Collation: Sun JDK 1.6 vs ICU4J 3.8
The performance test takes a locale and creates a RuleBasedCollator with
default options. A large list of names is used as data in each test, where the
names vary according to language. Each Collation operation over the whole list
is repeated 1000 times. The percentage values in the final column are the most
useful. They measure differences, where positive is better for ICU4J, and
negative is better for the compared implementation.
Key
| Operation |
Units |
Description |
| strcoll |
nanosecs |
Timing for string collation, an incremental compare of strings. |
| keygen |
nanosecs |
Timing for generation of sort keys, used to 'precompile' information so
that subsequent operations can use binary comparison. |
| keylen |
bytes/char |
The average length of the generated sort keys, in bytes per character
(Unicode/ISO 10646 code point). Generally this is the important field for sort
key performance, since it directly impacts the time necessary for binary
comparison, and the overhead of memory usage and retrieval time for sort
keys. |
Data
-------- ICU -------- ------ JAVA ------- (JAVA - ICU)/ICU
Locale Data file strcoll keygen keylen strcoll keygen keylen coll keygen keylen
------------------------------------------------------------------------------------------------------------
en_US TestNames_Latin.txt | 176 3940 1.59 | 2602 8003 5.92 | 1378% 103% 272%
da_DK TestNames_Latin.txt | 676 5629 1.59 | 7700 19744 5.91 | 1039% 251% 272%
de_DE TestNames_Latin.txt | 178 3951 1.59 | 2610 8013 5.92 | 1366% 103% 272%
de__PHONEB TestNames_Latin.txt | 189 5016 1.59 | 2635 8012 5.92 | 1294% 60% 272%
fr_FR TestNames_Latin.txt | 177 4111 1.59 | 2616 8826 5.92 | 1378% 115% 272%
ja_JP TestNames_Latin.txt | 176 3981 1.59 | 2918 8274 5.92 | 1558% 108% 272%
ja_JP TestNames_Japanese_h.txt | 486 2744 3.32 | 1253 4782 6.68 | 158% 74% 101%
ja_JP TestNames_Japanese_k.txt | 489 2765 3.32 | 5145 15923 6.68 | 952% 476% 101%
ja_JP TestNames_Asian.txt | 557 2356 3.49 | 2133 7767 6.81 | 283% 230% 95%
zh_CN TestNames_Latin.txt | 207 6555 1.59 | 2602 8003 5.92 | 1157% 22% 272%
zh_CN TestNames_Chinese.txt | 581 1697 4.50 | 1233 2969 9.43 | 112% 75% 110%
zh_TW TestNames_Latin.txt | 207 6555 1.59 | 2618 8022 5.92 | 1165% 22% 272%
zh_TW TestNames_Chinese.txt | 502 1393 4.28 | 1590 2747 7.34 | 217% 97% 71%
zh__PINYIN TestNames_Latin.txt | 204 6545 1.59 | 2604 8033 5.92 | 1176% 23% 272%
zh__PINYIN TestNames_Chinese.txt | 580 1717 4.50 | 1236 2959 9.43 | 113% 72% 110%
ko_KR TestNames_Latin.txt | 175 3951 1.59 | 2756 8284 5.92 | 1475% 110% 272%
ko_KR TestNames_Korean.txt | 837 2132 4.66 | 1535 2785 7.33 | 83% 31% 57%
ru_RU TestNames_Latin.txt | 177 3941 1.59 | 2614 7992 5.92 | 1377% 103% 272%
ru_RU TestNames_Russian.txt | 590 7324 1.66 | 3170 23588 5.85 | 437% 222% 252%
th TestNames_Latin.txt | 176 4655 1.67 | 2638 8072 5.90 | 1399% 73% 253%
th TestNames_Thai.txt | 1537 6630 1.67 | 2743 8136 5.76 | 78% 23% 245%
Notes
- The tests are performed on a Pentium M machine, 1.6GHz with 1GB RAM,
running Windows XP.
- We were responsible for the original implementation of collation in Java,
so we've learned quite a bit since then.
- As with all performance measurements, the results will vary according to
the hardware and compiler. The strcoll operation is particularly sensitive; we
have found that even slight changes in code alignment can produce 10%
differences.
- For more information on incremental vs. sort key comparison, the importance
of multi-level sorting, and other features of collation, see Unicode Collation (UCA).
- For general information on ICU collation see User Guide.
- For information on APIs, see C,
C++, or Java.
|
