zero-copy

Zero copy lets you avoid redundant data copies between intermediate buffers and reduces the number of context switches between user-space and kernel-space. Ideal Zero copy (zero cpu copy) is possible when your hardware (disk drive, network card, graphic card, sound card) supports DMA (Direct Memory Acess).

Example below demonstrates zero and non-zero file copy in Java and C.

Let's try out example in C

To try out zero and non-zero copy we need files. Let's quiclky create 2 files each 1G.

fallocate -l 1G /tmp/rbigfile.dat
fallocate -l 1G /tmp/wbigfile.dat

It's time to compile our c code:

gcc -Wall -W -Werror fcopy.c -o fcopy

Now let's execute non-zero file copy where 8K buffer is used to read and write data.

time ./fcopy /tmp/rbigfile.dat /tmp/wbigfile.dat 1

-----------------------
real  0m13.528s 
user  0m0.080s 
sys   0m2.160s
-----------------------

Now let's execute zero file copy where sendfile is used to copy data between one file descriptor and another.

time ./fcopy /tmp/rbigfile.dat /tmp/wbigfile.dat 2

-----------------------
real  0m12.725s
user  0m0.000s
sys   0m0.880s
-----------------------

As copying is done within the kernel space, sendfile is more efficient than the combination of read and write, which would require transferring data to and from user space.

Let's try out example in Java

Compiling a java class is as simple as:

javac JioChannel.java

Now let's execute non-zero file copy where 8K buffer is used to read and write data.

time java JioChannel /tmp/rbigfile.dat /tmp/wbigfile.dat 1

-----------------------
real  0m14.445s
user  0m0.364s
sys   0m1.956s
-----------------------

Now let's execute zero file copy where FileChannel.transferTo() is used to copy data between one file channel and another.

time java JioChannel /tmp/rbigfile.dat /tmp/wbigfile.dat 2

-----------------------
real  0m15.473s
user  0m0.124s
sys   0m1.516s
-----------------------

Again as copying is done within the kernel space, FileChannel.transferTo() is more efficient than the combination of read and write, which would require transferring data to and from user space.

###real vs user vs sys

real - elapsed "wall clock" time, like using a stop watch. It includes time slices used by other processes and time the process spends blocked.
user - amount of CPU time spent in user-space code (outside the kernel-space) within the process.
sys - amount of CPU time spent in the kernel-space within the process. CPU time spent in system calls.
user + sys - actual CPU time used by process.

Conclusion

You have seen performance advantages of using sendfile and transferTo() compared to reading from one channel and writing the same data to another. Frameworks like netty, grizzly, kafka and others rely on zero-copy to have better throughput, lower latency.