Efficient data transfer through zero copy
Zero copy, zero overhead
Many Web applications serve a significant amount of static content, which amounts to reading data off of a disk and writing the exact same data back to the response socket. This activity might appear to require relatively little CPU activity, but it's somewhat inefficient: the kernel reads the data off of disk and pushes it across the kernel-user boundary to the application, and then the application pushes it back across the kernel-user boundary to be written out to the socket. In effect, the application serves as an inefficient intermediary that gets the data from the disk file to the socket.
Each time data traverses the user-kernel boundary, it must be copied, which consumes CPU cycles and memory bandwidth. Fortunately, you can eliminate these copies through a technique called — appropriately enough — zero copy. Applications that use zero copy request that the kernel copy the data directly from the disk file to the socket, without going through the application. Zero copy greatly improves application performance and reduces the number of context switches between kernel and user mode.
The Java class libraries support zero copy on Linux and UNIX systems
transferTo() method in
java.nio.channels.FileChannel. You can use the
transferTo() method to transfer bytes directly from the
channel on which it is invoked to another writable byte channel, without
requiring data to flow through the application. This article first
demonstrates the overhead incurred by simple file transfer done through
traditional copy semantics, then shows how the zero-copy technique using
transferTo() achieves better performance.
Date transfer: The traditional approach
Consider the scenario of reading from a file and transferring the data to another program over the network. (This scenario describes the behavior of many server applications, including Web applications serving static content, FTP servers, mail servers, and so on.) The core of the operation is in the two calls in Listing 1 (see Download for a link to the complete sample code):
Listing 1. Copying bytes from a file to a socket
File.read(fileDesc, buf, len); Socket.send(socket, buf, len);
Although Listing 1 is conceptually simple, internally, the copy operation requires four context switches between user mode and kernel mode, and the data is copied four times before the operation is complete. Figure 1 shows how data is moved internally from the file to the socket:
Figure 1. Traditional data copying approach
Figure 2 shows the context switching:
Figure 2. Traditional context switches
The steps involved are:
read()call causes a context switch (see Figure 2) from user mode to kernel mode. Internally a
sys_read()(or equivalent) is issued to read the data from the file. The first copy (see Figure 1) is performed by the direct memory access (DMA) engine, which reads file contents from the disk and stores them into a kernel address space buffer.
- The requested amount of data is copied from the read buffer into the
user buffer, and the
read()call returns. The return from the call causes another context switch from kernel back to user mode. Now the data is stored in the user address space buffer.
send()socket call causes a context switch from user mode to kernel mode. A third copy is performed to put the data into a kernel address space buffer again. This time, though, the data is put into a different buffer, one that is associated with the destination socket.
send()system call returns, creating the fourth context switch. Independently and asynchronously, a fourth copy happens as the DMA engine passes the data from the kernel buffer to the protocol engine.
Use of the intermediate kernel buffer (rather than a direct transfer of the data into the user buffer) might seem inefficient. But intermediate kernel buffers were introduced into the process to improve performance. Using the intermediate buffer on the read side allows the kernel buffer to act as a "readahead cache" when the application hasn't asked for as much data as the kernel buffer holds. This significantly improves performance when the requested data amount is less than the kernel buffer size. The intermediate buffer on the write side allows the write to complete asynchronously.
Unfortunately, this approach itself can become a performance bottleneck if the size of the data requested is considerably larger than the kernel buffer size. The data gets copied multiple times among the disk, kernel buffer, and user buffer before it is finally delivered to the application.
Zero copy improves performance by eliminating these redundant data copies.
Data transfer: The zero-copy approach
If you re-examine the traditional scenario, you'll
notice that the second and third data copies are not actually required.
The application does nothing other than cache the data and transfer it
back to the socket buffer. Instead, the data could be transferred directly
from the read buffer to the socket buffer. The
method lets you do exactly this. Listing 2 shows the method signature of
Listing 2. The
public void transferTo(long position, long count, WritableByteChannel target);
transferTo() method transfers data from the file channel
to the given writable byte channel. Internally, it depends on the
underlying operating system's support for zero copy; in UNIX and various
flavors of Linux, this call is routed to the
system call, shown in Listing 3, which transfers data from one file
descriptor to another:
Listing 3. The
#include <sys/socket.h> ssize_t sendfile(int out_fd, int in_fd, off_t *offset, size_t count);
The action of the
calls in Listing 1 can be replaced by a single
transferTo() call, as shown in Listing 4:
Listing 4. Using
transferTo() to copy data from a disk file to a
transferTo(position, count, writableChannel);
Figure 3 shows the data path when the
transferTo() method is
Figure 3. Data copy with
Figure 4 shows the context switches when the
method is used:
Figure 4. Context switching with
The steps taken when you use
transferTo() as in Listing 4 are:
transferTo()method causes the file contents to be copied into a read buffer by the DMA engine. Then the data is copied by the kernel into the kernel buffer associated with the output socket.
- The third copy happens as the DMA engine passes the data from the kernel socket buffers to the protocol engine.
This is an improvement: we've reduced the number of context switches from four to two and reduced the number of data copies from four to three (only one of which involves the CPU). But this does not yet get us to our goal of zero copy. We can further reduce the data duplication done by the kernel if the underlying network interface card supports gather operations. In Linux kernels 2.4 and later, the socket buffer descriptor was modified to accommodate this requirement. This approach not only reduces multiple context switches but also eliminates the duplicated data copies that require CPU involvement. The user-side usage still remains the same, but the intrinsics have changed:
transferTo()method causes the file contents to be copied into a kernel buffer by the DMA engine.
- No data is copied into the socket buffer. Instead, only descriptors with information about the location and length of the data are appended to the socket buffer. The DMA engine passes data directly from the kernel buffer to the protocol engine, thus eliminating the remaining final CPU copy.
Figure 5 shows the data copies using
transferTo() with the
Figure 5. Data copies when
transferTo() and gather operations are used
Building a file server
Now let's put zero copy into practice, using the same example of
transferring a file between a client and a server (see Download for the sample code).
TraditionalServer.java are based on the traditional copy
TraditionalServer.java is a server program that listens on a
particular port for the client to connect, and then reads 4K bytes of data
at a time from the socket.
TraditionalClient.java connects to
the server, reads (using
File.read()) 4K bytes of data from a
file, and sends (using
socket.send()) the contents to the
server via the socket.
TransferToClient.java perform the same function, but instead
transferTo() method (and in turn the
sendfile() system call) to transfer the file from server to
We executed the sample programs on a Linux system running the 2.6 kernel
and measured the run time in milliseconds for both the traditional
approach and the
transferTo() approach for various sizes.
Table 1 shows the results:
Table 1. Performance comparison: Traditional approach vs. zero copy
|File size||Normal file transfer (ms)||transferTo (ms)|
As you can see, the
transferTo() API brings down the time
approximately 65 percent compared to the traditional approach. This has
the potential to increase performance significantly for applications that
do a great deal of copying of data from one I/O channel to another, such
as Web servers.
We have demonstrated the performance advantages of using
transferTo() compared to reading from one channel and writing
the same data to another. 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, the zero-copy
technique can offer a significant performance improvement.