Advanced Programming in the UNIX® Environment

ptg10805159

86 File I/O Chapter 3

If we add the line

set_fl(STDOUT_FILENO, O_SYNC);

to the beginning of the program shown in Figure3.5, we’ll turn on the synchronous-

write flag. This causes eachwriteto wait for the data to be written to disk before

returning. Normally in the UNIX System, awriteonly queues the data for writing; the

actual disk write operation can take place sometime later.Adatabase system is a likely

candidate for usingO_SYNC, so that it knows on return from awritethat the data is

actually on the disk, in case of an abnormal system failure.

We expect theO_SYNC flag to increase the system and clock times when the

program runs. Totest this, we can run the program in Figure3.5, copying 492.6 MB of

data from one file on disk to another and comparethis with a version that does the

same thing with theO_SYNCflag set. The results from a Linux system using theext4

file system areshown in Figure3.13.

User CPU System CPU Clock time

Operation (seconds) (seconds) (seconds)

read time from Figure3.6 forBUFFSIZE=4,096 0.03 0.58 8.62

normalwriteto disk file 0.00 1.05 9.70

writeto disk file withO_SYNCset 0.02 1.09 10.28

writeto disk followed byfdatasync 0.02 1.14 17.93

writeto disk followed byfsync 0.00 1.19 18.17

writeto disk withO_SYNCset followed byfsync 0.02 1.15 17.88

Figure 3.13Linuxext4timing results using various synchronization mechanisms

The six rows in Figure3.13 wereall measured with aBUFFSIZEof 4,096 bytes. The

results in Figure3.6 weremeasured while reading a disk file and writing to

/dev/null, so therewas no disk output. The second row in Figure3.13 corresponds to

reading a disk file and writing to another disk file. This is why the first and second

rows in Figure3.13 aredifferent. The system time increases when we write to a disk

file, because the kernel now copies the data from our process and queues the data for

writing by the disk driver.Weexpect the clock time to increase as well when we write

to a disk file.

When we enable synchronous writes, the system and clock times should increase

significantly.Asthe thirdrow shows, the system time for writing synchronously is not

much moreexpensive than when we used delayed writes. This implies that the Linux

operating system is doing the same amount of work for delayed and synchronous

writes (which is unlikely), or else theO_SYNCflag isn’t having the desired effect. In this

case, the Linux operating system isn’t allowing us to set theO_SYNCflag usingfcntl,

instead failing without returning an error (but it would have honored the flag if we

wereable to specify it when the file was opened).

The clock time in the last three rows reflects the extra time needed to wait for all of

the writes to be committed to disk. After writing a file synchronously, we expect that a

call tofsyncwill have no effect. This case is supposed to be represented by the last