[IOTDB-2205] Optimize unaligned int/long read/write functions in BytesUtils

[543202718]

注意到在现有的代码中，BytesUtils中非对齐的int/long的读写函数一次循环只能处理一个比特，效率较低，可以对其进行优化，使得一次循环可以处理一个字节。

该优化改动了下面的函数：

• void longToBytes(long srcNum, byte[] result, int pos, int width)
• void intToBytes(int srcNum, byte[] result, int pos, int width)
• long bytesToLong(byte[] result, int pos, int width)
• int bytesToInt(byte[] result, int pos, int width)

通过对用例的查找，发现该优化将会对使用TS_2DIFF编码的数据读写产生影响。

单元性能测试表明，该优化可以提高函数接近5倍的性能。总体性能测试表明，对于采用TS_2DIFF编码且不压缩的INT32/INT64数据，该优化可以减少约40%的读取时间开销。

---

In the current codes, the read and write functions of unaligned int/long in BytesUtils can only process one bit per loop, which is inefficient. It can be optimized to one byte per loop.

The following functions are optimized:

• void longToBytes(long srcNum, byte[] result, int pos, int width)
• void intToBytes(int srcNum, byte[] result, int pos, int width)
• long bytesToLong(byte[] result, int pos, int width)
• int bytesToInt(byte[] result, int pos, int width)

Through the use case search, the optimization influences the reading and writing of data encoded by TS_2DIFF.

In the unit performance test, the optimized functions are about 5x faster the current ones. In the overall test, about 40% time cost is reduced in the query of the uncompressed TS_2DIFF-coded INT32/INT64 data due to the optimization .

[JackieTien97]

Great job! I wonder that how much improvement does this change bring to write performance？ Since you only mentioned the query performance improvements in your pr description.

[543202718]

With the help of IoTDB-benchmark, I measured the performance improvement. Due to the difference in test data, the results are slightly different from those previously reported.

• The write throughput before improvement is 232,268 pt/s and 237,046 pt/s after improvement. The gap is within the fluctuation range because the write bottleneck is not the optimized function.
• The query throughput before improvement is 5,648,345 pt/s and 6,767,949 pt/s after improvement. Thers is a 20% improvement.

The modified config of IoTDB-benchmark is listed as follows while others are still the default.

LOOP=1000
SENSOR_NUMBER=4
CLIENT_NUMBER=1
POINT_STEP=1000
INSERT_DATATYPE_PROPORTION=0:1:1:1:1:0
COMPRESSOR=UNCOMPRESSED
ENCODING_INT32=TS_2DIFF
ENCODING_INT64=TS_2DIFF
ENCODING_FLOAT=TS_2DIFF
ENCODING_DOUBLE=TS_2DIFF
LINE_RATIO=0
SIN_RATIO=0
SQUARE_RATIO=0
RANDOM_RATIO=1
CONSTANT_RATIO=0
BATCH_SIZE_PER_WRITE=10000
QUERY_INTERVAL=100000000000

OPERATION_PROPORTION=1:0:0:0:0:0:0:0:0:0:0 (for write)
OPERATION_PROPORTION=0:0:1:0:0:0:0:0:0:0:0 (for query)

[HTHou]

Hi, plz run mvn spotless:apply to fix the format issue.

[543202718]

Hi, plz run mvn spotless:apply to fix the format issue.

Thanks. The format issue is fixed in the new commit.

[jihaitangguo]

Note: There was a bug in the code submitted here, IOTDB-4633, which has been fixed in #7669, where

byte y = (byte) (srcNum >> width);

has been corrected to

byte y = (byte) (srcNum >>> width);