dtrace discuss - Mar 2009 - Dtrace and applications I/O

Hye everybody,

I just try to write a simple DTrace program to trace I/O from the application
point of view (and not the Disk Driver point of view (io::*:* provider)).

My first idea is to trace calls to syscall read and write.
Is it enougth to trace efficiently all application I/O or not ?

Thanks
R.
-- 
This message posted from opensolaris.org

Generally, yes, using syscalls is a great way to profile the IO
load from your applications. You need to figure out which
calls they''re using - read(2), or pread(2), etc. Start with a
generic syscall profile, and go from there.

Once you have the system call profile, you can tweak your
dtrace to breakdown IO by file type, file system type, etc.
The fds[] array is a joy to use. Makes it very easy to build scripts
with simple predicates to profile disk versus network, etc.

For example

syscall::read:entry
/ fds[arg0].fi_fs == "ufs" /
{
    @ufs_ios = count();
}
syscall::read:entry
/ fds[arg0].fi_fs == "sockfs" /
{
    @net_ios = count();
}
....

You can profile on file pathnames easily using fds[arg0].fi_pathname.

Data rates, etc...all is possible with the proper dose of vitamin D.

:^)

/jim


Ferrand wrote:
> Hye everybody,
>
> I just try to write a simple DTrace program to trace I/O from the
application point of view (and not the Disk Driver point of view (io::*:*
provider)).
>
> My first idea is to trace calls to syscall read and write.
> Is it enougth to trace efficiently all application I/O or not ?
>
> Thanks
> R.
>

Thanks for your answer.

That''s true that i got real results with those syscalls (read, write,
pread, pwrite), but i also noticed that the FSINFO provider (fsinfo::fop_read ||
fsinfo::fop_write) can give me those informations from the VFS point of view.

According to me, the VFS point of view is better for we only use real Disk I/O
created by applications (and no write to PTs or TTY). What''s your point
of view ?

I also have another problem : the mmap() access in Write and Read to Disk,
neither of those two methods is efficient to give those access (because mmap is
also used for memory access and i can''t differenciate memory access
from disk access)

Thanks.
R.
-- 
This message posted from opensolaris.org