[threeten-dev] Ser performance

[Xueming Shen]

On 3/8/13 7:25 AM, roger riggs wrote:
> Hi Sherman,
>
> I would also be interested in the comparison with the serialization of 
> java.util.Date
> though it is likely to be similar to LocalDate since it just 
> serializes a long.
>
> Does Regex have an patterns for which it has optimized implementations?
> It should be possible to normalize the patterns and with some data 
> gathering
> code up some specific implementations.
>
> Regex has always seemed like overkill to me for most patterns.
>
> The code for checkname could be optimized a bit as Stephen suggested
> and could also put the individual comparisons in numeric order and
> if the character was > making a quick continue.

I will try the suggested one, but I remember, if my memory servers me
correctly, the current code is actually faster. I remember I tried the
similar "optimization" years ago on the charset version, but it ends up
the existing one is faster. But my memory might be wrong and the
hotspot might be different:-)



>
> Either way its a good improvement.
>
> Thanks, Roger
>
>
> On 3/8/2013 12:10 AM, Xueming Shen wrote:
>> I was asked to compare the ZDT and GregorianCalendar the other day, here
>> are the data.
>>
>> --------LocalDate--------
>> serSize=16
>> Writing: 6655
>> Reading: 3318
>> --------LocalTime--------
>> serSize=17
>> Writing: 6335
>> Reading: 2968
>> --------LocalDateTime--------
>> serSize=23
>> Writing: 6317
>> Reading: 3392
>> --------ZonedDateTime--------
>> serSize=46
>> Writing: 7731
>> Reading: 15841
>> --------GregorianCal--------
>> serSize=266
>> Writing: 28491
>> Reading: 43594
>> --------ZoneId.of()--------
>> Reading: 4500
>>
>> Obviously JSR310 ZDT is much faster than GregorianCalendar, both
>> serialization and de-ser. with a much smaller size, as expected.
>>
>> However I noticed that the ZDT reading/de-serialization time is kinda 
>> "unproportional"
>> bigger, given it only has a LDT, a ZoneOffset and ZoneId to read in. 
>> It appears we "spend"
>> most of the ZoneRegion.ofId() time on the name validation via the 
>> regex. the regex's
>> matches() is too slow for this "simple" validation, unfortunately. 
>> Though I hate to propose,
>> as the proud maintainer of jdk regex:-), but I think it might worth 
>> considering the simple
>> char by char direct match, as we did in java.nio.Charset for the 
>> similar name validation
>> situation, as showed in the following webrev.
>>
>> http://cr.openjdk.java.net/~sherman/jdk8_threeten/zidPerm/
>>
>> It boost the ZoneId.of() speed from 4500 to 1261, and cut the ZDT 
>> de-serialization
>> time almost in half.
>> (test case: 
>> http://cr.openjdk.java.net/~sherman/jdk8_threeten/zidPerm/SerTest.java)
>>
>> --------LocalDate--------
>> serSize=16
>> Writing: 6741
>> Reading: 3328
>> --------LocalTime--------
>> serSize=17
>> Writing: 6237
>> Reading: 3000
>> --------LocalDateTime--------
>> serSize=23
>> Writing: 6290
>> Reading: 3412
>> --------ZonedDateTime--------
>> serSize=46
>> Writing: 7566
>> Reading: 8054
>> --------GregorianCal--------
>> serSize=266
>> Writing: 29517
>> Reading: 42369
>> --------ZoneId.of()--------
>> Reading: 1261
>>
>> -Sherman
>>
>