  us/op
>> RecordDeserializationBench.deserializeRecords        10  avgt   10    
>> 15.270 ±    0.080  us/op
>> RecordDeserializationBench.deserializeRecords       100  avgt   10  
>> 1127.163 ± 1771.892  us/op
>> RecordDeserializationBench.deserializeRecords      1000  avgt   10  
>> 2003.235 ±  227.159  us/op
>>
>>
>> Yes, there is quite a bit of GCing going on when cache is thrashing. 
>> Note that I haven't tuned GC in any way and I'm running this on a 
>> machine with 64GiB of RAM so heap is allowed to grow quite big and G1 
>> is used by default I think.
>>
>>
>> This is still no worse than before the patch:
>>
>>
>> Benchmark                                 (length)  Mode  Cnt     
>> Score    Error  Units
>> RecordDeserialization.deserializeClasses        10  avgt   10     
>> 8.382 :  0.013  us/op
>> RecordDeserialization.deserializeClasses       100  avgt   10    
>> 33.736 :  0.171  us/op
>> RecordDeserialization.deserializeClasses      1000  avgt   10   
>> 271.224 :  0.953  us/op
>> RecordDeserialization.deserializeRecords        10  avgt   10    
>> 58.606 :  0.446  us/op
>> RecordDeserialization.deserializeRecords       100  avgt   10   
>> 530.044 :  1.752  us/op
>> RecordDeserialization.deserializeRecords      1000  avgt   10  
>> 5335.624 : 44.942  us/op
>>
>>
>> ... since caching of adapted method handle for multiple objects 
>> withing single stream is still effective.
>>
>> I doubt there will ever be more than 10 variants/versions of the same 
>> record class deserialized by the same VM and in rapid succession, so I 
>> think this should not be an issue. We could add a system property to 
>> control the MAX_SIZE of cache if you think it is needed.
> 
> Thanks for running the numbers on this, and I agree - it seems
> outlandishly improbable (most will only see one, and if you have to
> maintain 10+ different serialized shapes of some record you likely have 
> bigger problems).
> 
> I'd say let's keep it constant unless someone actually asks for it.
> 
> /Claes
> 
>>
>>
>> Regards, Peter
>>
>>
>> On 6/22/20 1:04 AM, Claes Redestad wrote:
>>> Hi Peter,
>>>
>>> patch and results look great!
>>>
>>> My only real comment on this is that I think the microbenchmark would be
>>> a valuable contribution, too.
>>>
>>> It'd also be interesting to explore how poor performance would become if
>>> we'd hit the (artificial) 11 layouts limit, e.g, by cycling through
>>> 10, 11, or 12 different shapes.
>>>
>>> /Claes
>>>
>>> On 2020-06-21 19:16, Peter Levart wrote:
>>>> Hi,
>>>>
>>>>
>>>> When re-running the benchmark [1] with different lengths of 
>>>> serialized arrays of records, I found that, compared to classical 
>>>> classes, lookup into the cache of adapted method handles starts to 
>>>> show when the length of array is larger (# of instances of same 
>>>> record type deserialized in single stream). Each record deserialized 
>>>> must lookup the method handle in a ConcurrentHashMap:
>>>>
>>>>
>>>> Benchmark                                    (length)  Mode Cnt 
>>>> Score   Error  Units
>>>> RecordSerializationBench.deserializeClasses        10  avgt 10 8.088 
>>>> ± 0.013  us/op
>>>> RecordSerializationBench.deserializeClasses       100  avgt 10 
>>>> 32.171 ± 0.324  us/op
>>>> RecordSerializationBench.deserializeClasses      1000  avgt 10 
>>>> 279.762 ± 3.072  us/op
>>>> RecordSerializationBench.deserializeRecords        10  avgt 10 9.011 
>>>> ± 0.027  us/op
>>>> RecordSerializationBench.deserializeRecords       100  avgt 10 
>>>> 33.206 ± 0.514  us/op
>>>> RecordSerializationBench.deserializeRecords      1000  avgt 10 
>>>> 325.137 ± 0.969  us/op
>>>>
>>>>
>>>> ...so keeping the correctly shaped adapted method handle in the 
>>>> per-serialization-session ObjectStreamClass instance [2] starts to 
>>>> make sense:
>>>>
>>>>
>>>> Benchmark                                    (length)  Mode Cnt 
>>>> Score   Error  Units
>>>> RecordSerializationBench.deserializeClasses        10  avgt 10 8.681 
>>>> ± 0.155  us/op
>>>> RecordSerializationBench.deserializeClasses       100  avgt 10 
>>>> 32.496 ± 0.087  us/op
>>>> RecordSerializationBench.deserializeClasses      1000  avgt 10 
>>>> 279.014 ± 1.189  us/op
>>>> RecordSerializationBench.deserializeRecords        10  avgt 10 8.537 
>>>> ± 0.032  us/op
>>>> RecordSerializationBench.deserializeRecords       100  avgt 10 
>>>> 31.451 ± 0.083  us/op
>>>> RecordSerializationBench.deserializeRecords      1000  avgt 10 
>>>> 250.854 ± 2.772  us/op
>>>>
>>>>
>>>> With that, more objects means advantage over classical classes 
>>>> instead of disadvantage.
>>>>
>>>>
>>>> [1] 
>>>> http://cr.openjdk.java.net/~plevart/jdk-dev/RecordsDeserialization/RecordSerializationBench.java 
>>>>
>>>>
>>>> [2] 
>>>> http://cr.openjdk.java.net/~plevart/jdk-dev/RecordsDeserialization/webrev.06/ 
>>>>
>>>>
>>>>
>>>> Regards, Peter
>>>>
>>>>


More information about the core-libs-dev mailing list