Comments on: strace: The Friend You Never Knew You Had

by: tothzp

Tue, 01 Apr 2008 09:25:00 +0000

1. A pretty big amount of data can be written to the buffer cache first. So when write() returns you don’t know if the data has been physically written to the disk or not. (Except if you flush your file to disk, e.g. with the fsync() or fdatasync() function.)

2. strace can be also used to collect statistics about the time spent in different system calls. Use the -c option for that:

# strace -p 16640 -c

Process 16640 attached - interrupt to quit Process 16640 detached % time seconds usecs/call calls errors syscall ------ ----------- ----------- --------- --------- ---------------- 36.72 0.649557 36 17944 epoll_wait 33.40 0.590959 7 85036 gettimeofday 21.75 0.384777 16 23568 writev 6.65 0.117595 6 19767 recv 1.48 0.026198 8 3354 time ------ ----------- ----------- --------- --------- ---------------- 100.00 1.769086 149669 total

by: tombauer

Tue, 01 Apr 2008 04:28:03 +0000

I think an interesting question would be why does the compiler create code that does the write with two write system calls. Why not one? And what is special about 388K? Hmmm.

by: ohmay

Fri, 28 Mar 2008 18:51:33 +0000

And for such “small” writes, how interacts the buffer-cache? In the article it is said that the difference between write rates are because the latency of the drive, the speed of the drive and this kind of things.

Buf for small writes the data writed shouldn’t be writed on buffer-cache?, and because of that at memory speed? Perhaps I’m wrong, but I think that the difference between write rates might be related to the initialization of I/O functions.

In any case, great article.

by: Marcin Owsiany

Fri, 28 Mar 2008 12:45:06 +0000

There is ‘ltrace’ as well. More useful sometimes.

by: parmando

Fri, 28 Mar 2008 10:31:20 +0000

Muy bueno, me di cuenta de varias cosas que pasaba por alto al ver la salida de strace!!

by: gserban

Fri, 28 Mar 2008 02:01:38 +0000

Cool. I actually had no idea this command existed.