Re: Performance patch for DateFormatSymbols.getZoneIndex(String)

Tue, 24 Apr 2012 01:54:42 -0700 

Hi Charles,

Yes, it's jdk8-tl.

Thanks,
Masayoshi

On 4/23/2012 6:01 PM, Charles Lee wrote:

Hi Masayoshi,
I'd like to sponsor this patch. But I did not figure out which repos the patch should go to. Is it jdk8-tl repository? 

On 04/23/2012 04:54 PM, Masayoshi Okutsu wrote:

Looks good!

Masayoshi

On 4/23/2012 5:51 PM, Deven You wrote:

Hi Masayoshi,
Thanks for your reminder. I have updated the webrev[1]. Please review it.
[1] http://cr.openjdk.java.net/~littlee/OJDK-548/webrev.03 <http://cr.openjdk.java.net/%7Elittlee/OJDK-548/webrev.03> 

Thanks a lot!
On 04/23/2012 01:19 PM, Masayoshi Okutsu wrote:

Hi Deven,
As Yoshito noted, lastZoneIndex needs to be transient because DateFormatSymbols is serializable. Otherwise, the change looks good to me. 

Thanks,
Masayoshi

On 4/20/2012 11:56 AM, Deven You wrote:

Yoshito,
It is good to me for your suggestion. So Masayoshi, what is your opinion, is cache the last zone index better than cache recent zone ID and zone indexes, do you have any concern about using last zone index?
[1] The updated version of this patch http://cr.openjdk.java.net/~youdwei/DateFormatSymbol_perf/webrev.02/ <http://cr.openjdk.java.net/%7Eyoudwei/DateFormatSymbol_perf/webrev.02/> 

Thanks a lot!

On 04/11/2012 10:35 PM, Yoshito Umaoka wrote:

Deven,
If last time zone ID is cached along with the previously resolved index, and getZoneIndex does not check the contents of array, it might not work when zone strings are updated (setZoneStrings). So, to make this change work properly, you have to reset lastZoneID/zoneIndex in setZoneStrings(String[][]).
In my opinion, we just need to keep last zone index. And it should be 'transient'.
For every getZoneIndex(String) call, input ID is compared to zoneStrings[lastZoneIndex][0] - and if they matches, simply return lastIndex. Otherwise, do the linear search and update lastZoneIndex- like below.
private transient int lastZoneIndex = 0; // with initial value to be 0, you do not need to check lastZoneIndex >= 0 every time 

final int getZoneIndex(String ID) {
  // fast path
// note: if you want to get rid of lastZoneIndex < zoneStrings.length, // setZoneStrings(String[][]) should reset lastZoneIndex to 0 or any valid index.
if (lastZoneIndex < zoneStrings.length && ID.equals(zoneStrings[lastZoneIndex ][0]) { 
    return lastIndex;
  }

  // slow path
  for (int index = 0; index < zoneStrings.length; index++) {
    if (ID.equals(zoneStrings[index][0]) {
      lastZoneIndex = index;
      return index;
    }
  }
}

-Yoshito

On 4/11/2012 3:02 AM, Deven You wrote:
I think Yoshito's suggestion make sense, since the getZoneIndex is an internal method, if there is no manual setting to change the timezone, The timezone ID won't be changed for one instance of SimpleDateFormt.
I updated the patch webrev[1] for this suggestion and test the GetZoneIndexTest.java, the performance is almost the same with webrev.00. 

Please review the updated patch[1].

Thanks a lot!
[1] http://cr.openjdk.java.net/~littlee/OJDK-548/webrev.01 <http://cr.openjdk.java.net/%7Elittlee/OJDK-548/webrev.01> 

On 04/11/2012 04:00 AM, Yoshito Umaoka wrote:
I'm wondering if caching index matched last time (per DateFormatSymbols instance) is sufficient, instead of keeping multiple results in a hash map. I guess majority of use cases repeatedly refer the index of the zone used by SimpleDateFormat. 

-Yoshito

On 4/9/2012 10:46 AM, Masayoshi Okutsu wrote:

Hi Deven,

Here are my comments on the proposed changes.
- zoneIndexCache should be an instance field because WeakHashMap isn't thread-safe and the order of IDs in zoneStrings differs in each DateFormatSymbols.
- equals shouldn't be replaced with equalsIgnoreCase because time zone IDs are (supposed to be) case-sensitive. 

- Utilize auto-boxing/unboxing.

Thanks,
Masayoshi

On 4/9/2012 12:10 PM, Deven You wrote:

Hi i18n-devs,
The getZoneIndex() method is expensive and it's performance depends on which timezone you're in. This is because the code steps through a list of timezones until it finds the current one. This is happening over and over again. An idea would be to either cache it or rewrite the way we store time zone ids, such as a hashmap instead of an array. 
This patch[1] is for the cache option.
Applications which format/parse dates using SimpleDateFormat repeatedly may obtain benefits from this patch, especially when run in a timezone far down the zoneStrings array the improvements will be even bigger.
I have written a test case[2] which shows when a timezone which is in lower end of the zone strings array, the performance improvement this patch can get. 

test results:

no patch:
The total time is 742 ms!

with this patch[1] :
The total time is 508 ms!

If increase the loop times to 1000000, the results are:

no patch:
The total time is 4743 ms!

with this patch[1] :
The total time is 2126 ms!

The java version is:
/home/deven/hgrps/jdk8/build/linux-i586/j2sdk-image/bin/java -version 
openjdk version "1.8.0-internal"
OpenJDK Runtime Environment (build 1.8.0-internal-deven_2012_03_14_16_14-b00) 
OpenJDK Server VM (build 24.0-b02, mixed mode)
[1] http://cr.openjdk.java.net/~youdwei/DateFormatSymbol_perf/webrev.00/ <http://cr.openjdk.java.net/%7Eyoudwei/DateFormatSymbol_perf/webrev.00/> 
[2] GetZoneIndexTest.java





--
Best Regards,

Deven





--
Best Regards,

Deven





--
Best Regards,

Deven

Reply via email to