In 4.3.1, LazyDecompressingInputStream was introduced. However, LazyDecompressingInputStream subclasses InputStream without overriding the multi-byte read() method, and the inherited method does a byte-by-byte read.
This is trace showing what happens:
copyContent() would love to read(byte,int,int) in a buffer, but since LazyDecompressingInputStream doesn't override it it invokes instead the read-byte-by-byte inherited method in InputStream, which in turn now calls for each byte the one-byte read() method from LazyDecompressingInputStream, which invokes the one-byte read method from InflaterInputStream, which does a multi-byte, length-one read from GZIPInputStream, which unleashes a similar call on InflaterInputStream, which unfortunately makes a similar read using the native inflateBytes() method.
Thus, for each byte there is a native-method call. The result is a 10-50x increase in CPU usage, which turns into a 10x-50x decrease in speed if, as in our case, you have 7000 threads downloading in parallel.
Overriding read(byte,int,int) in LazyDecompressingInputStream will solve the problem:
public int read(byte b, int off, int len) throws IOException