Check if this typecast IO causing entire object to be read into the buffer

[VitalyFedyunin]

data/torchdata/datapipes/iter/load/s3io.py

Line 125 in 25cc03c

yield url, StreamWrapper(BytesIO(self.handler.s3_read(url)))

CC @ejguan

[ejguan]

Using the same script from #500 to test the performance w/ or w/o BytesIO from S3FileLoader.

• With BytesIO

S3FileLoader: Mean(23.807878967700006), STD(3.3510862470041345)
FSSpecFileOpener: Mean(2.4083272642999987), STD(0.3079321041894781)

• Without BytesIO

S3FileLoader: Mean(25.671961921199994), STD(6.733909007381986)
FSSpecFileOpener: Mean(2.278061599999995), STD(0.31031518396637964)

It seems the performance is on parity and sometimes even worse than it with BytesIO. So, closing the issue for now unless more investigation on typecast is required.