This session explains how you can improve performance of I/O-intensive Java™ applications through a technique called zero copy. Zero copy lets you to avoid redundant data copies between intermediate buffers and reduces number of context switches between user and kernel space. Background: Many applications [Web Servers, FTP-like services] serve a significant amount of static content, which amounts to reading data off of a disk and writing the exact data back to the response socket. Each time data traverses the user-kernel boundary; it must be copied, which consumes CPU cycles and memory bandwidth. Fortunately, we can eliminate these copies through a technique called — zero copy. The Java class libraries support zero copy on Linux and UNIX systems through the transferTo() method in java.nio.channels.FileChannel. Session Agenda: The session will initially focus on “Zero Copy” concept and its relevance in Data transfer applications. The traditional approach of transferring data b/w processes using File and Socket I/O will be explained in detail. It will demonstrate the overhead incurred when using traditional copy semantics, and will show how transferTo() achieves better performance. transferTo() API brings down the time 65% compared to traditional approach. Summary The Session demonstrates performance advantages of using transferTo() compared to the traditional approach. Intermediate buffer copies — even those hidden in the kernel — can have a measurable cost. In applications that do a great deal of copying of data between channels, zero-copy technique can offer a significant performance improvement.

1. Zero Copy: A Green Approach
to Data Transfer
Pramod B Nagaraja
IBM India Pvt LTD

2. Agenda

Zero Copy

Data Transfer : Age-Old Practice

Data Transfer : Best Practice

Performance Comparison

Application Scenario

Limitations

Conclusion

3. What is Zero Copy ?
•
Data transfer from disk to user buffers without
an intermediate copy into OS Kernel buffers
•
Computer operations in which CPU does not
perform the task of copying data from one
memory area to another
• DMA
• MMU

4. What is Zero Copy ?
•
Fewer intermediate Data Copies, fewer
context switches
•
Linux :: sendfile, sendfile64

5. Agenda

Zero Copy

Data Transfer : Age-Old Practice

Data Transfer : Best Practice

Performance Comparison

Application Scenario

Limitations

Conclusion

6. Application Scenario

7. Data Transfer : Age-Old Practice
File.read(fd, buf, len);
Socket.send(socket, buf, len);
inputStream = new FileInputStream(fname);
output = new
DataOutputStream(socket.getOutputStream());
byte[] b = new byte[4096];
long read = 0, total = 0;
while((read = inputStream.read(b))>=0) {
total = total + read;
output.write(b);
}

8. Data Transfer : Age-Old Practice

4 Data Copies

9. Data Transfer : Age-Old Practice

4 Context Switches

10. Agenda

Zero Copy

Data Transfer : Age-Old Practice

Data Transfer : Best Practice

Performance Comparison

Application Scenario

Limitations

Conclusion

11. Data Transfer : Best Practice

FileChannel Approach
java.nio.channels.FileChannel.transferTo(pos, cnt, writableChannel);
SocketChannel sc = SocketChannel.open();
sc.connect(sad);
FileChannel fc = new
FileInputStream(fname).getChannel();
long nsent = 0, curnset = 0;
curnset = fc.transferTo(0, fsize, sc);

12. Data Transfer : Best Practice
 3 Data Copies

13. Data Transfer : Best Practice
 2 Context Switches

14. Agenda

Zero Copy

Data Transfer : Age-Old Practice

Data Transfer : Best Practice
•
Performance Comparison

Application Scenario

Limitations

Conclusion

15. Performance Comparison
• ~ 65% Performance Boost up !
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
63
M B
350
M B
1
GB
Traditio
nal
FileCha
nnel
Measurements take on Linux (xi32) with 2.6 kernel

16. Agenda

Zero Copy

Data Transfer : Age-Old Practice

Data Transfer : Best Practice

Performance Comparison
•
Application Scenario
•
Limitations
•
Conclusion

17. Application Scenario

FTP Servers

Web Servers

Mail Servers

Any Applications serving static content

18. Agenda

Zero Copy

Data Transfer : Age-Old Practice

Data Transfer : Best Practice

Performance Comparison

Application Scenario
•
Limitations
•
Conclusion

19. Theoretically Speaking there is no difference
b/w Theory and Practice .. but in Practice
there is !!

20. Limitations

Suitable only for static content

Not suitable if Data to be transferred needs to
be modified/appended with/appended to, etc

FileChannels can only be used to transfer
data from File to File, File to Socket objects.

21. References

More details can be found @
http://www.ibm.com/developerworks/library/j-
zerocopy/index.html

http://www.linuxjournal.com/article/6345

22. Q&A

23. Thank You !!!!